Skip to content

P
PaddleOCR

weixin_41840029 / PaddleOCR
与 Fork 源项目一致

Fork自 PaddlePaddle / PaddleOCR

通知 1
Star 1
Fork 0
P
PaddleOCR
提交 4b214948 编写于 作者: z37757's avatar z37757
浏览文件
操作

Merge branch 'dygraph' of https://github.com/PaddlePaddle/PaddleOCR into new_branch

上级 917606b4 6e607a0f
显示空白变更内容
内联 并排
Showing with 20794 addition and 1779 deletion
PPOCRLabel/PPOCRLabel.py
浏览文件 @ 4b214948
......@@ -10,95 +10,65 @@
# SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF
# CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
# THE SOFTWARE.
#!/usr/bin/env python
# !/usr/bin/env python
# -*- coding: utf-8 -*-
# pyrcc5 -o libs/resources.py resources.qrc
import argparse
import ast
import codecs
import json
import os.path
import platform
import subprocess
import sys
from functools import partial
from collections import defaultdict
import json
import cv2
from PyQt5.QtCore import QSize, Qt, QPoint, QByteArray, QTimer, QFileInfo, QPointF, QProcess
from PyQt5.QtGui import QImage, QCursor, QPixmap, QImageReader
from PyQt5.QtWidgets import QMainWindow, QListWidget, QVBoxLayout, QToolButton, QHBoxLayout, QDockWidget, QWidget, \
QSlider, QGraphicsOpacityEffect, QMessageBox, QListView, QScrollArea, QWidgetAction, QApplication, QLabel, \
QFileDialog, QListWidgetItem, QComboBox, QDialog
__dir__ = os.path.dirname(os.path.abspath(__file__))
import numpy as np
sys.path.append(__dir__)
sys.path.append(os.path.abspath(os.path.join(__dir__, '../..')))
sys.path.append(os.path.abspath(os.path.join(__dir__, '../PaddleOCR')))
sys.path.append("..")
from paddleocr import PaddleOCR
try:
from PyQt5 import QtCore, QtGui, QtWidgets
from PyQt5.QtGui import *
from PyQt5.QtCore import *
from PyQt5.QtWidgets import *
except ImportError:
# needed for py3+qt4
# Ref:
# http://pyqt.sourceforge.net/Docs/PyQt4/incompatible_apis.html
# http://stackoverflow.com/questions/21217399/pyqt4-qtcore-qvariant-object-instead-of-a-string
if sys.version_info.major >= 3:
import sip
sip.setapi('QVariant', 2)
from PyQt4.QtGui import *
from PyQt4.QtCore import *
from combobox import ComboBox
from libs.constants import *
from libs.utils import *
from libs.labelColor import label_colormap
from libs.settings import Settings
from libs.shape import Shape, DEFAULT_LINE_COLOR, DEFAULT_FILL_COLOR,DEFAULT_LOCK_COLOR
from libs.shape import Shape, DEFAULT_LINE_COLOR, DEFAULT_FILL_COLOR, DEFAULT_LOCK_COLOR
from libs.stringBundle import StringBundle
from libs.canvas import Canvas
from libs.zoomWidget import ZoomWidget
from libs.autoDialog import AutoDialog
from libs.labelDialog import LabelDialog
from libs.colorDialog import ColorDialog
from libs.toolBar import ToolBar
from libs.ustr import ustr
from libs.hashableQListWidgetItem import HashableQListWidgetItem
from libs.editinlist import EditInList
from libs.unique_label_qlist_widget import UniqueLabelQListWidget
from libs.keyDialog import KeyDialog
__appname__ = 'PPOCRLabel'
LABEL_COLORMAP = label_colormap()
class WindowMixin(object):
def menu(self, title, actions=None):
menu = self.menuBar().addMenu(title)
if actions:
addActions(menu, actions)
return menu
def toolbar(self, title, actions=None):
toolbar = ToolBar(title)
toolbar.setObjectName(u'%sToolBar' % title)
# toolbar.setOrientation(Qt.Vertical)
toolbar.setToolButtonStyle(Qt.ToolButtonTextUnderIcon)
if actions:
addActions(toolbar, actions)
self.addToolBar(Qt.LeftToolBarArea, toolbar)
return toolbar
class MainWindow(QMainWindow, WindowMixin):
class MainWindow(QMainWindow):
FIT_WINDOW, FIT_WIDTH, MANUAL_ZOOM = list(range(3))
def __init__(self, lang="ch", gpu=False, defaultFilename=None, defaultPrefdefClassFile=None, defaultSaveDir=None):
def __init__(self,
lang="ch",
gpu=False,
kie_mode=False,
default_filename=None,
default_predefined_class_file=None,
default_save_dir=None):
super(MainWindow, self).__init__()
self.setWindowTitle(__appname__)
self.setWindowState(Qt.WindowMaximized) # set window max
......@@ -109,14 +79,27 @@ class MainWindow(QMainWindow, WindowMixin):
self.settings.load()
settings = self.settings
self.lang = lang
# Load string bundle for i18n
if lang not in ['ch', 'en']:
lang = 'en'
self.stringBundle = StringBundle.getBundle(localeStr='zh-CN' if lang=='ch' else 'en') # 'en'
self.stringBundle = StringBundle.getBundle(localeStr='zh-CN' if lang == 'ch' else 'en') # 'en'
getStr = lambda strId: self.stringBundle.getString(strId)
self.defaultSaveDir = defaultSaveDir
self.ocr = PaddleOCR(use_pdserving=False, use_angle_cls=True, det=True, cls=True, use_gpu=gpu, lang=lang, show_log=False)
# KIE setting
self.kie_mode = kie_mode
self.key_previous_text = ""
self.existed_key_cls_set = set()
self.key_dialog_tip = getStr('keyDialogTip')
self.defaultSaveDir = default_save_dir
self.ocr = PaddleOCR(use_pdserving=False,
use_angle_cls=True,
det=True,
cls=True,
use_gpu=gpu,
lang=lang,
show_log=False)
if os.path.exists('./data/paddle.png'):
result = self.ocr.ocr('./data/paddle.png', cls=True, det=True)
......@@ -134,7 +117,6 @@ class MainWindow(QMainWindow, WindowMixin):
self.labelFile = None
self.currIndex = 0
# Whether we need to save or not.
self.dirty = False
......@@ -144,7 +126,7 @@ class MainWindow(QMainWindow, WindowMixin):
self.screencast = "https://github.com/PaddlePaddle/PaddleOCR"
# Load predefined classes to the list
self.loadPredefinedClasses(defaultPrefdefClassFile)
self.loadPredefinedClasses(default_predefined_class_file)
# Main widgets and related state.
self.labelDialog = LabelDialog(parent=self, listItem=self.labelHist)
......@@ -160,12 +142,14 @@ class MainWindow(QMainWindow, WindowMixin):
self.PPreader = None
self.autoSaveNum = 5
################# file list ###############
# ================== File List ==================
filelistLayout = QVBoxLayout()
filelistLayout.setContentsMargins(0, 0, 0, 0)
self.fileListWidget = QListWidget()
self.fileListWidget.itemClicked.connect(self.fileitemDoubleClicked)
self.fileListWidget.setIconSize(QSize(25, 25))
filelistLayout = QVBoxLayout()
filelistLayout.setContentsMargins(0, 0, 0, 0)
filelistLayout.addWidget(self.fileListWidget)
self.AutoRecognition = QToolButton()
......@@ -181,15 +165,29 @@ class MainWindow(QMainWindow, WindowMixin):
fileListContainer = QWidget()
fileListContainer.setLayout(filelistLayout)
self.fileListName = getStr('fileList')
self.filedock = QDockWidget(self.fileListName, self)
self.filedock.setObjectName(getStr('files'))
self.filedock.setWidget(fileListContainer)
self.addDockWidget(Qt.LeftDockWidgetArea, self.filedock)
######## Right area ##########
self.fileDock = QDockWidget(self.fileListName, self)
self.fileDock.setObjectName(getStr('files'))
self.fileDock.setWidget(fileListContainer)
self.addDockWidget(Qt.LeftDockWidgetArea, self.fileDock)
# ================== Key List ==================
if self.kie_mode:
# self.keyList = QListWidget()
self.keyList = UniqueLabelQListWidget()
# self.keyList.itemSelectionChanged.connect(self.keyListSelectionChanged)
# self.keyList.itemDoubleClicked.connect(self.editBox)
# self.keyList.itemChanged.connect(self.keyListItemChanged)
self.keyListDockName = getStr('keyListTitle')
self.keyListDock = QDockWidget(self.keyListDockName, self)
self.keyListDock.setWidget(self.keyList)
self.keyListDock.setFeatures(QDockWidget.NoDockWidgetFeatures)
filelistLayout.addWidget(self.keyListDock)
# ================== Right Area ==================
listLayout = QVBoxLayout()
listLayout.setContentsMargins(0, 0, 0, 0)
# Buttons
self.editButton = QToolButton()
self.reRecogButton = QToolButton()
self.reRecogButton.setIcon(newIcon('reRec', 30))
......@@ -202,44 +200,44 @@ class MainWindow(QMainWindow, WindowMixin):
self.DelButton = QToolButton()
self.DelButton.setToolButtonStyle(Qt.ToolButtonTextBesideIcon)
leftTopToolBox = QHBoxLayout()
leftTopToolBox.addWidget(self.newButton)
leftTopToolBox.addWidget(self.reRecogButton)
leftTopToolBoxContainer = QWidget()
leftTopToolBoxContainer.setLayout(leftTopToolBox)
listLayout.addWidget(leftTopToolBoxContainer)
lefttoptoolbox = QHBoxLayout()
lefttoptoolbox.addWidget(self.newButton)
lefttoptoolbox.addWidget(self.reRecogButton)
lefttoptoolboxcontainer = QWidget()
lefttoptoolboxcontainer.setLayout(lefttoptoolbox)
listLayout.addWidget(lefttoptoolboxcontainer)
################## label list ####################
# ================== Label List ==================
# Create and add a widget for showing current label items
self.labelList = EditInList()
labelListContainer = QWidget()
labelListContainer.setLayout(listLayout)
#self.labelList.itemActivated.connect(self.labelSelectionChanged)
self.labelList.itemSelectionChanged.connect(self.labelSelectionChanged)
self.labelList.clicked.connect(self.labelList.item_clicked)
# Connect to itemChanged to detect checkbox changes.
self.labelList.itemChanged.connect(self.labelItemChanged)
self.labelListDock = QDockWidget(getStr('recognitionResult'),self)
self.labelListDockName = getStr('recognitionResult')
self.labelListDock = QDockWidget(self.labelListDockName, self)
self.labelListDock.setWidget(self.labelList)
self.labelListDock.setFeatures(QDockWidget.NoDockWidgetFeatures)
listLayout.addWidget(self.labelListDock)
################## detection box ####################
# ================== Detection Box ==================
self.BoxList = QListWidget()
#self.BoxList.itemActivated.connect(self.boxSelectionChanged)
# self.BoxList.itemActivated.connect(self.boxSelectionChanged)
self.BoxList.itemSelectionChanged.connect(self.boxSelectionChanged)
self.BoxList.itemDoubleClicked.connect(self.editBox)
# Connect to itemChanged to detect checkbox changes.
self.BoxList.itemChanged.connect(self.boxItemChanged)
self.BoxListDock = QDockWidget(getStr('detectionBoxposition'), self)
self.BoxListDockName = getStr('detectionBoxposition')
self.BoxListDock = QDockWidget(self.BoxListDockName, self)
self.BoxListDock.setWidget(self.BoxList)
self.BoxListDock.setFeatures(QDockWidget.NoDockWidgetFeatures)
listLayout.addWidget(self.BoxListDock)
############ lower right area ############
# ================== Lower Right Area ==================
leftbtmtoolbox = QHBoxLayout()
leftbtmtoolbox.addWidget(self.SaveButton)
leftbtmtoolbox.addWidget(self.DelButton)
......@@ -251,26 +249,26 @@ class MainWindow(QMainWindow, WindowMixin):
self.dock.setObjectName(getStr('labels'))
self.dock.setWidget(labelListContainer)
# ================== Zoom Bar ==================
self.imageSlider = QSlider(Qt.Horizontal)
self.imageSlider.valueChanged.connect(self.CanvasSizeChange)
self.imageSlider.setMinimum(-9)
self.imageSlider.setMaximum(510)
self.imageSlider.setSingleStep(1)
self.imageSlider.setTickPosition(QSlider.TicksBelow)
self.imageSlider.setTickInterval(1)
########## zoom bar #########
self.imgsplider = QSlider(Qt.Horizontal)
self.imgsplider.valueChanged.connect(self.CanvasSizeChange)
self.imgsplider.setMinimum(-150)
self.imgsplider.setMaximum(150)
self.imgsplider.setSingleStep(1)
self.imgsplider.setTickPosition(QSlider.TicksBelow)
self.imgsplider.setTickInterval(1)
op = QGraphicsOpacityEffect()
op.setOpacity(0.2)
self.imgsplider.setGraphicsEffect(op)
# self.imgsplider.setAttribute(Qt.WA_TranslucentBackground)
self.imgsplider.setStyleSheet("background-color:transparent")
self.imgsliderDock = QDockWidget(getStr('ImageResize'), self)
self.imgsliderDock.setObjectName(getStr('IR'))
self.imgsliderDock.setWidget(self.imgsplider)
self.imgsliderDock.setFeatures(QDockWidget.DockWidgetFloatable)
self.imgsliderDock.setAttribute(Qt.WA_TranslucentBackground)
self.addDockWidget(Qt.RightDockWidgetArea, self.imgsliderDock)
self.imageSlider.setGraphicsEffect(op)
self.imageSlider.setStyleSheet("background-color:transparent")
self.imageSliderDock = QDockWidget(getStr('ImageResize'), self)
self.imageSliderDock.setObjectName(getStr('IR'))
self.imageSliderDock.setWidget(self.imageSlider)
self.imageSliderDock.setFeatures(QDockWidget.DockWidgetFloatable)
self.imageSliderDock.setAttribute(Qt.WA_TranslucentBackground)
self.addDockWidget(Qt.RightDockWidgetArea, self.imageSliderDock)
self.zoomWidget = ZoomWidget()
self.colorDialog = ColorDialog(parent=self)
......@@ -278,13 +276,13 @@ class MainWindow(QMainWindow, WindowMixin):
self.msgBox = QMessageBox()
########## thumbnail #########
# ================== Thumbnail ==================
hlayout = QHBoxLayout()
m = (0, 0, 0, 0)
hlayout.setSpacing(0)
hlayout.setContentsMargins(*m)
self.preButton = QToolButton()
self.preButton.setIcon(newIcon("prev",40))
self.preButton.setIcon(newIcon("prev", 40))
self.preButton.setIconSize(QSize(40, 100))
self.preButton.clicked.connect(self.openPrevImg)
self.preButton.setStyleSheet('border: none;')
......@@ -294,10 +292,10 @@ class MainWindow(QMainWindow, WindowMixin):
self.iconlist.setFlow(QListView.TopToBottom)
self.iconlist.setSpacing(10)
self.iconlist.setIconSize(QSize(50, 50))
self.iconlist.setMovement(False)
self.iconlist.setMovement(QListView.Static)
self.iconlist.setResizeMode(QListView.Adjust)
self.iconlist.itemClicked.connect(self.iconitemDoubleClicked)
self.iconlist.setStyleSheet("background-color:transparent; border: none;")
self.iconlist.setStyleSheet("QListWidget{ background-color:transparent; border: none;}")
self.iconlist.setHorizontalScrollBarPolicy(Qt.ScrollBarAlwaysOff)
self.nextButton = QToolButton()
self.nextButton.setIcon(newIcon("next", 40))
......@@ -310,12 +308,11 @@ class MainWindow(QMainWindow, WindowMixin):
hlayout.addWidget(self.iconlist)
hlayout.addWidget(self.nextButton)
iconListContainer = QWidget()
iconListContainer.setLayout(hlayout)
iconListContainer.setFixedHeight(100)
########### Canvas ###########
# ================== Canvas ==================
self.canvas = Canvas(parent=self)
self.canvas.zoomRequest.connect(self.zoomRequest)
self.canvas.setDrawingShapeToSquare(settings.get(SETTING_DRAW_SQUARE, False))
......@@ -338,32 +335,17 @@ class MainWindow(QMainWindow, WindowMixin):
centerLayout = QVBoxLayout()
centerLayout.setContentsMargins(0, 0, 0, 0)
centerLayout.addWidget(scroll)
#centerLayout.addWidget(self.icondock)
centerLayout.addWidget(iconListContainer,0,Qt.AlignCenter)
centercontainer = QWidget()
centercontainer.setLayout(centerLayout)
# self.scrolldock = QDockWidget('WorkSpace',self)
# self.scrolldock.setObjectName('WorkSpace')
# self.scrolldock.setWidget(centercontainer)
# self.scrolldock.setFeatures(QDockWidget.NoDockWidgetFeatures)
# orititle = self.scrolldock.titleBarWidget()
# tmpwidget = QWidget()
# self.scrolldock.setTitleBarWidget(tmpwidget)
# del orititle
self.setCentralWidget(centercontainer) #self.scrolldock
self.addDockWidget(Qt.RightDockWidgetArea, self.dock)
centerLayout.addWidget(iconListContainer, 0, Qt.AlignCenter)
centerContainer = QWidget()
centerContainer.setLayout(centerLayout)
self.setCentralWidget(centerContainer)
self.addDockWidget(Qt.RightDockWidgetArea, self.dock)
# self.filedock.setFeatures(QDockWidget.DockWidgetFloatable)
self.filedock.setFeatures(self.filedock.features() ^ QDockWidget.DockWidgetFloatable)
self.dockFeatures = QDockWidget.DockWidgetClosable | QDockWidget.DockWidgetFloatable
self.dock.setFeatures(self.dock.features() ^ self.dockFeatures)
self.filedock.setFeatures(QDockWidget.NoDockWidgetFeatures)
self.dock.setFeatures(QDockWidget.DockWidgetClosable | QDockWidget.DockWidgetFloatable)
self.fileDock.setFeatures(QDockWidget.NoDockWidgetFeatures)
###### Actions #######
# ================== Actions ==================
action = partial(newAction, self)
quit = action(getStr('quit'), self.close,
'Ctrl+Q', 'quit', getStr('quitApp'))
......@@ -385,7 +367,7 @@ class MainWindow(QMainWindow, WindowMixin):
resetAll = action(getStr('resetAll'), self.resetAll, None, 'resetall', getStr('resetAllDetail'))
color1 = action(getStr('boxLineColor'), self.chooseColor1,
color1 = action(getStr('boxLineColor'), self.chooseColor,
'Ctrl+L', 'color_line', getStr('boxLineColorDetail'))
createMode = action(getStr('crtBox'), self.setCreateMode,
......@@ -410,7 +392,6 @@ class MainWindow(QMainWindow, WindowMixin):
'Ctrl+A', 'hide', getStr('showAllBoxDetail'),
enabled=False)
help = action(getStr('tutorial'), self.showTutorialDialog, None, 'help', getStr('tutorialDetail'))
showInfo = action(getStr('info'), self.showInfoDialog, None, 'help', getStr('info'))
showSteps = action(getStr('steps'), self.showStepsDialog, None, 'help', getStr('steps'))
......@@ -447,11 +428,12 @@ class MainWindow(QMainWindow, WindowMixin):
self.MANUAL_ZOOM: lambda: 1,
}
# ================== New Actions ==================
edit = action(getStr('editLabel'), self.editLabel,
'Ctrl+E', 'edit', getStr('editLabelDetail'),
enabled=False)
######## New actions #######
AutoRec = action(getStr('autoRecognition'), self.autoRecognition,
'', 'Auto', getStr('autoRecognition'), enabled=False)
......@@ -473,15 +455,18 @@ class MainWindow(QMainWindow, WindowMixin):
undoLastPoint = action(getStr("undoLastPoint"), self.canvas.undoLastPoint,
'Ctrl+Z', "undo", getStr("undoLastPoint"), enabled=False)
rotateLeft = action(getStr("rotateLeft"), partial(self.rotateImgAction,1),
rotateLeft = action(getStr("rotateLeft"), partial(self.rotateImgAction, 1),
'Ctrl+Alt+L', "rotateLeft", getStr("rotateLeft"), enabled=False)
rotateRight = action(getStr("rotateRight"), partial(self.rotateImgAction,-1),
rotateRight = action(getStr("rotateRight"), partial(self.rotateImgAction, -1),
'Ctrl+Alt+R', "rotateRight", getStr("rotateRight"), enabled=False)
undo = action(getStr("undo"), self.undoShapeEdit,
'Ctrl+Z', "undo", getStr("undo"), enabled=False)
change_cls = action(getStr("keyChange"), self.change_box_key,
'Ctrl+B', "edit", getStr("keyChange"), enabled=False)
lock = action(getStr("lockBox"), self.lockSelectedShape,
None, "lock", getStr("lockBoxDetail"),
enabled=False)
......@@ -495,7 +480,7 @@ class MainWindow(QMainWindow, WindowMixin):
# self.preButton.setDefaultAction(openPrevImg)
# self.nextButton.setDefaultAction(openNextImg)
############# Zoom layout ##############
# ================== Zoom layout ==================
zoomLayout = QHBoxLayout()
zoomLayout.addStretch()
self.zoominButton = QToolButton()
......@@ -522,14 +507,12 @@ class MainWindow(QMainWindow, WindowMixin):
icon='color', tip=getStr('shapeFillColorDetail'),
enabled=False)
# Label list context menu.
labelMenu = QMenu()
addActions(labelMenu, (edit, delete))
self.labelList.setContextMenuPolicy(Qt.CustomContextMenu)
self.labelList.customContextMenuRequested.connect(
self.popLabelListMenu)
self.labelList.customContextMenuRequested.connect(self.popLabelListMenu)
# Draw squares/rectangles
self.drawSquaresOption = QAction(getStr('drawSquares'), self)
......@@ -540,37 +523,35 @@ class MainWindow(QMainWindow, WindowMixin):
# Store actions for further handling.
self.actions = struct(save=save, resetAll=resetAll, deleteImg=deleteImg,
lineColor=color1, create=create, delete=delete, edit=edit, copy=copy,
saveRec=saveRec, singleRere=singleRere,AutoRec=AutoRec,reRec=reRec,
saveRec=saveRec, singleRere=singleRere, AutoRec=AutoRec, reRec=reRec,
createMode=createMode, editMode=editMode,
shapeLineColor=shapeLineColor, shapeFillColor=shapeFillColor,
zoom=zoom, zoomIn=zoomIn, zoomOut=zoomOut, zoomOrg=zoomOrg,
fitWindow=fitWindow, fitWidth=fitWidth,
zoomActions=zoomActions, saveLabel=saveLabel,
undo=undo, undoLastPoint=undoLastPoint,open_dataset_dir=open_dataset_dir,
rotateLeft=rotateLeft,rotateRight=rotateRight,lock=lock,
fileMenuActions=(
opendir, open_dataset_dir, saveLabel, resetAll, quit),
zoomActions=zoomActions, saveLabel=saveLabel, change_cls=change_cls,
undo=undo, undoLastPoint=undoLastPoint, open_dataset_dir=open_dataset_dir,
rotateLeft=rotateLeft, rotateRight=rotateRight, lock=lock,
fileMenuActions=(opendir, open_dataset_dir, saveLabel, resetAll, quit),
beginner=(), advanced=(),
editMenu=(createpoly, edit, copy, delete,singleRere,None, undo, undoLastPoint,
None, rotateLeft, rotateRight, None, color1, self.drawSquaresOption,lock),
beginnerContext=(create, edit, copy, delete, singleRere, rotateLeft, rotateRight,lock),
editMenu=(createpoly, edit, copy, delete, singleRere, None, undo, undoLastPoint,
None, rotateLeft, rotateRight, None, color1, self.drawSquaresOption, lock),
beginnerContext=(
create, edit, copy, delete, singleRere, rotateLeft, rotateRight, lock, change_cls),
advancedContext=(createMode, editMode, edit, copy,
delete, shapeLineColor, shapeFillColor),
onLoadActive=(
create, createMode, editMode),
onLoadActive=(create, createMode, editMode),
onShapesPresent=(hideAll, showAll))
# menus
self.menus = struct(
file=self.menu('&'+getStr('mfile')),
edit=self.menu('&'+getStr('medit')),
view=self.menu('&'+getStr('mview')),
file=self.menu('&' + getStr('mfile')),
edit=self.menu('&' + getStr('medit')),
view=self.menu('&' + getStr('mview')),
autolabel=self.menu('&PaddleOCR'),
help=self.menu('&'+getStr('mhelp')),
help=self.menu('&' + getStr('mhelp')),
recentFiles=QMenu('Open &Recent'),
labelList=labelMenu)
self.lastLabel = None
# Add option to enable/disable labels being displayed at the top of bounding boxes
self.displayLabelOption = QAction(getStr('displayLabel'), self)
......@@ -591,9 +572,10 @@ class MainWindow(QMainWindow, WindowMixin):
self.autoSaveOption.triggered.connect(self.autoSaveFunc)
addActions(self.menus.file,
(opendir, open_dataset_dir, None, saveLabel, saveRec, self.autoSaveOption, None, resetAll, deleteImg, quit))
(opendir, open_dataset_dir, None, saveLabel, saveRec, self.autoSaveOption, None, resetAll, deleteImg,
quit))
addActions(self.menus.help, (showKeys,showSteps, showInfo))
addActions(self.menus.help, (showKeys, showSteps, showInfo))
addActions(self.menus.view, (
self.displayLabelOption, self.labelDialogOption,
None,
......@@ -601,23 +583,19 @@ class MainWindow(QMainWindow, WindowMixin):
zoomIn, zoomOut, zoomOrg, None,
fitWindow, fitWidth))
addActions(self.menus.autolabel, (AutoRec, reRec, alcm, None, help)) #
addActions(self.menus.autolabel, (AutoRec, reRec, alcm, None, help))
self.menus.file.aboutToShow.connect(self.updateFileMenu)
# Custom context menu for the canvas widget:
addActions(self.canvas.menus[0], self.actions.beginnerContext)
#addActions(self.canvas.menus[1], (
# action('&Copy here', self.copyShape),
# action('&Move here', self.moveShape)))
self.statusBar().showMessage('%s started.' % __appname__)
self.statusBar().show()
# Application state.
self.image = QImage()
self.filePath = ustr(defaultFilename)
self.filePath = ustr(default_filename)
self.lastOpenDir = None
self.recentFiles = []
self.maxRecent = 7
......@@ -628,7 +606,7 @@ class MainWindow(QMainWindow, WindowMixin):
# Add Chris
self.difficult = False
## Fix the compatible issue for qt4 and qt5. Convert the QStringList to python list
# Fix the compatible issue for qt4 and qt5. Convert the QStringList to python list
if settings.get(SETTING_RECENT_FILES):
if have_qstring():
recentFileQStringList = settings.get(SETTING_RECENT_FILES)
......@@ -657,7 +635,6 @@ class MainWindow(QMainWindow, WindowMixin):
# Add chris
Shape.difficult = self.difficult
# ADD:
# Populate the File menu dynamically.
self.updateFileMenu()
......@@ -668,6 +645,8 @@ class MainWindow(QMainWindow, WindowMixin):
elif self.filePath:
self.queueEvent(partial(self.loadFile, self.filePath or ""))
self.keyDialog = None
# Callbacks:
self.zoomWidget.valueChanged.connect(self.paintCanvas)
......@@ -681,6 +660,12 @@ class MainWindow(QMainWindow, WindowMixin):
if self.filePath and os.path.isdir(self.filePath):
self.openDirDialog(dirpath=self.filePath, silent=True)
def menu(self, title, actions=None):
menu = self.menuBar().addMenu(title)
if actions:
addActions(menu, actions)
return menu
def keyReleaseEvent(self, event):
if event.key() == Qt.Key_Control:
self.canvas.setDrawingShapeToSquare(False)
......@@ -690,11 +675,9 @@ class MainWindow(QMainWindow, WindowMixin):
# Draw rectangle if Ctrl is pressed
self.canvas.setDrawingShapeToSquare(True)
def noShapes(self):
return not self.itemsToShapes
def populateModeActions(self):
self.canvas.menus[0].clear()
addActions(self.canvas.menus[0], self.actions.beginnerContext)
......@@ -702,7 +685,6 @@ class MainWindow(QMainWindow, WindowMixin):
actions = (self.actions.create,) # if self.beginner() else (self.actions.createMode, self.actions.editMode)
addActions(self.menus.edit, actions + self.actions.editMenu)
def setDirty(self):
self.dirty = True
self.actions.save.setEnabled(True)
......@@ -816,10 +798,11 @@ class MainWindow(QMainWindow, WindowMixin):
def rotateImgWarn(self):
if self.lang == 'ch':
self.msgBox.warning (self, "提示", "\n 该图片已经有标注框,旋转操作会打乱标注,建议清除标注框后旋转。")
self.msgBox.warning(self, "提示", "\n 该图片已经有标注框,旋转操作会打乱标注,建议清除标注框后旋转。")
else:
self.msgBox.warning (self, "Warn", "\n The picture already has a label box, and rotation will disrupt the label.\
It is recommended to clear the label box and rotate it.")
self.msgBox.warning(self, "Warn", "\n The picture already has a label box, "
"and rotation will disrupt the label. "
"It is recommended to clear the label box and rotate it.")
def rotateImgAction(self, k=1, _value=False):
......@@ -894,14 +877,13 @@ class MainWindow(QMainWindow, WindowMixin):
self.setDirty()
self.updateComboBox()
######## detection box related functions #######
# =================== detection box related functions ===================
def boxItemChanged(self, item):
shape = self.itemsToShapesbox[item]
box = ast.literal_eval(item.text())
# print('shape in labelItemChanged is',shape.points)
if box != [(p.x(), p.y()) for p in shape.points]:
if box != [(int(p.x()), int(p.y())) for p in shape.points]:
# shape.points = box
shape.points = [QPointF(p[0], p[1]) for p in box]
......@@ -909,7 +891,7 @@ class MainWindow(QMainWindow, WindowMixin):
# shape.line_color = generateColorByText(shape.label)
self.setDirty()
else: # User probably changed item visibility
self.canvas.setShapeVisible(shape, True)#item.checkState() == Qt.Checked
self.canvas.setShapeVisible(shape, True) # item.checkState() == Qt.Checked
def editBox(self): # ADD
if not self.canvas.editing():
......@@ -959,11 +941,10 @@ class MainWindow(QMainWindow, WindowMixin):
def indexTo5Files(self, currIndex):
if currIndex < 2:
return self.mImgList[:5]
elif currIndex > len(self.mImgList)-3:
elif currIndex > len(self.mImgList) - 3:
return self.mImgList[-5:]
else:
return self.mImgList[currIndex - 2 : currIndex + 3]
return self.mImgList[currIndex - 2: currIndex + 3]
# Tzutalin 20160906 : Add file list and dock to move faster
def fileitemDoubleClicked(self, item=None):
......@@ -983,9 +964,8 @@ class MainWindow(QMainWindow, WindowMixin):
self.loadFile(filename)
def CanvasSizeChange(self):
if len(self.mImgList) > 0:
self.zoomWidget.setValue(self.zoomWidgetValue + self.imgsplider.value())
if len(self.mImgList) > 0 and self.imageSlider.hasFocus():
self.zoomWidget.setValue(self.imageSlider.value())
def shapeSelectionChanged(self, selected_shapes):
self._noSelectionSlot = True
......@@ -1001,6 +981,12 @@ class MainWindow(QMainWindow, WindowMixin):
self.labelList.scrollToItem(self.currentItem()) # QAbstractItemView.EnsureVisible
self.BoxList.scrollToItem(self.currentBox())
if self.kie_mode:
if len(self.canvas.selectedShapes) == 1 and self.keyList.count() > 0:
selected_key_item_row = self.keyList.findItemsByLabel(self.canvas.selectedShapes[0].key_cls,
get_row=True)
self.keyList.setCurrentRow(selected_key_item_row)
self._noSelectionSlot = False
n_selected = len(selected_shapes)
self.actions.singleRere.setEnabled(n_selected)
......@@ -1008,6 +994,7 @@ class MainWindow(QMainWindow, WindowMixin):
self.actions.copy.setEnabled(n_selected)
self.actions.edit.setEnabled(n_selected == 1)
self.actions.lock.setEnabled(n_selected)
self.actions.change_cls.setEnabled(n_selected)
def addLabel(self, shape):
shape.paintLabel = self.displayLabelOption.isChecked()
......@@ -1030,6 +1017,10 @@ class MainWindow(QMainWindow, WindowMixin):
action.setEnabled(True)
self.updateComboBox()
# update show counting
self.BoxListDock.setWindowTitle(self.BoxListDockName + f" ({self.BoxList.count()})")
self.labelListDock.setWindowTitle(self.labelListDockName + f" ({self.labelList.count()})")
def remLabels(self, shapes):
if shapes is None:
# print('rm empty label')
......@@ -1050,8 +1041,8 @@ class MainWindow(QMainWindow, WindowMixin):
def loadLabels(self, shapes):
s = []
for label, points, line_color, fill_color, difficult in shapes:
shape = Shape(label=label,line_color=line_color)
for label, points, line_color, key_cls, difficult in shapes:
shape = Shape(label=label, line_color=line_color, key_cls=key_cls)
for x, y in points:
# Ensure the labels are within the bounds of the image. If not, fix them.
......@@ -1061,26 +1052,16 @@ class MainWindow(QMainWindow, WindowMixin):
shape.addPoint(QPointF(x, y))
shape.difficult = difficult
#shape.locked = False
# shape.locked = False
shape.close()
s.append(shape)
# if line_color:
# shape.line_color = QColor(*line_color)
# else:
# shape.line_color = generateColorByText(label)
#
# if fill_color:
# shape.fill_color = QColor(*fill_color)
# else:
# shape.fill_color = generateColorByText(label)
self._update_shape_color(shape)
self.addLabel(shape)
self.updateComboBox()
self.canvas.loadShapes(s)
def singleLabel(self, shape):
if shape is None:
# print('rm empty label')
......@@ -1115,14 +1096,16 @@ class MainWindow(QMainWindow, WindowMixin):
line_color=s.line_color.getRgb(),
fill_color=s.fill_color.getRgb(),
points=[(int(p.x()), int(p.y())) for p in s.points], # QPonitF
# add chris
difficult=s.difficult) # bool
difficult=s.difficult,
key_cls=s.key_cls) # bool
shapes = [] if mode == 'Auto' else \
[format_shape(shape) for shape in self.canvas.shapes if shape.line_color != DEFAULT_LOCK_COLOR]
if mode == 'Auto':
shapes = []
else:
shapes = [format_shape(shape) for shape in self.canvas.shapes if shape.line_color != DEFAULT_LOCK_COLOR]
# Can add differrent annotation formats here
for box in self.result_dic :
trans_dic = {"label": box[1][0], "points": box[0], 'difficult': False}
for box in self.result_dic:
trans_dic = {"label": box[1][0], "points": box[0], "difficult": False, "key_cls": "None"}
if trans_dic["label"] == "" and mode == 'Auto':
continue
shapes.append(trans_dic)
......@@ -1130,7 +1113,8 @@ class MainWindow(QMainWindow, WindowMixin):
try:
trans_dic = []
for box in shapes:
trans_dic.append({"transcription": box['label'], "points": box['points'], 'difficult': box['difficult']})
trans_dic.append({"transcription": box['label'], "points": box['points'],
"difficult": box['difficult'], "key_cls": box['key_cls']})
self.PPlabel[annotationFilePath] = trans_dic
if mode == 'Auto':
self.Cachelabel[annotationFilePath] = trans_dic
......@@ -1148,8 +1132,7 @@ class MainWindow(QMainWindow, WindowMixin):
for shape in self.canvas.copySelectedShape():
self.addLabel(shape)
# fix copy and delete
#self.shapeSelectionChanged(True)
# self.shapeSelectionChanged(True)
def labelSelectionChanged(self):
if self._noSelectionSlot:
......@@ -1163,10 +1146,9 @@ class MainWindow(QMainWindow, WindowMixin):
else:
self.canvas.deSelectShape()
def boxSelectionChanged(self):
if self._noSelectionSlot:
#self.BoxList.scrollToItem(self.currentBox(), QAbstractItemView.PositionAtCenter)
# self.BoxList.scrollToItem(self.currentBox(), QAbstractItemView.PositionAtCenter)
return
if self.canvas.editing():
selected_shapes = []
......@@ -1177,7 +1159,6 @@ class MainWindow(QMainWindow, WindowMixin):
else:
self.canvas.deSelectShape()
def labelItemChanged(self, item):
shape = self.itemsToShapes[item]
label = item.text()
......@@ -1185,7 +1166,7 @@ class MainWindow(QMainWindow, WindowMixin):
shape.label = item.text()
# shape.line_color = generateColorByText(shape.label)
self.setDirty()
elif not ((item.checkState()== Qt.Unchecked) ^ (not shape.difficult)):
elif not ((item.checkState() == Qt.Unchecked) ^ (not shape.difficult)):
shape.difficult = True if item.checkState() == Qt.Unchecked else False
self.setDirty()
else: # User probably changed item visibility
......@@ -1199,8 +1180,7 @@ class MainWindow(QMainWindow, WindowMixin):
position MUST be in global coordinates.
"""
if len(self.labelHist) > 0:
self.labelDialog = LabelDialog(
parent=self, listItem=self.labelHist)
self.labelDialog = LabelDialog(parent=self, listItem=self.labelHist)
if value:
text = self.labelDialog.popUp(text=self.prevLabelText)
......@@ -1210,8 +1190,22 @@ class MainWindow(QMainWindow, WindowMixin):
if text is not None:
self.prevLabelText = self.stringBundle.getString('tempLabel')
# generate_color = generateColorByText(text)
shape = self.canvas.setLastLabel(text, None, None)#generate_color, generate_color
shape = self.canvas.setLastLabel(text, None, None, None) # generate_color, generate_color
if self.kie_mode:
key_text, _ = self.keyDialog.popUp(self.key_previous_text)
if key_text is not None:
shape = self.canvas.setLastLabel(text, None, None, key_text) # generate_color, generate_color
self.key_previous_text = key_text
if not self.keyList.findItemsByLabel(key_text):
item = self.keyList.createItemFromLabel(key_text)
self.keyList.addItem(item)
rgb = self._get_rgb_by_label(key_text, self.kie_mode)
self.keyList.setItemLabel(item, key_text, rgb)
self._update_shape_color(shape)
self.keyDialog.addLabelHistory(key_text)
self.addLabel(shape)
if self.beginner(): # Switch to edit mode.
self.canvas.setEditing(True)
......@@ -1226,6 +1220,25 @@ class MainWindow(QMainWindow, WindowMixin):
# self.canvas.undoLastLine()
self.canvas.resetAllLines()
def _update_shape_color(self, shape):
r, g, b = self._get_rgb_by_label(shape.key_cls, self.kie_mode)
shape.line_color = QColor(r, g, b)
shape.vertex_fill_color = QColor(r, g, b)
shape.hvertex_fill_color = QColor(255, 255, 255)
shape.fill_color = QColor(r, g, b, 128)
shape.select_line_color = QColor(255, 255, 255)
shape.select_fill_color = QColor(r, g, b, 155)
def _get_rgb_by_label(self, label, kie_mode):
shift_auto_shape_color = 2 # use for random color
if kie_mode and label != "None":
item = self.keyList.findItemsByLabel(label)[0]
label_id = self.keyList.indexFromItem(item).row() + 1
label_id += shift_auto_shape_color
return LABEL_COLORMAP[label_id % len(LABEL_COLORMAP)]
else:
return (0, 255, 0)
def scrollRequest(self, delta, orientation):
units = - delta / (8 * 15)
bar = self.scrollBars[orientation]
......@@ -1239,6 +1252,7 @@ class MainWindow(QMainWindow, WindowMixin):
def addZoom(self, increment=10):
self.setZoom(self.zoomWidget.value() + increment)
self.imageSlider.setValue(self.zoomWidget.value() + increment) # set zoom slider value
def zoomRequest(self, delta):
# get the current scrollbar positions
......@@ -1331,7 +1345,6 @@ class MainWindow(QMainWindow, WindowMixin):
fileWidgetItem = self.fileListWidget.item(index)
print('unicodeFilePath is', unicodeFilePath)
fileWidgetItem.setSelected(True)
###
self.iconlist.clear()
self.additems5(None)
......@@ -1394,9 +1407,11 @@ class MainWindow(QMainWindow, WindowMixin):
# show file list image count
select_indexes = self.fileListWidget.selectedIndexes()
if len(select_indexes) > 0:
self.filedock.setWindowTitle(self.fileListName + f" ({select_indexes[0].row() + 1}"
self.fileDock.setWindowTitle(self.fileListName + f" ({select_indexes[0].row() + 1}"
f"/{self.fileListWidget.count()})")
# update show counting
self.BoxListDock.setWindowTitle(self.BoxListDockName + f" ({self.BoxList.count()})")
self.labelListDock.setWindowTitle(self.labelListDockName + f" ({self.labelList.count()})")
self.canvas.setFocus(True)
return True
......@@ -1405,24 +1420,23 @@ class MainWindow(QMainWindow, WindowMixin):
def showBoundingBoxFromPPlabel(self, filePath):
width, height = self.image.width(), self.image.height()
imgidx = self.getImglabelidx(filePath)
shapes =[]
#box['ratio'] of the shapes saved in lockedShapes contains the ratio of the
shapes = []
# box['ratio'] of the shapes saved in lockedShapes contains the ratio of the
# four corner coordinates of the shapes to the height and width of the image
for box in self.canvas.lockedShapes:
if self.canvas.isInTheSameImage:
shapes.append((box['transcription'], [[s[0]*width,s[1]*height]for s in box['ratio']],
DEFAULT_LOCK_COLOR, None, box['difficult']))
shapes.append((box['transcription'], [[s[0] * width, s[1] * height] for s in box['ratio']],
DEFAULT_LOCK_COLOR, box['key_cls'], box['difficult']))
else:
shapes.append(('锁定框：待检测', [[s[0]*width,s[1]*height]for s in box['ratio']],
DEFAULT_LOCK_COLOR, None, box['difficult']))
shapes.append(('锁定框：待检测', [[s[0] * width, s[1] * height] for s in box['ratio']],
DEFAULT_LOCK_COLOR, box['key_cls'], box['difficult']))
if imgidx in self.PPlabel.keys():
for box in self.PPlabel[imgidx]:
shapes.append((box['transcription'], box['points'], None, None, box['difficult']))
shapes.append((box['transcription'], box['points'], None, box['key_cls'], box['difficult']))
self.loadLabels(shapes)
self.canvas.verified = False
def validFilestate(self, filePath):
if filePath not in self.fileStatedict.keys():
return None
......@@ -1451,7 +1465,7 @@ class MainWindow(QMainWindow, WindowMixin):
"""Figure out the size of the pixmap in order to fit the main widget."""
e = 2.0 # So that no scrollbars are generated.
w1 = self.centralWidget().width() - e
h1 = self.centralWidget().height() - e -110
h1 = self.centralWidget().height() - e - 110
a1 = w1 / h1
# Calculate a new scale value based on the pixmap's aspect ratio.
w2 = self.canvas.pixmap.width() - 0.0
......@@ -1502,7 +1516,7 @@ class MainWindow(QMainWindow, WindowMixin):
def loadRecent(self, filename):
if self.mayContinue():
print(filename,"======")
print(filename, "======")
self.loadFile(filename)
def scanAllImages(self, folderPath):
......@@ -1517,8 +1531,6 @@ class MainWindow(QMainWindow, WindowMixin):
natural_sort(images, key=lambda x: x.lower())
return images
def openDirDialog(self, _value=False, dirpath=None, silent=False):
if not self.mayContinue():
return
......@@ -1538,7 +1550,7 @@ class MainWindow(QMainWindow, WindowMixin):
self.lastOpenDir = targetDirPath
self.importDirImages(targetDirPath)
def openDatasetDirDialog(self,):
def openDatasetDirDialog(self):
if self.lastOpenDir and os.path.exists(self.lastOpenDir):
if platform.system() == 'Windows':
os.startfile(self.lastOpenDir)
......@@ -1550,12 +1562,46 @@ class MainWindow(QMainWindow, WindowMixin):
if self.lang == 'ch':
self.msgBox.warning(self, "提示", "\n 原文件夹已不存在,请从新选择数据集路径!")
else:
self.msgBox.warning(self, "Warn", "\n The original folder no longer exists, please choose the data set path again!")
self.msgBox.warning(self, "Warn",
"\n The original folder no longer exists, please choose the data set path again!")
self.actions.open_dataset_dir.setEnabled(False)
defaultOpenDirPath = os.path.dirname(self.filePath) if self.filePath else '.'
def importDirImages(self, dirpath, isDelete = False):
def init_key_list(self, label_dict):
if not self.kie_mode:
return
# load key_cls
for image, info in label_dict.items():
for box in info:
if "key_cls" not in box:
continue
self.existed_key_cls_set.add(box["key_cls"])
if len(self.existed_key_cls_set) > 0:
for key_text in self.existed_key_cls_set:
if not self.keyList.findItemsByLabel(key_text):
item = self.keyList.createItemFromLabel(key_text)
self.keyList.addItem(item)
rgb = self._get_rgb_by_label(key_text, self.kie_mode)
self.keyList.setItemLabel(item, key_text, rgb)
if self.keyDialog is None:
# key list dialog
self.keyDialog = KeyDialog(
text=self.key_dialog_tip,
parent=self,
labels=self.existed_key_cls_set,
sort_labels=True,
show_text_field=True,
completion="startswith",
fit_to_content={'column': True, 'row': False},
flags=None
)
else:
self.keyDialog.labelList.addItems(self.existed_key_cls_set)
def importDirImages(self, dirpath, isDelete=False):
if not self.mayContinue() or not dirpath:
return
if self.defaultSaveDir and self.defaultSaveDir != dirpath:
......@@ -1563,16 +1609,18 @@ class MainWindow(QMainWindow, WindowMixin):
if not isDelete:
self.loadFilestate(dirpath)
self.PPlabelpath = dirpath+ '/Label.txt'
self.PPlabelpath = dirpath + '/Label.txt'
self.PPlabel = self.loadLabelFile(self.PPlabelpath)
self.Cachelabelpath = dirpath + '/Cache.cach'
self.Cachelabel = self.loadLabelFile(self.Cachelabelpath)
if self.Cachelabel:
self.PPlabel = dict(self.Cachelabel, **self.PPlabel)
self.init_key_list(self.PPlabel)
self.lastOpenDir = dirpath
self.dirname = dirpath
self.defaultSaveDir = dirpath
self.statusBar().showMessage('%s started. Annotation will be saved to %s' %
(__appname__, self.defaultSaveDir))
......@@ -1607,7 +1655,7 @@ class MainWindow(QMainWindow, WindowMixin):
self.actions.rotateRight.setEnabled(True)
self.fileListWidget.setCurrentRow(0) # set list index to first
self.filedock.setWindowTitle(self.fileListName + f" (1/{self.fileListWidget.count()})") # show image count
self.fileDock.setWindowTitle(self.fileListName + f" (1/{self.fileListWidget.count()})") # show image count
def openPrevImg(self, _value=False):
if len(self.mImgList) <= 0:
......@@ -1643,7 +1691,7 @@ class MainWindow(QMainWindow, WindowMixin):
else:
self.mImgList5 = self.indexTo5Files(currIndex)
if filename:
print('file name in openNext is ',filename)
print('file name in openNext is ', filename)
self.loadFile(filename)
def updateFileListIcon(self, filename):
......@@ -1655,30 +1703,6 @@ class MainWindow(QMainWindow, WindowMixin):
imgidx = self.getImglabelidx(self.filePath)
self._saveFile(imgidx, mode=mode)
def saveFileAs(self, _value=False):
assert not self.image.isNull(), "cannot save empty image"
self._saveFile(self.saveFileDialog())
def saveFileDialog(self, removeExt=True):
caption = '%s - Choose File' % __appname__
filters = 'File (*%s)' % LabelFile.suffix
openDialogPath = self.currentPath()
dlg = QFileDialog(self, caption, openDialogPath, filters)
dlg.setDefaultSuffix(LabelFile.suffix[1:])
dlg.setAcceptMode(QFileDialog.AcceptSave)
filenameWithoutExtension = os.path.splitext(self.filePath)[0]
dlg.selectFile(filenameWithoutExtension)
dlg.setOption(QFileDialog.DontUseNativeDialog, False)
if dlg.exec_():
fullFilePath = ustr(dlg.selectedFiles()[0])
if removeExt:
return os.path.splitext(fullFilePath)[0] # Return file path without the extension.
else:
return fullFilePath
return ''
def saveLockedShapes(self):
self.canvas.lockedShapes = []
self.canvas.selectedShapes = []
......@@ -1691,7 +1715,6 @@ class MainWindow(QMainWindow, WindowMixin):
self.canvas.selectedShapes.remove(s)
self.canvas.shapes.remove(s)
def _saveFile(self, annotationFilePath, mode='Manual'):
if len(self.canvas.lockedShapes) != 0:
self.saveLockedShapes()
......@@ -1701,9 +1724,9 @@ class MainWindow(QMainWindow, WindowMixin):
img = cv2.imread(self.filePath)
width, height = self.image.width(), self.image.height()
for shape in self.canvas.lockedShapes:
box = [[int(p[0]*width), int(p[1]*height)] for p in shape['ratio']]
box = [[int(p[0] * width), int(p[1] * height)] for p in shape['ratio']]
assert len(box) == 4
result = [(shape['transcription'],1)]
result = [(shape['transcription'], 1)]
result.insert(0, box)
self.result_dic_locked.append(result)
self.result_dic += self.result_dic_locked
......@@ -1717,7 +1740,7 @@ class MainWindow(QMainWindow, WindowMixin):
item.setIcon(newIcon('done'))
self.fileStatedict[self.filePath] = 1
if len(self.fileStatedict)%self.autoSaveNum ==0:
if len(self.fileStatedict) % self.autoSaveNum == 0:
self.saveFilestate()
self.savePPlabel(mode='Auto')
......@@ -1814,7 +1837,7 @@ class MainWindow(QMainWindow, WindowMixin):
def currentPath(self):
return os.path.dirname(self.filePath) if self.filePath else '.'
def chooseColor1(self):
def chooseColor(self):
color = self.colorDialog.getColor(self.lineColor, u'Choose line color',
default=DEFAULT_LINE_COLOR)
if color:
......@@ -1831,6 +1854,8 @@ class MainWindow(QMainWindow, WindowMixin):
if self.noShapes():
for action in self.actions.onShapesPresent:
action.setEnabled(False)
self.BoxListDock.setWindowTitle(self.BoxListDockName + f" ({self.BoxList.count()})")
self.labelListDock.setWindowTitle(self.labelListDockName + f" ({self.labelList.count()})")
def chshapeLineColor(self):
color = self.colorDialog.getColor(self.lineColor, u'Choose line color',
......@@ -1867,7 +1892,6 @@ class MainWindow(QMainWindow, WindowMixin):
else:
self.labelHist.append(line)
def togglePaintLabelsOption(self):
for shape in self.canvas.shapes:
shape.paintLabel = self.displayLabelOption.isChecked()
......@@ -1896,7 +1920,7 @@ class MainWindow(QMainWindow, WindowMixin):
prelen = lentoken // 2
bfilename = prelen * " " + pfilename + (lentoken - prelen) * " "
# item = QListWidgetItem(QIcon(pix.scaled(100, 100, Qt.KeepAspectRatio, Qt.SmoothTransformation)),filename[:10])
item = QListWidgetItem(QIcon(pix.scaled(100, 100, Qt.IgnoreAspectRatio, Qt.FastTransformation)),pfilename)
item = QListWidgetItem(QIcon(pix.scaled(100, 100, Qt.IgnoreAspectRatio, Qt.FastTransformation)), pfilename)
# item.setForeground(QBrush(Qt.white))
item.setToolTip(file)
self.iconlist.addItem(item)
......@@ -1908,7 +1932,7 @@ class MainWindow(QMainWindow, WindowMixin):
self.iconlist.setMinimumWidth(owidth + 50)
def getImglabelidx(self, filePath):
if platform.system()=='Windows':
if platform.system() == 'Windows':
spliter = '\\'
else:
spliter = '/'
......@@ -1930,6 +1954,7 @@ class MainWindow(QMainWindow, WindowMixin):
self.setDirty()
self.saveCacheLabel()
self.init_key_list(self.Cachelabel)
def reRecognition(self):
img = cv2.imread(self.filePath)
......@@ -1959,9 +1984,9 @@ class MainWindow(QMainWindow, WindowMixin):
print('Can not recognise the box')
if shape.line_color == DEFAULT_LOCK_COLOR:
shape.label = result[0][0]
self.result_dic_locked.append([box,(self.noLabelText,0)])
self.result_dic_locked.append([box, (self.noLabelText, 0)])
else:
self.result_dic.append([box,(self.noLabelText,0)])
self.result_dic.append([box, (self.noLabelText, 0)])
try:
if self.noLabelText == shape.label or result[1][0] == shape.label:
print('label no change')
......@@ -1969,13 +1994,16 @@ class MainWindow(QMainWindow, WindowMixin):
rec_flag += 1
except IndexError as e:
print('Can not recognise the box')
if (len(self.result_dic) > 0 and rec_flag > 0)or self.canvas.lockedShapes:
if (len(self.result_dic) > 0 and rec_flag > 0) or self.canvas.lockedShapes:
self.canvas.isInTheSameImage = True
self.saveFile(mode='Auto')
self.loadFile(self.filePath)
self.canvas.isInTheSameImage = False
self.setDirty()
elif len(self.result_dic) == len(self.canvas.shapes) and rec_flag == 0:
if self.lang == 'ch':
QMessageBox.information(self, "Information", "识别结果保持一致！")
else:
QMessageBox.information(self, "Information", "The recognition result remains unchanged!")
else:
print('Can not recgonise in ', self.filePath)
......@@ -2041,7 +2069,6 @@ class MainWindow(QMainWindow, WindowMixin):
self.AutoRecognition.setEnabled(True)
self.actions.AutoRec.setEnabled(True)
def modelChoose(self):
print(self.comboBox.currentText())
lg_idx = {'Chinese & English': 'ch', 'English': 'en', 'French': 'french', 'German': 'german',
......@@ -2068,14 +2095,12 @@ class MainWindow(QMainWindow, WindowMixin):
self.actions.saveLabel.setEnabled(True)
self.actions.saveRec.setEnabled(True)
def saveFilestate(self):
with open(self.fileStatepath, 'w', encoding='utf-8') as f:
for key in self.fileStatedict:
f.write(key + '\t')
f.write(str(self.fileStatedict[key]) + '\n')
def loadLabelFile(self, labelpath):
labeldict = {}
if not os.path.exists(labelpath):
......@@ -2094,8 +2119,7 @@ class MainWindow(QMainWindow, WindowMixin):
labeldict[file] = []
return labeldict
def savePPlabel(self,mode='Manual'):
def savePPlabel(self, mode='Manual'):
savedfile = [self.getImglabelidx(i) for i in self.fileStatedict.keys()]
with open(self.PPlabelpath, 'w', encoding='utf-8') as f:
for key in self.PPlabel:
......@@ -2137,19 +2161,22 @@ class MainWindow(QMainWindow, WindowMixin):
try:
img = cv2.imread(key)
for i, label in enumerate(self.PPlabel[idx]):
if label['difficult']: continue
if label['difficult']:
continue
img_crop = get_rotate_crop_image(img, np.array(label['points'], np.float32))
img_name = os.path.splitext(os.path.basename(idx))[0] + '_crop_'+str(i)+'.jpg'
cv2.imwrite(crop_img_dir+img_name, img_crop)
f.write('crop_img/'+ img_name + '\t')
img_name = os.path.splitext(os.path.basename(idx))[0] + '_crop_' + str(i) + '.jpg'
cv2.imwrite(crop_img_dir + img_name, img_crop)
f.write('crop_img/' + img_name + '\t')
f.write(label['transcription'] + '\n')
except Exception as e:
ques_img.append(key)
print("Can not read image ",e)
print("Can not read image ", e)
if ques_img:
QMessageBox.information(self, "Information", "The following images can not be saved, "
"please check the image path and labels.\n" + "".join(str(i)+'\n' for i in ques_img))
QMessageBox.information(self, "Information", "Cropped images have been saved in "+str(crop_img_dir))
QMessageBox.information(self,
"Information",
"The following images can not be saved, please check the image path and labels.\n"
+ "".join(str(i) + '\n' for i in ques_img))
QMessageBox.information(self, "Information", "Cropped images have been saved in " + str(crop_img_dir))
def speedChoose(self):
if self.labelDialogOption.isChecked():
......@@ -2172,6 +2199,15 @@ class MainWindow(QMainWindow, WindowMixin):
self.autoSaveNum = 5 # Used for backup
print('The program will automatically save once after confirming 5 images (default)')
def change_box_key(self):
key_text, _ = self.keyDialog.popUp(self.key_previous_text)
if key_text is None:
return
self.key_previous_text = key_text
for shape in self.canvas.selectedShapes:
shape.key_cls = key_text
self._update_shape_color(shape)
def undoShapeEdit(self):
self.canvas.restoreShape()
self.labelList.clear()
......@@ -2186,25 +2222,27 @@ class MainWindow(QMainWindow, WindowMixin):
self.labelList.clearSelection()
self._noSelectionSlot = False
self.canvas.loadShapes(shapes, replace=replace)
print("loadShapes")#1
print("loadShapes") # 1
def lockSelectedShape(self):
"""lock the selsected shapes.
"""lock the selected shapes.
Add self.selectedShapes to lock self.canvas.lockedShapes,
which holds the ratio of the four coordinates of the locked shapes
to the width and height of the image
"""
width, height = self.image.width(), self.image.height()
def format_shape(s):
return dict(label=s.label, # str
line_color=s.line_color.getRgb(),
fill_color=s.fill_color.getRgb(),
ratio=[[int(p.x())/width, int(p.y())/height] for p in s.points], # QPonitF
# add chris
difficult=s.difficult) # bool
#lock
ratio=[[int(p.x()) / width, int(p.y()) / height] for p in s.points], # QPonitF
difficult=s.difficult, # bool
key_cls=s.key_cls, # bool
)
# lock
if len(self.canvas.lockedShapes) == 0:
for s in self.canvas.selectedShapes:
s.line_color = DEFAULT_LOCK_COLOR
......@@ -2212,11 +2250,13 @@ class MainWindow(QMainWindow, WindowMixin):
shapes = [format_shape(shape) for shape in self.canvas.selectedShapes]
trans_dic = []
for box in shapes:
trans_dic.append({"transcription": box['label'], "ratio": box['ratio'], 'difficult': box['difficult']})
trans_dic.append({"transcription": box['label'], "ratio": box['ratio'],
"difficult": box['difficult'],
"key_cls": "None" if "key_cls" not in box else box["key_cls"]})
self.canvas.lockedShapes = trans_dic
self.actions.save.setEnabled(True)
#unlock
# unlock
else:
for s in self.canvas.shapes:
s.line_color = DEFAULT_LINE_COLOR
......@@ -2237,9 +2277,11 @@ def read(filename, default=None):
except:
return default
def str2bool(v):
return v.lower() in ("true", "t", "1")
def get_main_app(argv=[]):
"""
Standard boilerplate Qt application code.
......@@ -2248,23 +2290,26 @@ def get_main_app(argv=[]):
app = QApplication(argv)
app.setApplicationName(__appname__)
app.setWindowIcon(newIcon("app"))
# Tzutalin 201705+: Accept extra agruments to change predefined class file
argparser = argparse.ArgumentParser()
argparser.add_argument("--lang", type=str, default='en', nargs="?")
argparser.add_argument("--gpu", type=str2bool, default=False, nargs="?")
argparser.add_argument("--predefined_classes_file",
# Tzutalin 201705+: Accept extra arguments to change predefined class file
arg_parser = argparse.ArgumentParser()
arg_parser.add_argument("--lang", type=str, default='en', nargs="?")
arg_parser.add_argument("--gpu", type=str2bool, default=True, nargs="?")
arg_parser.add_argument("--kie", type=str2bool, default=False, nargs="?")
arg_parser.add_argument("--predefined_classes_file",
default=os.path.join(os.path.dirname(__file__), "data", "predefined_classes.txt"),
nargs="?")
args = argparser.parse_args(argv[1:])
# Usage : labelImg.py image predefClassFile saveDir
win = MainWindow(lang=args.lang, gpu=args.gpu,
defaultPrefdefClassFile=args.predefined_classes_file)
args = arg_parser.parse_args(argv[1:])
win = MainWindow(lang=args.lang,
gpu=args.gpu,
kie_mode=args.kie,
default_predefined_class_file=args.predefined_classes_file)
win.show()
return app, win
def main():
'''construct main app and run it'''
"""construct main app and run it"""
app, _win = get_main_app(sys.argv)
return app.exec_()
......@@ -2276,5 +2321,5 @@ if __name__ == '__main__':
output = os.system('pyrcc5 -o libs/resources.py resources.qrc')
assert output == 0, "operate the cmd have some problems ,please check whether there is a in the lib " \
"directory resources.py "
import libs.resources
sys.exit(main())
PPOCRLabel/README.md
浏览文件 @ 4b214948
......@@ -8,6 +8,10 @@ PPOCRLabel is a semi-automatic graphic annotation tool suitable for OCR field, w
### Recent Update
- 2022.02：（by [PeterH0323](https://github.com/peterh0323) ）
- Added KIE mode, for [detection + identification + keyword extraction] labeling.
- 2022.01：（by [PeterH0323](https://github.com/peterh0323) ）
- Improve user experience: prompt for the number of files and labels, optimize interaction, and fix bugs such as only use CPU when inference
- 2021.11.17:
- Support install and start PPOCRLabel through the whl package (by [d2623587501](https://github.com/d2623587501))
- Dataset segmentation: Divide the annotation file into training, verification and testing parts (refer to section 3.5 below, by [MrCuiHao](https://github.com/MrCuiHao))
......@@ -70,7 +74,8 @@ PPOCRLabel
```bash
pip3 install PPOCRLabel
pip3 install opencv-contrib-python-headless==4.2.0.32
PPOCRLabel # run
PPOCRLabel # [Normal mode] for [detection + recognition] labeling
PPOCRLabel --kie True # [KIE mode] for [detection + recognition + keyword extraction] labeling
```
#### 1.2.2 Build and Install the Whl Package Locally
......@@ -85,7 +90,8 @@ pip3 install dist/PPOCRLabel-1.0.2-py2.py3-none-any.whl
```bash
cd ./PPOCRLabel # Switch to the PPOCRLabel directory
python PPOCRLabel.py
python PPOCRLabel.py # [Normal mode] for [detection + recognition] labeling
python PPOCRLabel.py --kie True # [KIE mode] for [detection + recognition + keyword extraction] labeling
```
......@@ -110,7 +116,7 @@ python PPOCRLabel.py
6. Click 're-Recognition', model will rewrite ALL recognition results in ALL detection box<sup>[3]</sup>.
7. Double click the result in 'recognition result' list to manually change inaccurate recognition results.
7. Single click the result in 'recognition result' list to manually change inaccurate recognition results.
8. **Click "Check", the image status will switch to "√",then the program automatically jump to the next.**
......@@ -143,15 +149,17 @@ python PPOCRLabel.py
### 3.1 Shortcut keys
| Shortcut keys | Description |
|--------------------------| ------------------------------------------------ |
|--------------------------|--------------------------------------------------|
| Ctrl + Shift + R | Re-recognize all the labels of the current image |
| W | Create a rect box |
| Q | Create a four-points box |
| X | Rotate the box anti-clockwise |
| C | Rotate the box clockwise |
| Ctrl + E | Edit label of the selected box |
| Ctrl + R | Re-recognize the selected box |
| Ctrl + C | Copy and paste the selected box |
| Ctrl + Left Mouse Button | Multi select the label box |
| Ctrl + X | Delete the selected box |
| Alt + X | Delete the selected box |
| Ctrl + V | Check image |
| Ctrl + Shift + d | Delete image |
| D | Next image |
......@@ -167,7 +175,7 @@ python PPOCRLabel.py
- Model language switching: Changing the built-in model language is supportable by clicking "PaddleOCR"-"Choose OCR Model" in the menu bar. Currently supported languages​include French, German, Korean, and Japanese.
For specific model download links, please refer to [PaddleOCR Model List](https://github.com/PaddlePaddle/PaddleOCR/blob/develop/doc/doc_en/models_list_en.md#multilingual-recognition-modelupdating)
- **Custom Model**: If users want to replace the built-in model with their own inference model, they can follow the [Custom Model Code Usage](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/doc/doc_en/whl_en.md#31-use-by-code) by modifying PPOCRLabel.py for [Instantiation of PaddleOCR class](https://github.com/PaddlePaddle/PaddleOCR/blob/release/ 2.3/PPOCRLabel/PPOCRLabel.py#L116) :
- **Custom Model**: If users want to replace the built-in model with their own inference model, they can follow the [Custom Model Code Usage](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/doc/doc_en/whl_en.md#31-use-by-code) by modifying PPOCRLabel.py for [Instantiation of PaddleOCR class](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/PPOCRLabel/PPOCRLabel.py#L86) :
add parameter `det_model_dir` in `self.ocr = PaddleOCR(use_pdserving=False, use_angle_cls=True, det=True, cls=True, use_gpu=gpu, lang=lang) `
......@@ -196,18 +204,28 @@ For some data that are difficult to recognize, the recognition results will not
```
cd ./PPOCRLabel # Change the directory to the PPOCRLabel folder
python gen_ocr_train_val_test.py --trainValTestRatio 6:2:2 --labelRootPath ../train_data/label --detRootPath ../train_data/det --recRootPath ../train_data/rec
python gen_ocr_train_val_test.py --trainValTestRatio 6:2:2 --datasetRootPath ../train_data
```
Parameter Description:
- `trainValTestRatio` is the division ratio of the number of images in the training set, validation set, and test set, set according to your actual situation, the default is `6:2:2`
- `labelRootPath` is the storage path of the dataset labeled by PPOCRLabel, the default is `../train_data/label`
- `detRootPath` is the path where the text detection dataset is divided according to the dataset marked by PPOCRLabel. The default is `../train_data/det`
- `recRootPath` is the path where the character recognition dataset is divided according to the dataset marked by PPOCRLabel. The default is `../train_data/rec`
- `datasetRootPath` is the storage path of the complete dataset labeled by PPOCRLabel. The default path is `PaddleOCR/train_data` .
```
|-train_data
|-crop_img
|- word_001_crop_0.png
|- word_002_crop_0.jpg
|- word_003_crop_0.jpg
| ...
| Label.txt
| rec_gt.txt
|- word_001.png
|- word_002.jpg
|- word_003.jpg
| ...
```
### 3.6 Error message
......
PPOCRLabel/README_ch.md
浏览文件 @ 4b214948
......@@ -8,6 +8,10 @@ PPOCRLabel是一款适用于OCR领域的半自动化图形标注工具，内置P
#### 近期更新
- 2022.02：（by [PeterH0323](https://github.com/peterh0323) ）
- 新增：KIE 功能，用于打【检测+识别+关键字提取】的标签
- 2022.01：（by [PeterH0323](https://github.com/peterh0323) ）
- 提升用户体验：新增文件与标记数目提示、优化交互、修复gpu使用等问题
- 2021.11.17：
- 新增支持通过whl包安装和启动PPOCRLabel（by [d2623587501](https://github.com/d2623587501)）
- 标注数据集切分：对标注数据进行训练、验证与测试集划分（参考下方3.5节，by [MrCuiHao](https://github.com/MrCuiHao)）
......@@ -24,7 +28,7 @@ PPOCRLabel是一款适用于OCR领域的半自动化图形标注工具，内置P
- 识别结果更改为单击修改。（如果无法修改，请切换为系统自带输入法，或再次切回原输入法）
- 2020.12.18： 支持对单个标记框进行重新识别（by [ninetailskim](https://github.com/ninetailskim)），完善快捷键。
如果您对以上内容感兴趣或对完善工具有不一样的想法，欢迎加入我们的SIG队伍与我们共同开发。可以在[此处](https://github.com/PaddlePaddle/PaddleOCR/issues/1728)完成问卷和前置任务，经过我们确认相关内容后即可正式加入，享受SIG福利，共同为OCR开源事业贡献（特别说明：针对PPOCRLabel的改进也属于PaddleOCR前置任务）
如果您对完善工具有不一样的想法，欢迎通过[社区常规赛](https://github.com/PaddlePaddle/PaddleOCR/issues/4982)报名相关更改，获得积分兑换奖励。
......@@ -68,7 +72,8 @@ PPOCRLabel --lang ch
```bash
pip3 install PPOCRLabel
pip3 install opencv-contrib-python-headless==4.2.0.32 # 如果下载过慢请添加"-i https://mirror.baidu.com/pypi/simple"
PPOCRLabel --lang ch # 启动
PPOCRLabel --lang ch # 启动【普通模式】，用于打【检测+识别】场景的标签
PPOCRLabel --lang ch --kie True # 启动 【KIE 模式】，用于打【检测+识别+关键字提取】场景的标签
```
> 如果上述安装出现问题，可以参考3.6节 错误提示
......@@ -87,7 +92,8 @@ pip3 install dist/PPOCRLabel-1.0.2-py2.py3-none-any.whl -i https://mirror.baidu.
```bash
cd ./PPOCRLabel # 切换到PPOCRLabel目录
python PPOCRLabel.py --lang ch
python PPOCRLabel.py --lang ch # 启动【普通模式】，用于打【检测+识别】场景的标签
python PPOCRLabel.py --lang ch --kie True # 启动 【KIE 模式】，用于打【检测+识别+关键字提取】场景的标签
```
......@@ -102,7 +108,7 @@ python PPOCRLabel.py --lang ch
4. 手动标注：点击 “矩形标注”（推荐直接在英文模式下点击键盘中的 “W”)，用户可对当前图片中模型未检出的部分进行手动绘制标记框。点击键盘Q，则使用四点标注模式（或点击“编辑” - “四点标注”），用户依次点击4个点后，双击左键表示标注完成。
5. 标记框绘制完成后，用户点击 “确认”，检测框会先被预分配一个 “待识别” 标签。
6. 重新识别：将图片中的所有检测画绘制/调整完成后，点击 “重新识别”，PPOCR模型会对当前图片中的**所有检测框**重新识别<sup>[3]</sup>。
7. 内容更改：双击识别结果，对不准确的识别结果进行手动更改。
7. 内容更改：单击识别结果，对不准确的识别结果进行手动更改。
8. **确认标记：点击 “确认”，图片状态切换为 “√”，跳转至下一张。**
9. 删除：点击 “删除图像”，图片将会被删除至回收站。
10. 导出结果：用户可以通过菜单中“文件-导出标记结果”手动导出，同时也可以点击“文件 - 自动导出标记结果”开启自动导出。手动确认过的标记将会被存放在所打开图片文件夹下的*Label.txt*中。在菜单栏点击 “文件” - "导出识别结果"后，会将此类图片的识别训练数据保存在*crop_img*文件夹下，识别标签保存在*rec_gt.txt*中<sup>[4]</sup>。
......@@ -132,15 +138,17 @@ python PPOCRLabel.py --lang ch
### 3.1 快捷键
| 快捷键 | 说明 |
|------------------| ---------------------------- |
|------------------|----------------|
| Ctrl + shift + R | 对当前图片的所有标记重新识别 |
| W | 新建矩形框 |
| Q | 新建四点框 |
| X | 框逆时针旋转 |
| C | 框顺时针旋转 |
| Ctrl + E | 编辑所选框标签 |
| Ctrl + R | 重新识别所选标记 |
| Ctrl + C | 复制并粘贴选中的标记框 |
| Ctrl + 鼠标左键 | 多选标记框 |
| Ctrl + X | 删除所选框 |
| Alt + X | 删除所选框 |
| Ctrl + V | 确认本张图片标记 |
| Ctrl + Shift + d | 删除本张图片 |
| D | 下一张图片 |
......@@ -181,18 +189,28 @@ PPOCRLabel支持三种导出方式：
```
cd ./PPOCRLabel # 将目录切换到PPOCRLabel文件夹下
python gen_ocr_train_val_test.py --trainValTestRatio 6:2:2 --labelRootPath ../train_data/label --detRootPath ../train_data/det --recRootPath ../train_data/rec
python gen_ocr_train_val_test.py --trainValTestRatio 6:2:2 --datasetRootPath ../train_data
```
参数说明：
- `trainValTestRatio` 是训练集、验证集、测试集的图像数量划分比例，根据实际情况设定，默认是`6:2:2`
- `labelRootPath` 是PPOCRLabel标注的数据集存放路径，默认是`../train_data/label`
- `detRootPath` 是根据PPOCRLabel标注的数据集划分后的文本检测数据集存放的路径，默认是`../train_data/det `
- `recRootPath` 是根据PPOCRLabel标注的数据集划分后的字符识别数据集存放的路径，默认是`../train_data/rec`
- `datasetRootPath` 是PPOCRLabel标注的完整数据集存放路径。默认路径是 `PaddleOCR/train_data` 分割数据集前应有如下结构：
```
|-train_data
|-crop_img
|- word_001_crop_0.png
|- word_002_crop_0.jpg
|- word_003_crop_0.jpg
| ...
| Label.txt
| rec_gt.txt
|- word_001.png
|- word_002.jpg
|- word_003.jpg
| ...
```
### 3.6 错误提示
......
PPOCRLabel/combobox.py 已删除 100644 → 0
浏览文件 @ 917606b4
# Copyright (c) <2015-Present> Tzutalin
# Copyright (C) 2013 MIT, Computer Science and Artificial Intelligence Laboratory. Bryan Russell, Antonio Torralba,
# William T. Freeman. Permission is hereby granted, free of charge, to any person obtaining a copy of this software and
# associated documentation files (the "Software"), to deal in the Software without restriction, including without
# limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the
# Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
# The above copyright notice and this permission notice shall be included in all copies or substantial portions of
# the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT
# NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT
# SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF
# CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
# THE SOFTWARE.
import sys
try:
from PyQt5.QtWidgets import QWidget, QHBoxLayout, QComboBox
except ImportError:
# needed for py3+qt4
# Ref:
# http://pyqt.sourceforge.net/Docs/PyQt4/incompatible_apis.html
# http://stackoverflow.com/questions/21217399/pyqt4-qtcore-qvariant-object-instead-of-a-string
if sys.version_info.major >= 3:
import sip
sip.setapi('QVariant', 2)
from PyQt4.QtGui import QWidget, QHBoxLayout, QComboBox
class ComboBox(QWidget):
def __init__(self, parent=None, items=[]):
super(ComboBox, self).__init__(parent)
layout = QHBoxLayout()
self.cb = QComboBox()
self.items = items
self.cb.addItems(self.items)
self.cb.currentIndexChanged.connect(parent.comboSelectionChanged)
layout.addWidget(self.cb)
self.setLayout(layout)
def update_items(self, items):
self.items = items
self.cb.clear()
self.cb.addItems(self.items)
PPOCRLabel/gen_ocr_train_val_test.py
浏览文件 @ 4b214948
......@@ -17,15 +17,14 @@ def isCreateOrDeleteFolder(path, flag):
return flagAbsPath
def splitTrainVal(root, dir, absTrainRootPath, absValRootPath, absTestRootPath, trainTxt, valTxt, testTxt, flag):
def splitTrainVal(root, absTrainRootPath, absValRootPath, absTestRootPath, trainTxt, valTxt, testTxt, flag):
# 按照指定的比例划分训练集、验证集、测试集
labelPath = os.path.join(root, dir)
labelAbsPath = os.path.abspath(labelPath)
dataAbsPath = os.path.abspath(root)
if flag == "det":
labelFilePath = os.path.join(labelAbsPath, args.detLabelFileName)
labelFilePath = os.path.join(dataAbsPath, args.detLabelFileName)
elif flag == "rec":
labelFilePath = os.path.join(labelAbsPath, args.recLabelFileName)
labelFilePath = os.path.join(dataAbsPath, args.recLabelFileName)
labelFileRead = open(labelFilePath, "r", encoding="UTF-8")
labelFileContent = labelFileRead.readlines()
......@@ -38,9 +37,9 @@ def splitTrainVal(root, dir, absTrainRootPath, absValRootPath, absTestRootPath,
imageName = os.path.basename(imageRelativePath)
if flag == "det":
imagePath = os.path.join(labelAbsPath, imageName)
imagePath = os.path.join(dataAbsPath, imageName)
elif flag == "rec":
imagePath = os.path.join(labelAbsPath, "{}\\{}".format(args.recImageDirName, imageName))
imagePath = os.path.join(dataAbsPath, "{}\\{}".format(args.recImageDirName, imageName))
# 按预设的比例划分训练集、验证集、测试集
trainValTestRatio = args.trainValTestRatio.split(":")
......@@ -90,15 +89,20 @@ def genDetRecTrainVal(args):
recValTxt = open(os.path.join(args.recRootPath, "val.txt"), "a", encoding="UTF-8")
recTestTxt = open(os.path.join(args.recRootPath, "test.txt"), "a", encoding="UTF-8")
for root, dirs, files in os.walk(args.labelRootPath):
for dir in dirs:
splitTrainVal(root, dir, detAbsTrainRootPath, detAbsValRootPath, detAbsTestRootPath, detTrainTxt, detValTxt,
splitTrainVal(args.datasetRootPath, detAbsTrainRootPath, detAbsValRootPath, detAbsTestRootPath, detTrainTxt, detValTxt,
detTestTxt, "det")
splitTrainVal(root, dir, recAbsTrainRootPath, recAbsValRootPath, recAbsTestRootPath, recTrainTxt, recValTxt,
for root, dirs, files in os.walk(args.datasetRootPath):
for dir in dirs:
if dir == 'crop_img':
splitTrainVal(root, recAbsTrainRootPath, recAbsValRootPath, recAbsTestRootPath, recTrainTxt, recValTxt,
recTestTxt, "rec")
else:
continue
break
if __name__ == "__main__":
# 功能描述：分别划分检测和识别的训练集、验证集、测试集
# 说明：可以根据自己的路径和需求调整参数，图像数据往往多人合作分批标注，每一批图像数据放在一个文件夹内用PPOCRLabel进行标注，
......@@ -110,9 +114,9 @@ if __name__ == "__main__":
default="6:2:2",
help="ratio of trainset:valset:testset")
parser.add_argument(
"--labelRootPath",
"--datasetRootPath",
type=str,
default="../train_data/label",
default="../train_data/",
help="path to the dataset marked by ppocrlabel, E.g, dataset folder named 1,2,3..."
)
parser.add_argument(
......
PPOCRLabel/libs/canvas.py
浏览文件 @ 4b214948
......@@ -11,19 +11,13 @@
# CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
# THE SOFTWARE.
try:
from PyQt5.QtGui import *
from PyQt5.QtCore import *
from PyQt5.QtWidgets import *
except ImportError:
from PyQt4.QtGui import *
from PyQt4.QtCore import *
#from PyQt4.QtOpenGL import *
import copy
from PyQt5.QtCore import Qt, pyqtSignal, QPointF, QPoint
from PyQt5.QtGui import QPainter, QBrush, QColor, QPixmap
from PyQt5.QtWidgets import QWidget, QMenu, QApplication
from libs.shape import Shape
from libs.utils import distance
import copy
CURSOR_DEFAULT = Qt.ArrowCursor
CURSOR_POINT = Qt.PointingHandCursor
......@@ -31,8 +25,6 @@ CURSOR_DRAW = Qt.CrossCursor
CURSOR_MOVE = Qt.ClosedHandCursor
CURSOR_GRAB = Qt.OpenHandCursor
# class Canvas(QGLWidget):
class Canvas(QWidget):
zoomRequest = pyqtSignal(int)
......@@ -129,7 +121,6 @@ class Canvas(QWidget):
def selectedVertex(self):
return self.hVertex is not None
def mouseMoveEvent(self, ev):
"""Update line with last point and current coordinates."""
pos = self.transformPos(ev.pos())
......@@ -333,7 +324,6 @@ class Canvas(QWidget):
self.movingShape = False
def endMove(self, copy=False):
assert self.selectedShapes and self.selectedShapesCopy
assert len(self.selectedShapesCopy) == len(self.selectedShapes)
......@@ -410,7 +400,6 @@ class Canvas(QWidget):
self.selectionChanged.emit(shapes)
self.update()
def selectShapePoint(self, point, multiple_selection_mode):
"""Select the first shape created which contains this point."""
if self.selectedVertex(): # A vertex is marked for selection.
......@@ -494,7 +483,6 @@ class Canvas(QWidget):
else:
shape.moveVertexBy(index, shiftPos)
def boundedMoveShape(self, shapes, pos):
if type(shapes).__name__ != 'list': shapes = [shapes]
if self.outOfPixmap(pos):
......@@ -515,6 +503,7 @@ class Canvas(QWidget):
if dp:
for shape in shapes:
shape.moveBy(dp)
shape.close()
self.prevPoint = pos
return True
return False
......@@ -728,6 +717,31 @@ class Canvas(QWidget):
self.moveOnePixel('Up')
elif key == Qt.Key_Down and self.selectedShapes:
self.moveOnePixel('Down')
elif key == Qt.Key_X and self.selectedShapes:
for i in range(len(self.selectedShapes)):
self.selectedShape = self.selectedShapes[i]
if self.rotateOutOfBound(0.01):
continue
self.selectedShape.rotate(0.01)
self.shapeMoved.emit()
self.update()
elif key == Qt.Key_C and self.selectedShapes:
for i in range(len(self.selectedShapes)):
self.selectedShape = self.selectedShapes[i]
if self.rotateOutOfBound(-0.01):
continue
self.selectedShape.rotate(-0.01)
self.shapeMoved.emit()
self.update()
def rotateOutOfBound(self, angle):
for shape in range(len(self.selectedShapes)):
self.selectedShape = self.selectedShapes[shape]
for i, p in enumerate(self.selectedShape.points):
if self.outOfPixmap(self.selectedShape.rotatePoint(p, angle)):
return True
return False
def moveOnePixel(self, direction):
# print(self.selectedShape.points)
......@@ -769,7 +783,7 @@ class Canvas(QWidget):
points = [p1+p2 for p1, p2 in zip(self.selectedShape.points, [step]*4)]
return True in map(self.outOfPixmap, points)
def setLastLabel(self, text, line_color = None, fill_color = None):
def setLastLabel(self, text, line_color=None, fill_color=None, key_cls=None):
assert text
self.shapes[-1].label = text
if line_color:
......@@ -777,6 +791,10 @@ class Canvas(QWidget):
if fill_color:
self.shapes[-1].fill_color = fill_color
if key_cls:
self.shapes[-1].key_cls = key_cls
self.storeShapes()
return self.shapes[-1]
......
PPOCRLabel/libs/editinlist.py
浏览文件 @ 4b214948
import sys, time
from PyQt5 import QtWidgets
from PyQt5.QtGui import *
from PyQt5.QtCore import *
from PyQt5.QtWidgets import *
# !/usr/bin/env python
# -*- coding: utf-8 -*-
from PyQt5.QtCore import QModelIndex
from PyQt5.QtWidgets import QListWidget
class EditInList(QListWidget):
def __init__(self):
super(EditInList,self).__init__()
# click to edit
self.clicked.connect(self.item_clicked)
super(EditInList, self).__init__()
self.edited_item = None
def item_clicked(self, modelindex: QModelIndex) -> None:
self.edited_item = self.currentItem()
def item_clicked(self, modelindex: QModelIndex):
try:
if self.edited_item is not None:
self.closePersistentEditor(self.edited_item)
item = self.item(modelindex.row())
# time.sleep(0.2)
self.edited_item = item
self.openPersistentEditor(item)
# time.sleep(0.2)
self.editItem(item)
except:
self.edited_item = self.currentItem()
self.edited_item = self.item(modelindex.row())
self.openPersistentEditor(self.edited_item)
self.editItem(self.edited_item)
def mouseDoubleClickEvent(self, event):
# close edit
for i in range(self.count()):
self.closePersistentEditor(self.item(i))
pass
def leaveEvent(self, event):
# close edit
......
PPOCRLabel/libs/keyDialog.py 0 → 100644
浏览文件 @ 4b214948
import re
from PyQt5 import QtCore
from PyQt5 import QtGui
from PyQt5 import QtWidgets
from PyQt5.Qt import QT_VERSION_STR
from libs.utils import newIcon, labelValidator
QT5 = QT_VERSION_STR[0] == '5'
# TODO(unknown):
# - Calculate optimal position so as not to go out of screen area.
class KeyQLineEdit(QtWidgets.QLineEdit):
def setListWidget(self, list_widget):
self.list_widget = list_widget
def keyPressEvent(self, e):
if e.key() in [QtCore.Qt.Key_Up, QtCore.Qt.Key_Down]:
self.list_widget.keyPressEvent(e)
else:
super(KeyQLineEdit, self).keyPressEvent(e)
class KeyDialog(QtWidgets.QDialog):
def __init__(
self,
text="Enter object label",
parent=None,
labels=None,
sort_labels=True,
show_text_field=True,
completion="startswith",
fit_to_content=None,
flags=None,
):
if fit_to_content is None:
fit_to_content = {"row": False, "column": True}
self._fit_to_content = fit_to_content
super(KeyDialog, self).__init__(parent)
self.edit = KeyQLineEdit()
self.edit.setPlaceholderText(text)
self.edit.setValidator(labelValidator())
self.edit.editingFinished.connect(self.postProcess)
if flags:
self.edit.textChanged.connect(self.updateFlags)
layout = QtWidgets.QVBoxLayout()
if show_text_field:
layout_edit = QtWidgets.QHBoxLayout()
layout_edit.addWidget(self.edit, 6)
layout.addLayout(layout_edit)
# buttons
self.buttonBox = bb = QtWidgets.QDialogButtonBox(
QtWidgets.QDialogButtonBox.Ok | QtWidgets.QDialogButtonBox.Cancel,
QtCore.Qt.Horizontal,
self,
)
bb.button(bb.Ok).setIcon(newIcon("done"))
bb.button(bb.Cancel).setIcon(newIcon("undo"))
bb.accepted.connect(self.validate)
bb.rejected.connect(self.reject)
layout.addWidget(bb)
# label_list
self.labelList = QtWidgets.QListWidget()
if self._fit_to_content["row"]:
self.labelList.setHorizontalScrollBarPolicy(
QtCore.Qt.ScrollBarAlwaysOff
)
if self._fit_to_content["column"]:
self.labelList.setVerticalScrollBarPolicy(
QtCore.Qt.ScrollBarAlwaysOff
)
self._sort_labels = sort_labels
if labels:
self.labelList.addItems(labels)
if self._sort_labels:
self.labelList.sortItems()
else:
self.labelList.setDragDropMode(
QtWidgets.QAbstractItemView.InternalMove
)
self.labelList.currentItemChanged.connect(self.labelSelected)
self.labelList.itemDoubleClicked.connect(self.labelDoubleClicked)
self.edit.setListWidget(self.labelList)
layout.addWidget(self.labelList)
# label_flags
if flags is None:
flags = {}
self._flags = flags
self.flagsLayout = QtWidgets.QVBoxLayout()
self.resetFlags()
layout.addItem(self.flagsLayout)
self.edit.textChanged.connect(self.updateFlags)
self.setLayout(layout)
# completion
completer = QtWidgets.QCompleter()
if not QT5 and completion != "startswith":
completion = "startswith"
if completion == "startswith":
completer.setCompletionMode(QtWidgets.QCompleter.InlineCompletion)
# Default settings.
# completer.setFilterMode(QtCore.Qt.MatchStartsWith)
elif completion == "contains":
completer.setCompletionMode(QtWidgets.QCompleter.PopupCompletion)
completer.setFilterMode(QtCore.Qt.MatchContains)
else:
raise ValueError("Unsupported completion: {}".format(completion))
completer.setModel(self.labelList.model())
self.edit.setCompleter(completer)
def addLabelHistory(self, label):
if self.labelList.findItems(label, QtCore.Qt.MatchExactly):
return
self.labelList.addItem(label)
if self._sort_labels:
self.labelList.sortItems()
def labelSelected(self, item):
self.edit.setText(item.text())
def validate(self):
text = self.edit.text()
if hasattr(text, "strip"):
text = text.strip()
else:
text = text.trimmed()
if text:
self.accept()
def labelDoubleClicked(self, item):
self.validate()
def postProcess(self):
text = self.edit.text()
if hasattr(text, "strip"):
text = text.strip()
else:
text = text.trimmed()
self.edit.setText(text)
def updateFlags(self, label_new):
# keep state of shared flags
flags_old = self.getFlags()
flags_new = {}
for pattern, keys in self._flags.items():
if re.match(pattern, label_new):
for key in keys:
flags_new[key] = flags_old.get(key, False)
self.setFlags(flags_new)
def deleteFlags(self):
for i in reversed(range(self.flagsLayout.count())):
item = self.flagsLayout.itemAt(i).widget()
self.flagsLayout.removeWidget(item)
item.setParent(None)
def resetFlags(self, label=""):
flags = {}
for pattern, keys in self._flags.items():
if re.match(pattern, label):
for key in keys:
flags[key] = False
self.setFlags(flags)
def setFlags(self, flags):
self.deleteFlags()
for key in flags:
item = QtWidgets.QCheckBox(key, self)
item.setChecked(flags[key])
self.flagsLayout.addWidget(item)
item.show()
def getFlags(self):
flags = {}
for i in range(self.flagsLayout.count()):
item = self.flagsLayout.itemAt(i).widget()
flags[item.text()] = item.isChecked()
return flags
def popUp(self, text=None, move=True, flags=None):
if self._fit_to_content["row"]:
self.labelList.setMinimumHeight(
self.labelList.sizeHintForRow(0) * self.labelList.count() + 2
)
if self._fit_to_content["column"]:
self.labelList.setMinimumWidth(
self.labelList.sizeHintForColumn(0) + 2
)
# if text is None, the previous label in self.edit is kept
if text is None:
text = self.edit.text()
if flags:
self.setFlags(flags)
else:
self.resetFlags(text)
self.edit.setText(text)
self.edit.setSelection(0, len(text))
items = self.labelList.findItems(text, QtCore.Qt.MatchFixedString)
if items:
if len(items) != 1:
self.labelList.setCurrentItem(items[0])
row = self.labelList.row(items[0])
self.edit.completer().setCurrentRow(row)
self.edit.setFocus(QtCore.Qt.PopupFocusReason)
if move:
self.move(QtGui.QCursor.pos())
if self.exec_():
return self.edit.text(), self.getFlags()
else:
return None, None
PPOCRLabel/libs/labelColor.py 0 → 100644
浏览文件 @ 4b214948
import PIL.Image
import numpy as np
def rgb2hsv(rgb):
# type: (np.ndarray) -> np.ndarray
"""Convert rgb to hsv.
Parameters
----------
rgb: numpy.ndarray, (H, W, 3), np.uint8
Input rgb image.
Returns
-------
hsv: numpy.ndarray, (H, W, 3), np.uint8
Output hsv image.
"""
hsv = PIL.Image.fromarray(rgb, mode="RGB")
hsv = hsv.convert("HSV")
hsv = np.array(hsv)
return hsv
def hsv2rgb(hsv):
# type: (np.ndarray) -> np.ndarray
"""Convert hsv to rgb.
Parameters
----------
hsv: numpy.ndarray, (H, W, 3), np.uint8
Input hsv image.
Returns
-------
rgb: numpy.ndarray, (H, W, 3), np.uint8
Output rgb image.
"""
rgb = PIL.Image.fromarray(hsv, mode="HSV")
rgb = rgb.convert("RGB")
rgb = np.array(rgb)
return rgb
def label_colormap(n_label=256, value=None):
"""Label colormap.
Parameters
----------
n_label: int
Number of labels (default: 256).
value: float or int
Value scale or value of label color in HSV space.
Returns
-------
cmap: numpy.ndarray, (N, 3), numpy.uint8
Label id to colormap.
"""
def bitget(byteval, idx):
return (byteval & (1 << idx)) != 0
cmap = np.zeros((n_label, 3), dtype=np.uint8)
for i in range(0, n_label):
id = i
r, g, b = 0, 0, 0
for j in range(0, 8):
r = np.bitwise_or(r, (bitget(id, 0) << 7 - j))
g = np.bitwise_or(g, (bitget(id, 1) << 7 - j))
b = np.bitwise_or(b, (bitget(id, 2) << 7 - j))
id = id >> 3
cmap[i, 0] = r
cmap[i, 1] = g
cmap[i, 2] = b
if value is not None:
hsv = rgb2hsv(cmap.reshape(1, -1, 3))
if isinstance(value, float):
hsv[:, 1:, 2] = hsv[:, 1:, 2].astype(float) * value
else:
assert isinstance(value, int)
hsv[:, 1:, 2] = value
cmap = hsv2rgb(hsv).reshape(-1, 3)
return cmap
PPOCRLabel/libs/shape.py
浏览文件 @ 4b214948
......@@ -10,19 +10,14 @@
# SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF
# CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
# THE SOFTWARE.
#!/usr/bin/python
# !/usr/bin/python
# -*- coding: utf-8 -*-
import math
import sys
try:
from PyQt5.QtGui import *
from PyQt5.QtCore import *
except ImportError:
from PyQt4.QtGui import *
from PyQt4.QtCore import *
from PyQt5.QtCore import QPointF
from PyQt5.QtGui import QColor, QPen, QPainterPath, QFont
from libs.utils import distance
import sys
DEFAULT_LINE_COLOR = QColor(0, 255, 0, 128)
DEFAULT_FILL_COLOR = QColor(255, 0, 0, 128)
......@@ -51,14 +46,17 @@ class Shape(object):
point_size = 8
scale = 1.0
def __init__(self, label=None, line_color=None, difficult=False, paintLabel=False):
def __init__(self, label=None, line_color=None, difficult=False, key_cls="None", paintLabel=False):
self.label = label
self.points = []
self.fill = False
self.selected = False
self.difficult = difficult
self.key_cls = key_cls
self.paintLabel = paintLabel
self.locked = False
self.direction = 0
self.center = None
self._highlightIndex = None
self._highlightMode = self.NEAR_VERTEX
self._highlightSettings = {
......@@ -74,7 +72,24 @@ class Shape(object):
# is used for drawing the pending line a different color.
self.line_color = line_color
def rotate(self, theta):
for i, p in enumerate(self.points):
self.points[i] = self.rotatePoint(p, theta)
self.direction -= theta
self.direction = self.direction % (2 * math.pi)
def rotatePoint(self, p, theta):
order = p - self.center
cosTheta = math.cos(theta)
sinTheta = math.sin(theta)
pResx = cosTheta * order.x() + sinTheta * order.y()
pResy = - sinTheta * order.x() + cosTheta * order.y()
pRes = QPointF(self.center.x() + pResx, self.center.y() + pResy)
return pRes
def close(self):
self.center = QPointF((self.points[0].x() + self.points[2].x()) / 2,
(self.points[0].y() + self.points[2].y()) / 2)
self._closed = True
def reachMaxPoints(self):
......@@ -83,7 +98,9 @@ class Shape(object):
return False
def addPoint(self, point):
if not self.reachMaxPoints(): # 4个点时发出close信号
if self.reachMaxPoints():
self.close()
else:
self.points.append(point)
def popPoint(self):
......@@ -112,7 +129,7 @@ class Shape(object):
# Uncommenting the following line will draw 2 paths
# for the 1st vertex, and make it non-filled, which
# may be desirable.
#self.drawVertex(vrtx_path, 0)
# self.drawVertex(vrtx_path, 0)
for i, p in enumerate(self.points):
line_path.lineTo(p)
......@@ -136,9 +153,9 @@ class Shape(object):
font.setPointSize(8)
font.setBold(True)
painter.setFont(font)
if(self.label == None):
if self.label is None:
self.label = ""
if(min_y < MIN_Y_LABEL):
if min_y < MIN_Y_LABEL:
min_y += MIN_Y_LABEL
painter.drawText(min_x, min_y, self.label)
......@@ -198,6 +215,8 @@ class Shape(object):
def copy(self):
shape = Shape("%s" % self.label)
shape.points = [p for p in self.points]
shape.center = self.center
shape.direction = self.direction
shape.fill = self.fill
shape.selected = self.selected
shape._closed = self._closed
......@@ -206,6 +225,7 @@ class Shape(object):
if self.fill_color != Shape.fill_color:
shape.fill_color = self.fill_color
shape.difficult = self.difficult
shape.key_cls = self.key_cls
return shape
def __len__(self):
......
PPOCRLabel/libs/unique_label_qlist_widget.py 0 → 100644
浏览文件 @ 4b214948
# -*- encoding: utf-8 -*-
from PyQt5.QtCore import Qt
from PyQt5 import QtWidgets
class EscapableQListWidget(QtWidgets.QListWidget):
def keyPressEvent(self, event):
super(EscapableQListWidget, self).keyPressEvent(event)
if event.key() == Qt.Key_Escape:
self.clearSelection()
class UniqueLabelQListWidget(EscapableQListWidget):
def mousePressEvent(self, event):
super(UniqueLabelQListWidget, self).mousePressEvent(event)
if not self.indexAt(event.pos()).isValid():
self.clearSelection()
def findItemsByLabel(self, label, get_row=False):
items = []
for row in range(self.count()):
item = self.item(row)
if item.data(Qt.UserRole) == label:
items.append(item)
if get_row:
return row
return items
def createItemFromLabel(self, label):
item = QtWidgets.QListWidgetItem()
item.setData(Qt.UserRole, label)
return item
def setItemLabel(self, item, label, color=None):
qlabel = QtWidgets.QLabel()
if color is None:
qlabel.setText(f"{label}")
else:
qlabel.setText('<font color="#{:02x}{:02x}{:02x}">●</font> {} '.format(*color, label))
qlabel.setAlignment(Qt.AlignBottom)
item.setSizeHint(qlabel.sizeHint())
self.setItemWidget(item, qlabel)
PPOCRLabel/libs/utils.py
浏览文件 @ 4b214948
......@@ -10,30 +10,26 @@
# SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF
# CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
# THE SOFTWARE.
from math import sqrt
from libs.ustr import ustr
import hashlib
import os
import re
import sys
from math import sqrt
import cv2
import numpy as np
import os
from PyQt5.QtCore import QRegExp, QT_VERSION_STR
from PyQt5.QtGui import QIcon, QRegExpValidator, QColor
from PyQt5.QtWidgets import QPushButton, QAction, QMenu
from libs.ustr import ustr
__dir__ = os.path.dirname(os.path.abspath(__file__)) # 获取本程序文件路径
__iconpath__ = os.path.abspath(os.path.join(__dir__, '../resources/icons'))
try:
from PyQt5.QtGui import *
from PyQt5.QtCore import *
from PyQt5.QtWidgets import *
except ImportError:
from PyQt4.QtGui import *
from PyQt4.QtCore import *
def newIcon(icon, iconSize=None):
if iconSize is not None:
return QIcon(QIcon(__iconpath__ + "/" + icon + ".png").pixmap(iconSize,iconSize))
return QIcon(QIcon(__iconpath__ + "/" + icon + ".png").pixmap(iconSize, iconSize))
else:
return QIcon(__iconpath__ + "/" + icon + ".png")
......@@ -109,20 +105,21 @@ def generateColorByText(text):
b = int((hashCode / 16581375) % 255)
return QColor(r, g, b, 100)
def have_qstring():
'''p3/qt5 get rid of QString wrapper as py3 has native unicode str type'''
return not (sys.version_info.major >= 3 or QT_VERSION_STR.startswith('5.'))
def util_qt_strlistclass():
return QStringList if have_qstring() else list
def natural_sort(list, key=lambda s:s):
def natural_sort(list, key=lambda s: s):
"""
Sort the list into natural alphanumeric order.
"""
def get_alphanum_key_func(key):
convert = lambda text: int(text) if text.isdigit() else text
return lambda s: [convert(c) for c in re.split('([0-9]+)', key(s))]
sort_key = get_alphanum_key_func(key)
list.sort(key=sort_key)
......@@ -163,10 +160,11 @@ def get_rotate_crop_image(img, points):
except Exception as e:
print(e)
def stepsInfo(lang='en'):
if lang == 'ch':
msg = "1. 安装与运行：使用上述命令安装与运行程序。\n" \
"2. 打开文件夹：在菜单栏点击 “文件” - 打开目录 选择待标记图片的文件夹.\n"\
"2. 打开文件夹：在菜单栏点击 “文件” - 打开目录 选择待标记图片的文件夹.\n" \
"3. 自动标注：点击 ”自动标注“，使用PPOCR超轻量模型对图片文件名前图片状态为 “X” 的图片进行自动标注。\n" \
"4. 手动标注：点击 “矩形标注”（推荐直接在英文模式下点击键盘中的 “W”)，用户可对当前图片中模型未检出的部分进行手动" \
"绘制标记框。点击键盘P，则使用四点标注模式（或点击“编辑” - “四点标注”），用户依次点击4个点后，双击左键表示标注完成。\n" \
......@@ -181,25 +179,26 @@ def stepsInfo(lang='en'):
else:
msg = "1. Build and launch using the instructions above.\n" \
"2. Click 'Open Dir' in Menu/File to select the folder of the picture.\n"\
"3. Click 'Auto recognition', use PPOCR model to automatically annotate images which marked with 'X' before the file name."\
"4. Create Box:\n"\
"4.1 Click 'Create RectBox' or press 'W' in English keyboard mode to draw a new rectangle detection box. Click and release left mouse to select a region to annotate the text area.\n"\
"4.2 Press 'P' to enter four-point labeling mode which enables you to create any four-point shape by clicking four points with the left mouse button in succession and DOUBLE CLICK the left mouse as the signal of labeling completion.\n"\
"5. After the marking frame is drawn, the user clicks 'OK', and the detection frame will be pre-assigned a TEMPORARY label.\n"\
"6. Click re-Recognition, model will rewrite ALL recognition results in ALL detection box.\n"\
"7. Double click the result in 'recognition result' list to manually change inaccurate recognition results.\n"\
"8. Click 'Save', the image status will switch to '√',then the program automatically jump to the next.\n"\
"9. Click 'Delete Image' and the image will be deleted to the recycle bin.\n"\
"10. Labeling result: After closing the application or switching the file path, the manually saved label will be stored in *Label.txt* under the opened picture folder.\n"\
"2. Click 'Open Dir' in Menu/File to select the folder of the picture.\n" \
"3. Click 'Auto recognition', use PPOCR model to automatically annotate images which marked with 'X' before the file name." \
"4. Create Box:\n" \
"4.1 Click 'Create RectBox' or press 'W' in English keyboard mode to draw a new rectangle detection box. Click and release left mouse to select a region to annotate the text area.\n" \
"4.2 Press 'P' to enter four-point labeling mode which enables you to create any four-point shape by clicking four points with the left mouse button in succession and DOUBLE CLICK the left mouse as the signal of labeling completion.\n" \
"5. After the marking frame is drawn, the user clicks 'OK', and the detection frame will be pre-assigned a TEMPORARY label.\n" \
"6. Click re-Recognition, model will rewrite ALL recognition results in ALL detection box.\n" \
"7. Double click the result in 'recognition result' list to manually change inaccurate recognition results.\n" \
"8. Click 'Save', the image status will switch to '√',then the program automatically jump to the next.\n" \
"9. Click 'Delete Image' and the image will be deleted to the recycle bin.\n" \
"10. Labeling result: After closing the application or switching the file path, the manually saved label will be stored in *Label.txt* under the opened picture folder.\n" \
" Click PaddleOCR-Save Recognition Results in the menu bar, the recognition training data of such pictures will be saved in the *crop_img* folder, and the recognition label will be saved in *rec_gt.txt*.\n"
return msg
def keysInfo(lang='en'):
if lang == 'ch':
msg = "快捷键\t\t\t说明\n" \
"———————————————————————\n"\
"———————————————————————\n" \
"Ctrl + shift + R\t\t对当前图片的所有标记重新识别\n" \
"W\t\t\t新建矩形框\n" \
"Q\t\t\t新建四点框\n" \
......@@ -223,17 +222,17 @@ def keysInfo(lang='en'):
"———————————————————————\n" \
"Ctrl + shift + R\t\tRe-recognize all the labels\n" \
"\t\t\tof the current image\n" \
"\n"\
"\n" \
"W\t\t\tCreate a rect box\n" \
"Q\t\t\tCreate a four-points box\n" \
"Ctrl + E\t\tEdit label of the selected box\n" \
"Ctrl + R\t\tRe-recognize the selected box\n" \
"Ctrl + C\t\tCopy and paste the selected\n" \
"\t\t\tbox\n" \
"\n"\
"\n" \
"Ctrl + Left Mouse\tMulti select the label\n" \
"Button\t\t\tbox\n" \
"\n"\
"\n" \
"Backspace\t\tDelete the selected box\n" \
"Ctrl + V\t\tCheck image\n" \
"Ctrl + Shift + d\tDelete image\n" \
......
PPOCRLabel/resources/strings/strings-en.properties
浏览文件 @ 4b214948
......@@ -107,3 +107,6 @@ undoLastPoint=Undo Last Point
autoSaveMode=Auto Export Label Mode
lockBox=Lock selected box/Unlock all box
lockBoxDetail=Lock selected box/Unlock all box
keyListTitle=Key List
keyDialogTip=Enter object label
keyChange=Change Box Key
PPOCRLabel/resources/strings/strings-zh-CN.properties
浏览文件 @ 4b214948
......@@ -107,3 +107,6 @@ undoLastPoint=撤销上个点
autoSaveMode=自动导出标记结果
lockBox=锁定框/解除锁定框
lockBoxDetail=若当前没有框处于锁定状态则锁定选中的框，若存在锁定框则解除所有锁定框的锁定状态
keyListTitle=关键词列表
keyDialogTip=请输入类型名称
keyChange=更改Box关键字类别
\ No newline at end of file
README.md
浏览文件 @ 4b214948
......@@ -92,7 +92,7 @@ Mobile DEMO experience (based on EasyEdge and Paddle-Lite, supports iOS and Andr
| ------------------------------------------------------------ | ---------------------------- | ----------------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ |
| Chinese and English ultra-lightweight PP-OCRv2 model（11.6M） | ch_PP-OCRv2_xx |Mobile & Server|[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_distill_train.tar)| [inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar)|
| Chinese and English ultra-lightweight PP-OCR model (9.4M) | ch_ppocr_mobile_v2.0_xx | Mobile & server |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_train.tar)|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_train.tar) |
| Chinese and English general PP-OCR model (143.4M) | ch_ppocr_server_v2.0_xx | Server |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_train.tar) |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_traingit.tar) |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_train.tar) |
| Chinese and English general PP-OCR model (143.4M) | ch_ppocr_server_v2.0_xx | Server |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_train.tar) |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_train.tar) |
For more model downloads (including multiple languages), please refer to [PP-OCR series model downloads](./doc/doc_en/models_list_en.md).
......@@ -152,7 +152,7 @@ For a new language request, please refer to [Guideline for new language_requests
[1] PP-OCR is a practical ultra-lightweight OCR system. It is mainly composed of three parts: DB text detection, detection frame correction and CRNN text recognition. The system adopts 19 effective strategies from 8 aspects including backbone network selection and adjustment, prediction head design, data augmentation, learning rate transformation strategy, regularization parameter selection, pre-training model use, and automatic model tailoring and quantization to optimize and slim down the models of each module (as shown in the green box above). The final results are an ultra-lightweight Chinese and English OCR model with an overall size of 3.5M and a 2.8M English digital OCR model. For more details, please refer to the PP-OCR technical article (https://arxiv.org/abs/2009.09941).
[2] On the basis of PP-OCR, PP-OCRv2 is further optimized in five aspects. The detection model adopts CML(Collaborative Mutual Learning) knowledge distillation strategy and CopyPaste data expansion strategy. The recognition model adopts LCNet lightweight backbone network, U-DML knowledge distillation strategy and enhanced CTC loss function improvement (as shown in the red box above), which further improves the inference speed and prediction effect. For more details, please refer to the technical report of PP-OCRv2 (arXiv link is coming soon).
[2] On the basis of PP-OCR, PP-OCRv2 is further optimized in five aspects. The detection model adopts CML(Collaborative Mutual Learning) knowledge distillation strategy and CopyPaste data expansion strategy. The recognition model adopts LCNet lightweight backbone network, U-DML knowledge distillation strategy and enhanced CTC loss function improvement (as shown in the red box above), which further improves the inference speed and prediction effect. For more details, please refer to the technical report of PP-OCRv2 (https://arxiv.org/abs/2109.03144).
......@@ -181,16 +181,11 @@ For a new language request, please refer to [Guideline for new language_requests
<a name="language_requests"></a>
## Guideline for New Language Requests
If you want to request a new language support, a PR with 2 following files are needed：
If you want to request a new language support, a PR with 1 following files are needed：
1. In folder [ppocr/utils/dict](./ppocr/utils/dict),
it is necessary to submit the dict text to this path and name it with `{language}_dict.txt` that contains a list of all characters. Please see the format example from other files in that folder.
2. In folder [ppocr/utils/corpus](./ppocr/utils/corpus),
it is necessary to submit the corpus to this path and name it with `{language}_corpus.txt` that contains a list of words in your language.
Maybe, 50000 words per language is necessary at least.
Of course, the more, the better.
If your language has unique elements, please tell me in advance within any way, such as useful links, wikipedia and so on.
More details, please refer to [Multilingual OCR Development Plan](https://github.com/PaddlePaddle/PaddleOCR/issues/1048).
......
README_ch.md
浏览文件 @ 4b214948
......@@ -99,7 +99,7 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
- [PP-Structure信息提取](./ppstructure/README_ch.md)
- [版面分析](./ppstructure/layout/README_ch.md)
- [表格识别](./ppstructure/table/README_ch.md)
- [DocVQA](./ppstructure/vqa/README_ch.md)
- [DocVQA](./ppstructure/vqa/README.md)
- [关键信息提取](./ppstructure/docs/kie.md)
- OCR学术圈
- [两阶段模型介绍与下载](./doc/doc_ch/algorithm_overview.md)
......
benchmark/analysis.py
浏览文件 @ 4b214948
......@@ -26,35 +26,57 @@ def parse_args():
parser.add_argument(
"--filename", type=str, help="The name of log which need to analysis.")
parser.add_argument(
"--log_with_profiler", type=str, help="The path of train log with profiler")
"--log_with_profiler",
type=str,
help="The path of train log with profiler")
parser.add_argument(
"--profiler_path", type=str, help="The path of profiler timeline log.")
parser.add_argument(
"--keyword", type=str, help="Keyword to specify analysis data")
parser.add_argument(
"--separator", type=str, default=None, help="Separator of different field in log")
"--separator",
type=str,
default=None,
help="Separator of different field in log")
parser.add_argument(
'--position', type=int, default=None, help='The position of data field')
parser.add_argument(
'--range', type=str, default="", help='The range of data field to intercept')
'--range',
type=str,
default="",
help='The range of data field to intercept')
parser.add_argument(
'--base_batch_size', type=int, help='base_batch size on gpu')
parser.add_argument(
'--skip_steps', type=int, default=0, help='The number of steps to be skipped')
'--skip_steps',
type=int,
default=0,
help='The number of steps to be skipped')
parser.add_argument(
'--model_mode', type=int, default=-1, help='Analysis mode, default value is -1')
'--model_mode',
type=int,
default=-1,
help='Analysis mode, default value is -1')
parser.add_argument('--ips_unit', type=str, default=None, help='IPS unit')
parser.add_argument(
'--ips_unit', type=str, default=None, help='IPS unit')
parser.add_argument(
'--model_name', type=str, default=0, help='training model_name, transformer_base')
'--model_name',
type=str,
default=0,
help='training model_name, transformer_base')
parser.add_argument(
'--mission_name', type=str, default=0, help='training mission name')
parser.add_argument(
'--direction_id', type=int, default=0, help='training direction_id')
parser.add_argument(
'--run_mode', type=str, default="sp", help='multi process or single process')
'--run_mode',
type=str,
default="sp",
help='multi process or single process')
parser.add_argument(
'--index', type=int, default=1, help='{1: speed, 2:mem, 3:profiler, 6:max_batch_size}')
'--index',
type=int,
default=1,
help='{1: speed, 2:mem, 3:profiler, 6:max_batch_size}')
parser.add_argument(
'--gpu_num', type=int, default=1, help='nums of training gpus')
args = parser.parse_args()
......@@ -72,7 +94,12 @@ def _is_number(num):
class TimeAnalyzer(object):
def __init__(self, filename, keyword=None, separator=None, position=None, range="-1"):
def __init__(self,
filename,
keyword=None,
separator=None,
position=None,
range="-1"):
if filename is None:
raise Exception("Please specify the filename!")
......@@ -99,7 +126,8 @@ class TimeAnalyzer(object):
# Distil the string from a line.
line = line.strip()
line_words = line.split(self.separator) if self.separator else line.split()
line_words = line.split(
self.separator) if self.separator else line.split()
if args.position:
result = line_words[self.position]
else:
......@@ -113,16 +141,25 @@ class TimeAnalyzer(object):
if not self.range:
result = result[0:]
elif _is_number(self.range):
result = result[0: int(self.range)]
result = result[0:int(self.range)]
else:
result = result[int(self.range.split(":")[0]): int(self.range.split(":")[1])]
result = result[int(self.range.split(":")[0]):int(
self.range.split(":")[1])]
self.records.append(float(result))
except Exception as exc:
print("line is: {}; separator={}; position={}".format(line, self.separator, self.position))
print("line is: {}; separator={}; position={}".format(
line, self.separator, self.position))
print("Extract {} records: separator={}; position={}".format(len(self.records), self.separator, self.position))
print("Extract {} records: separator={}; position={}".format(
len(self.records), self.separator, self.position))
def _get_fps(self, mode, batch_size, gpu_num, avg_of_records, run_mode, unit=None):
def _get_fps(self,
mode,
batch_size,
gpu_num,
avg_of_records,
run_mode,
unit=None):
if mode == -1 and run_mode == 'sp':
assert unit, "Please set the unit when mode is -1."
fps = gpu_num * avg_of_records
......@@ -155,12 +192,20 @@ class TimeAnalyzer(object):
return fps, unit
def analysis(self, batch_size, gpu_num=1, skip_steps=0, mode=-1, run_mode='sp', unit=None):
def analysis(self,
batch_size,
gpu_num=1,
skip_steps=0,
mode=-1,
run_mode='sp',
unit=None):
if batch_size <= 0:
print("base_batch_size should larger than 0.")
return 0, ''
if len(self.records) <= skip_steps: # to address the condition which item of log equals to skip_steps
if len(
self.records
) <= skip_steps: # to address the condition which item of log equals to skip_steps
print("no records")
return 0, ''
......@@ -180,16 +225,20 @@ class TimeAnalyzer(object):
skip_max = self.records[i]
avg_of_records = sum_of_records / float(count)
avg_of_records_skipped = sum_of_records_skipped / float(count - skip_steps)
avg_of_records_skipped = sum_of_records_skipped / float(count -
skip_steps)
fps, fps_unit = self._get_fps(mode, batch_size, gpu_num, avg_of_records, run_mode, unit)
fps_skipped, _ = self._get_fps(mode, batch_size, gpu_num, avg_of_records_skipped, run_mode, unit)
fps, fps_unit = self._get_fps(mode, batch_size, gpu_num, avg_of_records,
run_mode, unit)
fps_skipped, _ = self._get_fps(mode, batch_size, gpu_num,
avg_of_records_skipped, run_mode, unit)
if mode == -1:
print("average ips of %d steps, skip 0 step:" % count)
print("\tAvg: %.3f %s" % (avg_of_records, fps_unit))
print("\tFPS: %.3f %s" % (fps, fps_unit))
if skip_steps > 0:
print("average ips of %d steps, skip %d steps:" % (count, skip_steps))
print("average ips of %d steps, skip %d steps:" %
(count, skip_steps))
print("\tAvg: %.3f %s" % (avg_of_records_skipped, fps_unit))
print("\tMin: %.3f %s" % (skip_min, fps_unit))
print("\tMax: %.3f %s" % (skip_max, fps_unit))
......@@ -199,7 +248,8 @@ class TimeAnalyzer(object):
print("\tAvg: %.3f steps/s" % avg_of_records)
print("\tFPS: %.3f %s" % (fps, fps_unit))
if skip_steps > 0:
print("average latency of %d steps, skip %d steps:" % (count, skip_steps))
print("average latency of %d steps, skip %d steps:" %
(count, skip_steps))
print("\tAvg: %.3f steps/s" % avg_of_records_skipped)
print("\tMin: %.3f steps/s" % skip_min)
print("\tMax: %.3f steps/s" % skip_max)
......@@ -209,7 +259,8 @@ class TimeAnalyzer(object):
print("\tAvg: %.3f s/step" % avg_of_records)
print("\tFPS: %.3f %s" % (fps, fps_unit))
if skip_steps > 0:
print("average latency of %d steps, skip %d steps:" % (count, skip_steps))
print("average latency of %d steps, skip %d steps:" %
(count, skip_steps))
print("\tAvg: %.3f s/step" % avg_of_records_skipped)
print("\tMin: %.3f s/step" % skip_min)
print("\tMax: %.3f s/step" % skip_max)
......@@ -236,7 +287,8 @@ if __name__ == "__main__":
if args.gpu_num == 1:
run_info["log_with_profiler"] = args.log_with_profiler
run_info["profiler_path"] = args.profiler_path
analyzer = TimeAnalyzer(args.filename, args.keyword, args.separator, args.position, args.range)
analyzer = TimeAnalyzer(args.filename, args.keyword, args.separator,
args.position, args.range)
run_info["FINAL_RESULT"], run_info["UNIT"] = analyzer.analysis(
batch_size=args.base_batch_size,
gpu_num=args.gpu_num,
......@@ -245,29 +297,50 @@ if __name__ == "__main__":
run_mode=args.run_mode,
unit=args.ips_unit)
try:
if int(os.getenv('job_fail_flag')) == 1 or int(run_info["FINAL_RESULT"]) == 0:
if int(os.getenv('job_fail_flag')) == 1 or int(run_info[
"FINAL_RESULT"]) == 0:
run_info["JOB_FAIL_FLAG"] = 1
except:
pass
elif args.index == 3:
run_info["FINAL_RESULT"] = {}
records_fo_total = TimeAnalyzer(args.filename, 'Framework overhead', None, 3, '').records
records_fo_ratio = TimeAnalyzer(args.filename, 'Framework overhead', None, 5).records
records_ct_total = TimeAnalyzer(args.filename, 'Computation time', None, 3, '').records
records_gm_total = TimeAnalyzer(args.filename, 'GpuMemcpy Calls', None, 4, '').records
records_gm_ratio = TimeAnalyzer(args.filename, 'GpuMemcpy Calls', None, 6).records
records_gmas_total = TimeAnalyzer(args.filename, 'GpuMemcpyAsync Calls', None, 4, '').records
records_gms_total = TimeAnalyzer(args.filename, 'GpuMemcpySync Calls', None, 4, '').records
run_info["FINAL_RESULT"]["Framework_Total"] = records_fo_total[0] if records_fo_total else 0
run_info["FINAL_RESULT"]["Framework_Ratio"] = records_fo_ratio[0] if records_fo_ratio else 0
run_info["FINAL_RESULT"]["ComputationTime_Total"] = records_ct_total[0] if records_ct_total else 0
run_info["FINAL_RESULT"]["GpuMemcpy_Total"] = records_gm_total[0] if records_gm_total else 0
run_info["FINAL_RESULT"]["GpuMemcpy_Ratio"] = records_gm_ratio[0] if records_gm_ratio else 0
run_info["FINAL_RESULT"]["GpuMemcpyAsync_Total"] = records_gmas_total[0] if records_gmas_total else 0
run_info["FINAL_RESULT"]["GpuMemcpySync_Total"] = records_gms_total[0] if records_gms_total else 0
records_fo_total = TimeAnalyzer(args.filename, 'Framework overhead',
None, 3, '').records
records_fo_ratio = TimeAnalyzer(args.filename, 'Framework overhead',
None, 5).records
records_ct_total = TimeAnalyzer(args.filename, 'Computation time',
None, 3, '').records
records_gm_total = TimeAnalyzer(args.filename,
'GpuMemcpy Calls',
None, 4, '').records
records_gm_ratio = TimeAnalyzer(args.filename,
'GpuMemcpy Calls',
None, 6).records
records_gmas_total = TimeAnalyzer(args.filename,
'GpuMemcpyAsync Calls',
None, 4, '').records
records_gms_total = TimeAnalyzer(args.filename,
'GpuMemcpySync Calls',
None, 4, '').records
run_info["FINAL_RESULT"]["Framework_Total"] = records_fo_total[
0] if records_fo_total else 0
run_info["FINAL_RESULT"]["Framework_Ratio"] = records_fo_ratio[
0] if records_fo_ratio else 0
run_info["FINAL_RESULT"][
"ComputationTime_Total"] = records_ct_total[
0] if records_ct_total else 0
run_info["FINAL_RESULT"]["GpuMemcpy_Total"] = records_gm_total[
0] if records_gm_total else 0
run_info["FINAL_RESULT"]["GpuMemcpy_Ratio"] = records_gm_ratio[
0] if records_gm_ratio else 0
run_info["FINAL_RESULT"][
"GpuMemcpyAsync_Total"] = records_gmas_total[
0] if records_gmas_total else 0
run_info["FINAL_RESULT"]["GpuMemcpySync_Total"] = records_gms_total[
0] if records_gms_total else 0
else:
print("Not support!")
except Exception:
traceback.print_exc()
print("{}".format(json.dumps(run_info))) # it's required, for the log file path insert to the database
print("{}".format(json.dumps(run_info))
) # it's required, for the log file path insert to the database
benchmark/run_benchmark_det.sh
浏览文件 @ 4b214948
......@@ -58,3 +58,4 @@ source ${BENCHMARK_ROOT}/scripts/run_model.sh # 在该脚本中会对符合
_set_params $@
#_train # 如果只想产出训练log,不解析,可取消注释
_run # 该函数在run_model.sh中,执行时会调用_train; 如果不联调只想要产出训练log可以注掉本行,提交时需打开
benchmark/run_det.sh
浏览文件 @ 4b214948
......@@ -36,3 +36,4 @@ for model_mode in ${model_mode_list[@]}; do
done
configs/det/det_mv3_db.yml
浏览文件 @ 4b214948
Global:
use_gpu: true
use_xpu: false
epoch_num: 1200
log_smooth_window: 20
print_batch_step: 10
......
configs/rec/rec_efficientb3_fpn_pren.yml 0 → 100644
浏览文件 @ 4b214948
Global:
use_gpu: True
epoch_num: 8
log_smooth_window: 20
print_batch_step: 5
save_model_dir: ./output/rec/pren_new
save_epoch_step: 3
# evaluation is run every 2000 iterations after the 4000th iteration
eval_batch_step: [4000, 2000]
cal_metric_during_train: True
pretrained_model:
checkpoints:
save_inference_dir:
use_visualdl: False
infer_img: doc/imgs_words/ch/word_1.jpg
# for data or label process
character_dict_path:
max_text_length: &max_text_length 25
infer_mode: False
use_space_char: False
save_res_path: ./output/rec/predicts_pren.txt
Optimizer:
name: Adadelta
lr:
name: Piecewise
decay_epochs: [2, 5, 7]
values: [0.5, 0.1, 0.01, 0.001]
Architecture:
model_type: rec
algorithm: PREN
in_channels: 3
Backbone:
name: EfficientNetb3_PREN
Neck:
name: PRENFPN
n_r: 5
d_model: 384
max_len: *max_text_length
dropout: 0.1
Head:
name: PRENHead
Loss:
name: PRENLoss
PostProcess:
name: PRENLabelDecode
Metric:
name: RecMetric
main_indicator: acc
Train:
dataset:
name: LMDBDataSet
data_dir: ./train_data/data_lmdb_release/training/
transforms:
- DecodeImage:
img_mode: BGR
channel_first: False
- PRENLabelEncode:
- RecAug:
- PRENResizeImg:
image_shape: [64, 256] # h,w
- KeepKeys:
keep_keys: ['image', 'label']
loader:
shuffle: True
batch_size_per_card: 128
drop_last: True
num_workers: 8
Eval:
dataset:
name: LMDBDataSet
data_dir: ./train_data/data_lmdb_release/validation/
transforms:
- DecodeImage:
img_mode: BGR
channel_first: False
- PRENLabelEncode:
- PRENResizeImg:
image_shape: [64, 256] # h,w
- KeepKeys:
keep_keys: ['image', 'label']
loader:
shuffle: False
drop_last: False
batch_size_per_card: 64
num_workers: 8
configs/vqa/re/layoutlmv2.yml 0 → 100644
浏览文件 @ 4b214948
Global:
use_gpu: True
epoch_num: &epoch_num 200
log_smooth_window: 10
print_batch_step: 10
save_model_dir: ./output/re_layoutlmv2/
save_epoch_step: 2000
# evaluation is run every 10 iterations after the 0th iteration
eval_batch_step: [ 0, 19 ]
cal_metric_during_train: False
save_inference_dir:
use_visualdl: False
seed: 2048
infer_img: doc/vqa/input/zh_val_21.jpg
save_res_path: ./output/re/
Architecture:
model_type: vqa
algorithm: &algorithm "LayoutLMv2"
Transform:
Backbone:
name: LayoutLMv2ForRe
pretrained: True
checkpoints:
Loss:
name: LossFromOutput
key: loss
reduction: mean
Optimizer:
name: AdamW
beta1: 0.9
beta2: 0.999
clip_norm: 10
lr:
learning_rate: 0.00005
warmup_epoch: 10
regularizer:
name: L2
factor: 0.00000
PostProcess:
name: VQAReTokenLayoutLMPostProcess
Metric:
name: VQAReTokenMetric
main_indicator: hmean
Train:
dataset:
name: SimpleDataSet
data_dir: train_data/XFUND/zh_train/image
label_file_list:
- train_data/XFUND/zh_train/xfun_normalize_train.json
ratio_list: [ 1.0 ]
transforms:
- DecodeImage: # load image
img_mode: RGB
channel_first: False
- VQATokenLabelEncode: # Class handling label
contains_re: True
algorithm: *algorithm
class_path: &class_path ppstructure/vqa/labels/labels_ser.txt
- VQATokenPad:
max_seq_len: &max_seq_len 512
return_attention_mask: True
- VQAReTokenRelation:
- VQAReTokenChunk:
max_seq_len: *max_seq_len
- Resize:
size: [224,224]
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: [ 'input_ids', 'bbox', 'image', 'attention_mask', 'token_type_ids','entities', 'relations'] # dataloader will return list in this order
loader:
shuffle: True
drop_last: False
batch_size_per_card: 8
num_workers: 8
collate_fn: ListCollator
Eval:
dataset:
name: SimpleDataSet
data_dir: train_data/XFUND/zh_val/image
label_file_list:
- train_data/XFUND/zh_val/xfun_normalize_val.json
transforms:
- DecodeImage: # load image
img_mode: RGB
channel_first: False
- VQATokenLabelEncode: # Class handling label
contains_re: True
algorithm: *algorithm
class_path: *class_path
- VQATokenPad:
max_seq_len: *max_seq_len
return_attention_mask: True
- VQAReTokenRelation:
- VQAReTokenChunk:
max_seq_len: *max_seq_len
- Resize:
size: [224,224]
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: [ 'input_ids', 'bbox', 'image', 'attention_mask', 'token_type_ids','entities', 'relations'] # dataloader will return list in this order
loader:
shuffle: False
drop_last: False
batch_size_per_card: 8
num_workers: 8
collate_fn: ListCollator
configs/vqa/re/layoutxlm.yml
浏览文件 @ 4b214948
......@@ -35,6 +35,7 @@ Optimizer:
clip_norm: 10
lr:
learning_rate: 0.00005
warmup_epoch: 10
regularizer:
name: L2
factor: 0.00000
......@@ -81,7 +82,7 @@ Train:
shuffle: True
drop_last: False
batch_size_per_card: 8
num_workers: 4
num_workers: 8
collate_fn: ListCollator
Eval:
......@@ -118,5 +119,5 @@ Eval:
shuffle: False
drop_last: False
batch_size_per_card: 8
num_workers: 4
num_workers: 8
collate_fn: ListCollator
configs/vqa/ser/layoutlmv2.yml 0 → 100644
浏览文件 @ 4b214948
Global:
use_gpu: True
epoch_num: &epoch_num 200
log_smooth_window: 10
print_batch_step: 10
save_model_dir: ./output/ser_layoutlmv2/
save_epoch_step: 2000
# evaluation is run every 10 iterations after the 0th iteration
eval_batch_step: [ 0, 19 ]
cal_metric_during_train: False
save_inference_dir:
use_visualdl: False
seed: 2022
infer_img: doc/vqa/input/zh_val_0.jpg
save_res_path: ./output/ser/
Architecture:
model_type: vqa
algorithm: &algorithm "LayoutLMv2"
Transform:
Backbone:
name: LayoutLMv2ForSer
pretrained: True
checkpoints:
num_classes: &num_classes 7
Loss:
name: VQASerTokenLayoutLMLoss
num_classes: *num_classes
Optimizer:
name: AdamW
beta1: 0.9
beta2: 0.999
lr:
name: Linear
learning_rate: 0.00005
epochs: *epoch_num
warmup_epoch: 2
regularizer:
name: L2
factor: 0.00000
PostProcess:
name: VQASerTokenLayoutLMPostProcess
class_path: &class_path ppstructure/vqa/labels/labels_ser.txt
Metric:
name: VQASerTokenMetric
main_indicator: hmean
Train:
dataset:
name: SimpleDataSet
data_dir: train_data/XFUND/zh_train/image
label_file_list:
- train_data/XFUND/zh_train/xfun_normalize_train.json
transforms:
- DecodeImage: # load image
img_mode: RGB
channel_first: False
- VQATokenLabelEncode: # Class handling label
contains_re: False
algorithm: *algorithm
class_path: *class_path
- VQATokenPad:
max_seq_len: &max_seq_len 512
return_attention_mask: True
- VQASerTokenChunk:
max_seq_len: *max_seq_len
- Resize:
size: [224,224]
- NormalizeImage:
scale: 1
mean: [ 123.675, 116.28, 103.53 ]
std: [ 58.395, 57.12, 57.375 ]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: [ 'input_ids','labels', 'bbox', 'image', 'attention_mask', 'token_type_ids'] # dataloader will return list in this order
loader:
shuffle: True
drop_last: False
batch_size_per_card: 8
num_workers: 4
Eval:
dataset:
name: SimpleDataSet
data_dir: train_data/XFUND/zh_val/image
label_file_list:
- train_data/XFUND/zh_val/xfun_normalize_val.json
transforms:
- DecodeImage: # load image
img_mode: RGB
channel_first: False
- VQATokenLabelEncode: # Class handling label
contains_re: False
algorithm: *algorithm
class_path: *class_path
- VQATokenPad:
max_seq_len: *max_seq_len
return_attention_mask: True
- VQASerTokenChunk:
max_seq_len: *max_seq_len
- Resize:
size: [224,224]
- NormalizeImage:
scale: 1
mean: [ 123.675, 116.28, 103.53 ]
std: [ 58.395, 57.12, 57.375 ]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: [ 'input_ids', 'labels', 'bbox', 'image', 'attention_mask', 'token_type_ids'] # dataloader will return list in this order
loader:
shuffle: False
drop_last: False
batch_size_per_card: 8
num_workers: 4
deploy/android_demo/README.md
浏览文件 @ 4b214948
# 如何快速测试
### 1. 安装最新版本的Android Studio
- [Android Demo](#android-demo)
- [1. 简介](#1-简介)
- [2. 近期更新](#2-近期更新)
- [3. 快速使用](#3-快速使用)
- [3.1 安装最新版本的Android Studio](#31-安装最新版本的android-studio)
- [3.2 安装 NDK 20 以上版本](#32-安装-ndk-20-以上版本)
- [3.3 导入项目](#33-导入项目)
- [4 更多支持](#4-更多支持)
# Android Demo
## 1. 简介
此为PaddleOCR的Android Demo，目前支持文本检测，文本方向分类器和文本识别模型的使用。使用 [PaddleLite v2.10](https://github.com/PaddlePaddle/Paddle-Lite/tree/release/v2.10) 进行开发。
## 2. 近期更新
* 2022.02.27
* 预测库更新到PaddleLite v2.10
* 支持6种运行模式：
* 检测+分类+识别
* 检测+识别
* 分类+识别
* 检测
* 识别
* 分类
## 3. 快速使用
### 3.1 安装最新版本的Android Studio
可以从 https://developer.android.com/studio 下载。本Demo使用是4.0版本Android Studio编写。
### 2. 按照NDK 20 以上版本
### 3.2 安装 NDK 20 以上版本
Demo测试的时候使用的是NDK 20b版本，20版本以上均可以支持编译成功。
如果您是初学者，可以用以下方式安装和测试NDK编译环境。
点击 File -> New ->New Project， 新建 "Native C++" project
### 3. 导入项目
### 3.3 导入项目
点击 File->New->Import Project...， 然后跟着Android Studio的引导导入
## 4 更多支持
# 获得更多支持
前往[端计算模型生成平台EasyEdge](https://ai.baidu.com/easyedge/app/open_source_demo?referrerUrl=paddlelite)，获得更多开发支持：
前往[Paddle-Lite](https://github.com/PaddlePaddle/Paddle-Lite)，获得更多开发支持
- Demo APP：可使用手机扫码安装，方便手机端快速体验文字识别
- SDK：模型被封装为适配不同芯片硬件和操作系统SDK，包括完善的接口，方便进行二次开发
deploy/android_demo/app/build.gradle
浏览文件 @ 4b214948
......@@ -8,8 +8,8 @@ android {
applicationId "com.baidu.paddle.lite.demo.ocr"
minSdkVersion 23
targetSdkVersion 29
versionCode 1
versionName "1.0"
versionCode 2
versionName "2.0"
testInstrumentationRunner "android.support.test.runner.AndroidJUnitRunner"
externalNativeBuild {
cmake {
......@@ -17,11 +17,6 @@ android {
arguments '-DANDROID_PLATFORM=android-23', '-DANDROID_STL=c++_shared' ,"-DANDROID_ARM_NEON=TRUE"
}
}
ndk {
// abiFilters "arm64-v8a", "armeabi-v7a"
abiFilters "arm64-v8a", "armeabi-v7a"
ldLibs "jnigraphics"
}
}
buildTypes {
release {
......@@ -48,7 +43,7 @@ dependencies {
def archives = [
[
'src' : 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/paddle_lite_libs_v2_9_0.tar.gz',
'src' : 'https://paddleocr.bj.bcebos.com/libs/paddle_lite_libs_v2_10.tar.gz',
'dest': 'PaddleLite'
],
[
......@@ -56,7 +51,7 @@ def archives = [
'dest': 'OpenCV'
],
[
'src' : 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ocr_v2_for_cpu.tar.gz',
'src' : 'https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2.tar.gz',
'dest' : 'src/main/assets/models'
],
[
......
deploy/android_demo/app/src/main/AndroidManifest.xml
浏览文件 @ 4b214948
......@@ -14,7 +14,6 @@
android:roundIcon="@mipmap/ic_launcher_round"
android:supportsRtl="true"
android:theme="@style/AppTheme">
<!-- to test MiniActivity, change this to com.baidu.paddle.lite.demo.ocr.MiniActivity -->
<activity android:name="com.baidu.paddle.lite.demo.ocr.MainActivity">
<intent-filter>
<action android:name="android.intent.action.MAIN"/>
......
deploy/android_demo/app/src/main/cpp/native.cpp
浏览文件 @ 4b214948
......@@ -13,7 +13,7 @@ static paddle::lite_api::PowerMode str_to_cpu_mode(const std::string &cpu_mode);
extern "C" JNIEXPORT jlong JNICALL
Java_com_baidu_paddle_lite_demo_ocr_OCRPredictorNative_init(
JNIEnv *env, jobject thiz, jstring j_det_model_path,
jstring j_rec_model_path, jstring j_cls_model_path, jint j_thread_num,
jstring j_rec_model_path, jstring j_cls_model_path, jint j_use_opencl, jint j_thread_num,
jstring j_cpu_mode) {
std::string det_model_path = jstring_to_cpp_string(env, j_det_model_path);
std::string rec_model_path = jstring_to_cpp_string(env, j_rec_model_path);
......@@ -21,6 +21,7 @@ Java_com_baidu_paddle_lite_demo_ocr_OCRPredictorNative_init(
int thread_num = j_thread_num;
std::string cpu_mode = jstring_to_cpp_string(env, j_cpu_mode);
ppredictor::OCR_Config conf;
conf.use_opencl = j_use_opencl;
conf.thread_num = thread_num;
conf.mode = str_to_cpu_mode(cpu_mode);
ppredictor::OCR_PPredictor *orc_predictor =
......@@ -57,32 +58,31 @@ str_to_cpu_mode(const std::string &cpu_mode) {
extern "C" JNIEXPORT jfloatArray JNICALL
Java_com_baidu_paddle_lite_demo_ocr_OCRPredictorNative_forward(
JNIEnv *env, jobject thiz, jlong java_pointer, jfloatArray buf,
jfloatArray ddims, jobject original_image) {
JNIEnv *env, jobject thiz, jlong java_pointer, jobject original_image,jint j_max_size_len, jint j_run_det, jint j_run_cls, jint j_run_rec) {
LOGI("begin to run native forward");
if (java_pointer == 0) {
LOGE("JAVA pointer is NULL");
return cpp_array_to_jfloatarray(env, nullptr, 0);
}
cv::Mat origin = bitmap_to_cv_mat(env, original_image);
if (origin.size == 0) {
LOGE("origin bitmap cannot convert to CV Mat");
return cpp_array_to_jfloatarray(env, nullptr, 0);
}
int max_size_len = j_max_size_len;
int run_det = j_run_det;
int run_cls = j_run_cls;
int run_rec = j_run_rec;
ppredictor::OCR_PPredictor *ppredictor =
(ppredictor::OCR_PPredictor *)java_pointer;
std::vector<float> dims_float_arr = jfloatarray_to_float_vector(env, ddims);
std::vector<int64_t> dims_arr;
dims_arr.resize(dims_float_arr.size());
std::copy(dims_float_arr.cbegin(), dims_float_arr.cend(), dims_arr.begin());
// 这里值有点大，就不调用jfloatarray_to_float_vector了
int64_t buf_len = (int64_t)env->GetArrayLength(buf);
jfloat *buf_data = env->GetFloatArrayElements(buf, JNI_FALSE);
float *data = (jfloat *)buf_data;
std::vector<ppredictor::OCRPredictResult> results =
ppredictor->infer_ocr(dims_arr, data, buf_len, NET_OCR, origin);
ppredictor->infer_ocr(origin, max_size_len, run_det, run_cls, run_rec);
LOGI("infer_ocr finished with boxes %ld", results.size());
// 这里将std::vector<ppredictor::OCRPredictResult> 序列化成
// float数组，传输到java层再反序列化
std::vector<float> float_arr;
......@@ -90,13 +90,18 @@ Java_com_baidu_paddle_lite_demo_ocr_OCRPredictorNative_forward(
float_arr.push_back(r.points.size());
float_arr.push_back(r.word_index.size());
float_arr.push_back(r.score);
// add det point
for (const std::vector<int> &point : r.points) {
float_arr.push_back(point.at(0));
float_arr.push_back(point.at(1));
}
// add rec word idx
for (int index : r.word_index) {
float_arr.push_back(index);
}
// add cls result
float_arr.push_back(r.cls_label);
float_arr.push_back(r.cls_score);
}
return cpp_array_to_jfloatarray(env, float_arr.data(), float_arr.size());
}
......
deploy/android_demo/app/src/main/cpp/ocr_ppredictor.cpp
浏览文件 @ 4b214948
......@@ -17,15 +17,15 @@ int OCR_PPredictor::init(const std::string &det_model_content,
const std::string &rec_model_content,
const std::string &cls_model_content) {
_det_predictor = std::unique_ptr<PPredictor>(
new PPredictor{_config.thread_num, NET_OCR, _config.mode});
new PPredictor{_config.use_opencl,_config.thread_num, NET_OCR, _config.mode});
_det_predictor->init_nb(det_model_content);
_rec_predictor = std::unique_ptr<PPredictor>(
new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
new PPredictor{_config.use_opencl,_config.thread_num, NET_OCR_INTERNAL, _config.mode});
_rec_predictor->init_nb(rec_model_content);
_cls_predictor = std::unique_ptr<PPredictor>(
new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
new PPredictor{_config.use_opencl,_config.thread_num, NET_OCR_INTERNAL, _config.mode});
_cls_predictor->init_nb(cls_model_content);
return RETURN_OK;
}
......@@ -34,15 +34,16 @@ int OCR_PPredictor::init_from_file(const std::string &det_model_path,
const std::string &rec_model_path,
const std::string &cls_model_path) {
_det_predictor = std::unique_ptr<PPredictor>(
new PPredictor{_config.thread_num, NET_OCR, _config.mode});
new PPredictor{_config.use_opencl, _config.thread_num, NET_OCR, _config.mode});
_det_predictor->init_from_file(det_model_path);
_rec_predictor = std::unique_ptr<PPredictor>(
new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
new PPredictor{_config.use_opencl,_config.thread_num, NET_OCR_INTERNAL, _config.mode});
_rec_predictor->init_from_file(rec_model_path);
_cls_predictor = std::unique_ptr<PPredictor>(
new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
new PPredictor{_config.use_opencl,_config.thread_num, NET_OCR_INTERNAL, _config.mode});
_cls_predictor->init_from_file(cls_model_path);
return RETURN_OK;
}
......@@ -77,33 +78,126 @@ visual_img(const std::vector<std::vector<std::vector<int>>> &filter_boxes,
}
std::vector<OCRPredictResult>
OCR_PPredictor::infer_ocr(const std::vector<int64_t> &dims,
const float *input_data, int input_len, int net_flag,
cv::Mat &origin) {
OCR_PPredictor::infer_ocr(cv::Mat &origin,int max_size_len, int run_det, int run_cls, int run_rec) {
LOGI("ocr cpp start *****************");
LOGI("ocr cpp det: %d, cls: %d, rec: %d", run_det, run_cls, run_rec);
std::vector<OCRPredictResult> ocr_results;
if(run_det){
infer_det(origin, max_size_len, ocr_results);
}
if(run_rec){
if(ocr_results.size()==0){
OCRPredictResult res;
ocr_results.emplace_back(std::move(res));
}
for(int i = 0; i < ocr_results.size();i++) {
infer_rec(origin, run_cls, ocr_results[i]);
}
}else if(run_cls){
ClsPredictResult cls_res = infer_cls(origin);
OCRPredictResult res;
res.cls_score = cls_res.cls_score;
res.cls_label = cls_res.cls_label;
ocr_results.push_back(res);
}
LOGI("ocr cpp end *****************");
return ocr_results;
}
cv::Mat DetResizeImg(const cv::Mat img, int max_size_len,
std::vector<float> &ratio_hw) {
int w = img.cols;
int h = img.rows;
float ratio = 1.f;
int max_wh = w >= h ? w : h;
if (max_wh > max_size_len) {
if (h > w) {
ratio = static_cast<float>(max_size_len) / static_cast<float>(h);
} else {
ratio = static_cast<float>(max_size_len) / static_cast<float>(w);
}
}
int resize_h = static_cast<int>(float(h) * ratio);
int resize_w = static_cast<int>(float(w) * ratio);
if (resize_h % 32 == 0)
resize_h = resize_h;
else if (resize_h / 32 < 1 + 1e-5)
resize_h = 32;
else
resize_h = (resize_h / 32 - 1) * 32;
if (resize_w % 32 == 0)
resize_w = resize_w;
else if (resize_w / 32 < 1 + 1e-5)
resize_w = 32;
else
resize_w = (resize_w / 32 - 1) * 32;
cv::Mat resize_img;
cv::resize(img, resize_img, cv::Size(resize_w, resize_h));
ratio_hw.push_back(static_cast<float>(resize_h) / static_cast<float>(h));
ratio_hw.push_back(static_cast<float>(resize_w) / static_cast<float>(w));
return resize_img;
}
void OCR_PPredictor::infer_det(cv::Mat &origin, int max_size_len, std::vector<OCRPredictResult> &ocr_results) {
std::vector<float> mean = {0.485f, 0.456f, 0.406f};
std::vector<float> scale = {1 / 0.229f, 1 / 0.224f, 1 / 0.225f};
PredictorInput input = _det_predictor->get_first_input();
input.set_dims(dims);
input.set_data(input_data, input_len);
std::vector<float> ratio_hw;
cv::Mat input_image = DetResizeImg(origin, max_size_len, ratio_hw);
input_image.convertTo(input_image, CV_32FC3, 1 / 255.0f);
const float *dimg = reinterpret_cast<const float *>(input_image.data);
int input_size = input_image.rows * input_image.cols;
input.set_dims({1, 3, input_image.rows, input_image.cols});
neon_mean_scale(dimg, input.get_mutable_float_data(), input_size, mean,
scale);
LOGI("ocr cpp det shape %d,%d", input_image.rows,input_image.cols);
std::vector<PredictorOutput> results = _det_predictor->infer();
PredictorOutput &res = results.at(0);
std::vector<std::vector<std::vector<int>>> filtered_box = calc_filtered_boxes(
res.get_float_data(), res.get_size(), (int)dims[2], (int)dims[3], origin);
LOGI("Filter_box size %ld", filtered_box.size());
return infer_rec(filtered_box, origin);
res.get_float_data(), res.get_size(), input_image.rows, input_image.cols, origin);
LOGI("ocr cpp det Filter_box size %ld", filtered_box.size());
for(int i = 0;i<filtered_box.size();i++){
LOGI("ocr cpp box %d,%d,%d,%d,%d,%d,%d,%d", filtered_box[i][0][0],filtered_box[i][0][1], filtered_box[i][1][0],filtered_box[i][1][1], filtered_box[i][2][0],filtered_box[i][2][1], filtered_box[i][3][0],filtered_box[i][3][1]);
OCRPredictResult res;
res.points = filtered_box[i];
ocr_results.push_back(res);
}
}
std::vector<OCRPredictResult> OCR_PPredictor::infer_rec(
const std::vector<std::vector<std::vector<int>>> &boxes,
const cv::Mat &origin_img) {
void OCR_PPredictor::infer_rec(const cv::Mat &origin_img, int run_cls, OCRPredictResult& ocr_result) {
std::vector<float> mean = {0.5f, 0.5f, 0.5f};
std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
std::vector<int64_t> dims = {1, 3, 0, 0};
std::vector<OCRPredictResult> ocr_results;
PredictorInput input = _rec_predictor->get_first_input();
for (auto bp = boxes.crbegin(); bp != boxes.crend(); ++bp) {
const std::vector<std::vector<int>> &box = *bp;
cv::Mat crop_img = get_rotate_crop_image(origin_img, box);
crop_img = infer_cls(crop_img);
const std::vector<std::vector<int>> &box = ocr_result.points;
cv::Mat crop_img;
if(box.size()>0){
crop_img = get_rotate_crop_image(origin_img, box);
}
else{
crop_img = origin_img;
}
if(run_cls){
ClsPredictResult cls_res = infer_cls(crop_img);
crop_img = cls_res.img;
ocr_result.cls_score = cls_res.cls_score;
ocr_result.cls_label = cls_res.cls_label;
}
float wh_ratio = float(crop_img.cols) / float(crop_img.rows);
cv::Mat input_image = crnn_resize_img(crop_img, wh_ratio);
......@@ -122,8 +216,6 @@ std::vector<OCRPredictResult> OCR_PPredictor::infer_rec(
const float *predict_batch = results.at(0).get_float_data();
const std::vector<int64_t> predict_shape = results.at(0).get_shape();
OCRPredictResult res;
// ctc decode
int argmax_idx;
int last_index = 0;
......@@ -140,27 +232,19 @@ std::vector<OCRPredictResult> OCR_PPredictor::infer_rec(
if (argmax_idx > 0 && (!(n > 0 && argmax_idx == last_index))) {
score += max_value;
count += 1;
res.word_index.push_back(argmax_idx);
ocr_result.word_index.push_back(argmax_idx);
}
last_index = argmax_idx;
}
score /= count;
if (res.word_index.empty()) {
continue;
}
res.score = score;
res.points = box;
ocr_results.emplace_back(std::move(res));
}
LOGI("ocr_results finished %lu", ocr_results.size());
return ocr_results;
ocr_result.score = score;
LOGI("ocr cpp rec word size %ld", count);
}
cv::Mat OCR_PPredictor::infer_cls(const cv::Mat &img, float thresh) {
ClsPredictResult OCR_PPredictor::infer_cls(const cv::Mat &img, float thresh) {
std::vector<float> mean = {0.5f, 0.5f, 0.5f};
std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
std::vector<int64_t> dims = {1, 3, 0, 0};
std::vector<OCRPredictResult> ocr_results;
PredictorInput input = _cls_predictor->get_first_input();
......@@ -182,7 +266,7 @@ cv::Mat OCR_PPredictor::infer_cls(const cv::Mat &img, float thresh) {
float score = 0;
int label = 0;
for (int64_t i = 0; i < results.at(0).get_size(); i++) {
LOGI("output scores [%f]", scores[i]);
LOGI("ocr cpp cls output scores [%f]", scores[i]);
if (scores[i] > score) {
score = scores[i];
label = i;
......@@ -193,7 +277,12 @@ cv::Mat OCR_PPredictor::infer_cls(const cv::Mat &img, float thresh) {
if (label % 2 == 1 && score > thresh) {
cv::rotate(srcimg, srcimg, 1);
}
return srcimg;
ClsPredictResult res;
res.cls_label = label;
res.cls_score = score;
res.img = srcimg;
LOGI("ocr cpp cls word cls %ld, %f", label, score);
return res;
}
std::vector<std::vector<std::vector<int>>>
......
deploy/android_demo/app/src/main/cpp/ocr_ppredictor.h
浏览文件 @ 4b214948
......@@ -15,6 +15,7 @@ namespace ppredictor {
* Config
*/
struct OCR_Config {
int use_opencl = 0;
int thread_num = 4; // Thread num
paddle::lite_api::PowerMode mode =
paddle::lite_api::LITE_POWER_HIGH; // PaddleLite Mode
......@@ -27,8 +28,15 @@ struct OCRPredictResult {
std::vector<int> word_index;
std::vector<std::vector<int>> points;
float score;
float cls_score;
int cls_label=-1;
};
struct ClsPredictResult {
float cls_score;
int cls_label=-1;
cv::Mat img;
};
/**
* OCR there are 2 models
* 1. First model（det），select polygones to show where are the texts
......@@ -62,8 +70,7 @@ public:
* @return
*/
virtual std::vector<OCRPredictResult>
infer_ocr(const std::vector<int64_t> &dims, const float *input_data,
int input_len, int net_flag, cv::Mat &origin);
infer_ocr(cv::Mat &origin, int max_size_len, int run_det, int run_cls, int run_rec);
virtual NET_TYPE get_net_flag() const;
......@@ -80,16 +87,17 @@ private:
calc_filtered_boxes(const float *pred, int pred_size, int output_height,
int output_width, const cv::Mat &origin);
void
infer_det(cv::Mat &origin, int max_side_len, std::vector<OCRPredictResult>& ocr_results);
/**
* infer for second model
* infer for rec model
*
* @param boxes
* @param origin
* @return
*/
std::vector<OCRPredictResult>
infer_rec(const std::vector<std::vector<std::vector<int>>> &boxes,
const cv::Mat &origin);
void
infer_rec(const cv::Mat &origin, int run_cls, OCRPredictResult& ocr_result);
/**
* infer for cls model
......@@ -98,7 +106,7 @@ private:
* @param origin
* @return
*/
cv::Mat infer_cls(const cv::Mat &origin, float thresh = 0.9);
ClsPredictResult infer_cls(const cv::Mat &origin, float thresh = 0.9);
/**
* Postprocess or sencod model to extract text
......
deploy/android_demo/app/src/main/cpp/ppredictor.cpp
浏览文件 @ 4b214948
......@@ -2,9 +2,9 @@
#include "common.h"
namespace ppredictor {
PPredictor::PPredictor(int thread_num, int net_flag,
PPredictor::PPredictor(int use_opencl, int thread_num, int net_flag,
paddle::lite_api::PowerMode mode)
: _thread_num(thread_num), _net_flag(net_flag), _mode(mode) {}
: _use_opencl(use_opencl), _thread_num(thread_num), _net_flag(net_flag), _mode(mode) {}
int PPredictor::init_nb(const std::string &model_content) {
paddle::lite_api::MobileConfig config;
......@@ -19,10 +19,40 @@ int PPredictor::init_from_file(const std::string &model_content) {
}
template <typename ConfigT> int PPredictor::_init(ConfigT &config) {
bool is_opencl_backend_valid = paddle::lite_api::IsOpenCLBackendValid(/*check_fp16_valid = false*/);
if (is_opencl_backend_valid) {
if (_use_opencl != 0) {
// Make sure you have write permission of the binary path.
// We strongly recommend each model has a unique binary name.
const std::string bin_path = "/data/local/tmp/";
const std::string bin_name = "lite_opencl_kernel.bin";
config.set_opencl_binary_path_name(bin_path, bin_name);
// opencl tune option
// CL_TUNE_NONE: 0
// CL_TUNE_RAPID: 1
// CL_TUNE_NORMAL: 2
// CL_TUNE_EXHAUSTIVE: 3
const std::string tuned_path = "/data/local/tmp/";
const std::string tuned_name = "lite_opencl_tuned.bin";
config.set_opencl_tune(paddle::lite_api::CL_TUNE_NORMAL, tuned_path, tuned_name);
// opencl precision option
// CL_PRECISION_AUTO: 0, first fp16 if valid, default
// CL_PRECISION_FP32: 1, force fp32
// CL_PRECISION_FP16: 2, force fp16
config.set_opencl_precision(paddle::lite_api::CL_PRECISION_FP32);
LOGI("ocr cpp device: running on gpu.");
}
} else {
LOGI("ocr cpp device: running on cpu.");
// you can give backup cpu nb model instead
// config.set_model_from_file(cpu_nb_model_dir);
}
config.set_threads(_thread_num);
config.set_power_mode(_mode);
_predictor = paddle::lite_api::CreatePaddlePredictor(config);
LOGI("paddle instance created");
LOGI("ocr cpp paddle instance created");
return RETURN_OK;
}
......@@ -43,18 +73,18 @@ std::vector<PredictorInput> PPredictor::get_inputs(int num) {
PredictorInput PPredictor::get_first_input() { return get_input(0); }
std::vector<PredictorOutput> PPredictor::infer() {
LOGI("infer Run start %d", _net_flag);
LOGI("ocr cpp infer Run start %d", _net_flag);
std::vector<PredictorOutput> results;
if (!_is_input_get) {
return results;
}
_predictor->Run();
LOGI("infer Run end");
LOGI("ocr cpp infer Run end");
for (int i = 0; i < _predictor->GetOutputNames().size(); i++) {
std::unique_ptr<const paddle::lite_api::Tensor> output_tensor =
_predictor->GetOutput(i);
LOGI("output tensor[%d] size %ld", i, product(output_tensor->shape()));
LOGI("ocr cpp output tensor[%d] size %ld", i, product(output_tensor->shape()));
PredictorOutput result{std::move(output_tensor), i, _net_flag};
results.emplace_back(std::move(result));
}
......
deploy/android_demo/app/src/main/cpp/ppredictor.h
浏览文件 @ 4b214948
......@@ -22,7 +22,7 @@ public:
class PPredictor : public PPredictor_Interface {
public:
PPredictor(
int thread_num, int net_flag = 0,
int use_opencl, int thread_num, int net_flag = 0,
paddle::lite_api::PowerMode mode = paddle::lite_api::LITE_POWER_HIGH);
virtual ~PPredictor() {}
......@@ -54,6 +54,7 @@ protected:
template <typename ConfigT> int _init(ConfigT &config);
private:
int _use_opencl;
int _thread_num;
paddle::lite_api::PowerMode _mode;
std::shared_ptr<paddle::lite_api::PaddlePredictor> _predictor;
......
deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/MainActivity.java
浏览文件 @ 4b214948
......@@ -13,6 +13,7 @@ import android.graphics.BitmapFactory;
import android.graphics.drawable.BitmapDrawable;
import android.media.ExifInterface;
import android.content.res.AssetManager;
import android.media.FaceDetector;
import android.net.Uri;
import android.os.Bundle;
import android.os.Environment;
......@@ -27,7 +28,9 @@ import android.view.Menu;
import android.view.MenuInflater;
import android.view.MenuItem;
import android.view.View;
import android.widget.CheckBox;
import android.widget.ImageView;
import android.widget.Spinner;
import android.widget.TextView;
import android.widget.Toast;
......@@ -68,23 +71,24 @@ public class MainActivity extends AppCompatActivity {
protected ImageView ivInputImage;
protected TextView tvOutputResult;
protected TextView tvInferenceTime;
protected CheckBox cbOpencl;
protected Spinner spRunMode;
// Model settings of object detection
// Model settings of ocr
protected String modelPath = "";
protected String labelPath = "";
protected String imagePath = "";
protected int cpuThreadNum = 1;
protected String cpuPowerMode = "";
protected String inputColorFormat = "";
protected long[] inputShape = new long[]{};
protected float[] inputMean = new float[]{};
protected float[] inputStd = new float[]{};
protected int detLongSize = 960;
protected float scoreThreshold = 0.1f;
private String currentPhotoPath;
private AssetManager assetManager =null;
private AssetManager assetManager = null;
protected Predictor predictor = new Predictor();
private Bitmap cur_predict_image = null;
@Override
protected void onCreate(Bundle savedInstanceState) {
super.onCreate(savedInstanceState);
......@@ -98,10 +102,12 @@ public class MainActivity extends AppCompatActivity {
// Setup the UI components
tvInputSetting = findViewById(R.id.tv_input_setting);
cbOpencl = findViewById(R.id.cb_opencl);
tvStatus = findViewById(R.id.tv_model_img_status);
ivInputImage = findViewById(R.id.iv_input_image);
tvInferenceTime = findViewById(R.id.tv_inference_time);
tvOutputResult = findViewById(R.id.tv_output_result);
spRunMode = findViewById(R.id.sp_run_mode);
tvInputSetting.setMovementMethod(ScrollingMovementMethod.getInstance());
tvOutputResult.setMovementMethod(ScrollingMovementMethod.getInstance());
......@@ -111,26 +117,26 @@ public class MainActivity extends AppCompatActivity {
public void handleMessage(Message msg) {
switch (msg.what) {
case RESPONSE_LOAD_MODEL_SUCCESSED:
if(pbLoadModel!=null && pbLoadModel.isShowing()){
if (pbLoadModel != null && pbLoadModel.isShowing()) {
pbLoadModel.dismiss();
}
onLoadModelSuccessed();
break;
case RESPONSE_LOAD_MODEL_FAILED:
if(pbLoadModel!=null && pbLoadModel.isShowing()){
if (pbLoadModel != null && pbLoadModel.isShowing()) {
pbLoadModel.dismiss();
}
Toast.makeText(MainActivity.this, "Load model failed!", Toast.LENGTH_SHORT).show();
onLoadModelFailed();
break;
case RESPONSE_RUN_MODEL_SUCCESSED:
if(pbRunModel!=null && pbRunModel.isShowing()){
if (pbRunModel != null && pbRunModel.isShowing()) {
pbRunModel.dismiss();
}
onRunModelSuccessed();
break;
case RESPONSE_RUN_MODEL_FAILED:
if(pbRunModel!=null && pbRunModel.isShowing()){
if (pbRunModel != null && pbRunModel.isShowing()) {
pbRunModel.dismiss();
}
Toast.makeText(MainActivity.this, "Run model failed!", Toast.LENGTH_SHORT).show();
......@@ -175,71 +181,47 @@ public class MainActivity extends AppCompatActivity {
super.onResume();
SharedPreferences sharedPreferences = PreferenceManager.getDefaultSharedPreferences(this);
boolean settingsChanged = false;
boolean model_settingsChanged = false;
String model_path = sharedPreferences.getString(getString(R.string.MODEL_PATH_KEY),
getString(R.string.MODEL_PATH_DEFAULT));
String label_path = sharedPreferences.getString(getString(R.string.LABEL_PATH_KEY),
getString(R.string.LABEL_PATH_DEFAULT));
String image_path = sharedPreferences.getString(getString(R.string.IMAGE_PATH_KEY),
getString(R.string.IMAGE_PATH_DEFAULT));
settingsChanged |= !model_path.equalsIgnoreCase(modelPath);
model_settingsChanged |= !model_path.equalsIgnoreCase(modelPath);
settingsChanged |= !label_path.equalsIgnoreCase(labelPath);
settingsChanged |= !image_path.equalsIgnoreCase(imagePath);
int cpu_thread_num = Integer.parseInt(sharedPreferences.getString(getString(R.string.CPU_THREAD_NUM_KEY),
getString(R.string.CPU_THREAD_NUM_DEFAULT)));
settingsChanged |= cpu_thread_num != cpuThreadNum;
model_settingsChanged |= cpu_thread_num != cpuThreadNum;
String cpu_power_mode =
sharedPreferences.getString(getString(R.string.CPU_POWER_MODE_KEY),
getString(R.string.CPU_POWER_MODE_DEFAULT));
settingsChanged |= !cpu_power_mode.equalsIgnoreCase(cpuPowerMode);
String input_color_format =
sharedPreferences.getString(getString(R.string.INPUT_COLOR_FORMAT_KEY),
getString(R.string.INPUT_COLOR_FORMAT_DEFAULT));
settingsChanged |= !input_color_format.equalsIgnoreCase(inputColorFormat);
long[] input_shape =
Utils.parseLongsFromString(sharedPreferences.getString(getString(R.string.INPUT_SHAPE_KEY),
getString(R.string.INPUT_SHAPE_DEFAULT)), ",");
float[] input_mean =
Utils.parseFloatsFromString(sharedPreferences.getString(getString(R.string.INPUT_MEAN_KEY),
getString(R.string.INPUT_MEAN_DEFAULT)), ",");
float[] input_std =
Utils.parseFloatsFromString(sharedPreferences.getString(getString(R.string.INPUT_STD_KEY)
, getString(R.string.INPUT_STD_DEFAULT)), ",");
settingsChanged |= input_shape.length != inputShape.length;
settingsChanged |= input_mean.length != inputMean.length;
settingsChanged |= input_std.length != inputStd.length;
if (!settingsChanged) {
for (int i = 0; i < input_shape.length; i++) {
settingsChanged |= input_shape[i] != inputShape[i];
}
for (int i = 0; i < input_mean.length; i++) {
settingsChanged |= input_mean[i] != inputMean[i];
}
for (int i = 0; i < input_std.length; i++) {
settingsChanged |= input_std[i] != inputStd[i];
}
}
model_settingsChanged |= !cpu_power_mode.equalsIgnoreCase(cpuPowerMode);
int det_long_size = Integer.parseInt(sharedPreferences.getString(getString(R.string.DET_LONG_SIZE_KEY),
getString(R.string.DET_LONG_SIZE_DEFAULT)));
settingsChanged |= det_long_size != detLongSize;
float score_threshold =
Float.parseFloat(sharedPreferences.getString(getString(R.string.SCORE_THRESHOLD_KEY),
getString(R.string.SCORE_THRESHOLD_DEFAULT)));
settingsChanged |= scoreThreshold != score_threshold;
if (settingsChanged) {
modelPath = model_path;
labelPath = label_path;
imagePath = image_path;
detLongSize = det_long_size;
scoreThreshold = score_threshold;
set_img();
}
if (model_settingsChanged) {
modelPath = model_path;
cpuThreadNum = cpu_thread_num;
cpuPowerMode = cpu_power_mode;
inputColorFormat = input_color_format;
inputShape = input_shape;
inputMean = input_mean;
inputStd = input_std;
scoreThreshold = score_threshold;
// Update UI
tvInputSetting.setText("Model: " + modelPath.substring(modelPath.lastIndexOf("/") + 1) + "\n" + "CPU" +
" Thread Num: " + Integer.toString(cpuThreadNum) + "\n" + "CPU Power Mode: " + cpuPowerMode);
tvInputSetting.setText("Model: " + modelPath.substring(modelPath.lastIndexOf("/") + 1) + "\nOPENCL: " + cbOpencl.isChecked() + "\nCPU Thread Num: " + cpuThreadNum + "\nCPU Power Mode: " + cpuPowerMode);
tvInputSetting.scrollTo(0, 0);
// Reload model if configure has been changed
// loadModel();
set_img();
loadModel();
}
}
......@@ -254,20 +236,28 @@ public class MainActivity extends AppCompatActivity {
}
public boolean onLoadModel() {
return predictor.init(MainActivity.this, modelPath, labelPath, cpuThreadNum,
if (predictor.isLoaded()) {
predictor.releaseModel();
}
return predictor.init(MainActivity.this, modelPath, labelPath, cbOpencl.isChecked() ? 1 : 0, cpuThreadNum,
cpuPowerMode,
inputColorFormat,
inputShape, inputMean,
inputStd, scoreThreshold);
detLongSize, scoreThreshold);
}
public boolean onRunModel() {
return predictor.isLoaded() && predictor.runModel();
String run_mode = spRunMode.getSelectedItem().toString();
int run_det = run_mode.contains("检测") ? 1 : 0;
int run_cls = run_mode.contains("分类") ? 1 : 0;
int run_rec = run_mode.contains("识别") ? 1 : 0;
return predictor.isLoaded() && predictor.runModel(run_det, run_cls, run_rec);
}
public void onLoadModelSuccessed() {
// Load test image from path and run model
tvInputSetting.setText("Model: " + modelPath.substring(modelPath.lastIndexOf("/") + 1) + "\nOPENCL: " + cbOpencl.isChecked() + "\nCPU Thread Num: " + cpuThreadNum + "\nCPU Power Mode: " + cpuPowerMode);
tvInputSetting.scrollTo(0, 0);
tvStatus.setText("STATUS: load model successed");
}
public void onLoadModelFailed() {
......@@ -290,20 +280,13 @@ public class MainActivity extends AppCompatActivity {
tvStatus.setText("STATUS: run model failed");
}
public void onImageChanged(Bitmap image) {
// Rerun model if users pick test image from gallery or camera
if (image != null && predictor.isLoaded()) {
predictor.setInputImage(image);
runModel();
}
}
public void set_img() {
// Load test image from path and run model
try {
assetManager= getAssets();
InputStream in=assetManager.open(imagePath);
Bitmap bmp=BitmapFactory.decodeStream(in);
assetManager = getAssets();
InputStream in = assetManager.open(imagePath);
Bitmap bmp = BitmapFactory.decodeStream(in);
cur_predict_image = bmp;
ivInputImage.setImageBitmap(bmp);
} catch (IOException e) {
Toast.makeText(MainActivity.this, "Load image failed!", Toast.LENGTH_SHORT).show();
......@@ -430,7 +413,7 @@ public class MainActivity extends AppCompatActivity {
Cursor cursor = managedQuery(uri, proj, null, null, null);
cursor.moveToFirst();
if (image != null) {
// onImageChanged(image);
cur_predict_image = image;
ivInputImage.setImageBitmap(image);
}
} catch (IOException e) {
......@@ -451,7 +434,7 @@ public class MainActivity extends AppCompatActivity {
Bitmap image = BitmapFactory.decodeFile(currentPhotoPath);
image = Utils.rotateBitmap(image, orientation);
if (image != null) {
// onImageChanged(image);
cur_predict_image = image;
ivInputImage.setImageBitmap(image);
}
} else {
......@@ -464,28 +447,28 @@ public class MainActivity extends AppCompatActivity {
}
}
public void btn_load_model_click(View view) {
if (predictor.isLoaded()){
tvStatus.setText("STATUS: model has been loaded");
}else{
public void btn_reset_img_click(View view) {
ivInputImage.setImageBitmap(cur_predict_image);
}
public void cb_opencl_click(View view) {
tvStatus.setText("STATUS: load model ......");
loadModel();
}
}
public void btn_run_model_click(View view) {
Bitmap image =((BitmapDrawable)ivInputImage.getDrawable()).getBitmap();
if(image == null) {
Bitmap image = ((BitmapDrawable) ivInputImage.getDrawable()).getBitmap();
if (image == null) {
tvStatus.setText("STATUS: image is not exists");
}
else if (!predictor.isLoaded()){
} else if (!predictor.isLoaded()) {
tvStatus.setText("STATUS: model is not loaded");
}else{
} else {
tvStatus.setText("STATUS: run model ...... ");
predictor.setInputImage(image);
runModel();
}
}
public void btn_choice_img_click(View view) {
if (requestAllPermissions()) {
openGallery();
......@@ -506,4 +489,32 @@ public class MainActivity extends AppCompatActivity {
worker.quit();
super.onDestroy();
}
public int get_run_mode() {
String run_mode = spRunMode.getSelectedItem().toString();
int mode;
switch (run_mode) {
case "检测+分类+识别":
mode = 1;
break;
case "检测+识别":
mode = 2;
break;
case "识别+分类":
mode = 3;
break;
case "检测":
mode = 4;
break;
case "识别":
mode = 5;
break;
case "分类":
mode = 6;
break;
default:
mode = 1;
}
return mode;
}
}
deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/MiniActivity.java 已删除 100644 → 0
浏览文件 @ 917606b4
package com.baidu.paddle.lite.demo.ocr;
import android.graphics.Bitmap;
import android.graphics.BitmapFactory;
import android.os.Build;
import android.os.Bundle;
import android.os.Handler;
import android.os.HandlerThread;
import android.os.Message;
import android.util.Log;
import android.view.View;
import android.widget.Button;
import android.widget.ImageView;
import android.widget.TextView;
import android.widget.Toast;
import androidx.appcompat.app.AppCompatActivity;
import java.io.IOException;
import java.io.InputStream;
public class MiniActivity extends AppCompatActivity {
public static final int REQUEST_LOAD_MODEL = 0;
public static final int REQUEST_RUN_MODEL = 1;
public static final int REQUEST_UNLOAD_MODEL = 2;
public static final int RESPONSE_LOAD_MODEL_SUCCESSED = 0;
public static final int RESPONSE_LOAD_MODEL_FAILED = 1;
public static final int RESPONSE_RUN_MODEL_SUCCESSED = 2;
public static final int RESPONSE_RUN_MODEL_FAILED = 3;
private static final String TAG = "MiniActivity";
protected Handler receiver = null; // Receive messages from worker thread
protected Handler sender = null; // Send command to worker thread
protected HandlerThread worker = null; // Worker thread to load&run model
protected volatile Predictor predictor = null;
private String assetModelDirPath = "models/ocr_v2_for_cpu";
private String assetlabelFilePath = "labels/ppocr_keys_v1.txt";
private Button button;
private ImageView imageView; // image result
private TextView textView; // text result
@Override
protected void onCreate(Bundle savedInstanceState) {
super.onCreate(savedInstanceState);
setContentView(R.layout.activity_mini);
Log.i(TAG, "SHOW in Logcat");
// Prepare the worker thread for mode loading and inference
worker = new HandlerThread("Predictor Worker");
worker.start();
sender = new Handler(worker.getLooper()) {
public void handleMessage(Message msg) {
switch (msg.what) {
case REQUEST_LOAD_MODEL:
// Load model and reload test image
if (!onLoadModel()) {
runOnUiThread(new Runnable() {
@Override
public void run() {
Toast.makeText(MiniActivity.this, "Load model failed!", Toast.LENGTH_SHORT).show();
}
});
}
break;
case REQUEST_RUN_MODEL:
// Run model if model is loaded
final boolean isSuccessed = onRunModel();
runOnUiThread(new Runnable() {
@Override
public void run() {
if (isSuccessed){
onRunModelSuccessed();
}else{
Toast.makeText(MiniActivity.this, "Run model failed!", Toast.LENGTH_SHORT).show();
}
}
});
break;
}
}
};
sender.sendEmptyMessage(REQUEST_LOAD_MODEL); // corresponding to REQUEST_LOAD_MODEL， to call onLoadModel()
imageView = findViewById(R.id.imageView);
textView = findViewById(R.id.sample_text);
button = findViewById(R.id.button);
button.setOnClickListener(new View.OnClickListener() {
@Override
public void onClick(View v) {
sender.sendEmptyMessage(REQUEST_RUN_MODEL);
}
});
}
@Override
protected void onDestroy() {
onUnloadModel();
if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.JELLY_BEAN_MR2) {
worker.quitSafely();
} else {
worker.quit();
}
super.onDestroy();
}
/**
* call in onCreate, model init
*
* @return
*/
private boolean onLoadModel() {
if (predictor == null) {
predictor = new Predictor();
}
return predictor.init(this, assetModelDirPath, assetlabelFilePath);
}
/**
* init engine
* call in onCreate
*
* @return
*/
private boolean onRunModel() {
try {
String assetImagePath = "images/0.jpg";
InputStream imageStream = getAssets().open(assetImagePath);
Bitmap image = BitmapFactory.decodeStream(imageStream);
// Input is Bitmap
predictor.setInputImage(image);
return predictor.isLoaded() && predictor.runModel();
} catch (IOException e) {
e.printStackTrace();
return false;
}
}
private void onRunModelSuccessed() {
Log.i(TAG, "onRunModelSuccessed");
textView.setText(predictor.outputResult);
imageView.setImageBitmap(predictor.outputImage);
}
private void onUnloadModel() {
if (predictor != null) {
predictor.releaseModel();
}
}
}
deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/OCRPredictorNative.java
浏览文件 @ 4b214948
......@@ -29,22 +29,22 @@ public class OCRPredictorNative {
public OCRPredictorNative(Config config) {
this.config = config;
loadLibrary();
nativePointer = init(config.detModelFilename, config.recModelFilename,config.clsModelFilename,
nativePointer = init(config.detModelFilename, config.recModelFilename, config.clsModelFilename, config.useOpencl,
config.cpuThreadNum, config.cpuPower);
Log.i("OCRPredictorNative", "load success " + nativePointer);
}
public ArrayList<OcrResultModel> runImage(float[] inputData, int width, int height, int channels, Bitmap originalImage) {
Log.i("OCRPredictorNative", "begin to run image " + inputData.length + " " + width + " " + height);
float[] dims = new float[]{1, channels, height, width};
float[] rawResults = forward(nativePointer, inputData, dims, originalImage);
public ArrayList<OcrResultModel> runImage(Bitmap originalImage, int max_size_len, int run_det, int run_cls, int run_rec) {
Log.i("OCRPredictorNative", "begin to run image ");
float[] rawResults = forward(nativePointer, originalImage, max_size_len, run_det, run_cls, run_rec);
ArrayList<OcrResultModel> results = postprocess(rawResults);
return results;
}
public static class Config {
public int useOpencl;
public int cpuThreadNum;
public String cpuPower;
public String detModelFilename;
......@@ -53,16 +53,16 @@ public class OCRPredictorNative {
}
public void destory(){
public void destory() {
if (nativePointer > 0) {
release(nativePointer);
nativePointer = 0;
}
}
protected native long init(String detModelPath, String recModelPath,String clsModelPath, int threadNum, String cpuMode);
protected native long init(String detModelPath, String recModelPath, String clsModelPath, int useOpencl, int threadNum, String cpuMode);
protected native float[] forward(long pointer, float[] buf, float[] ddims, Bitmap originalImage);
protected native float[] forward(long pointer, Bitmap originalImage,int max_size_len, int run_det, int run_cls, int run_rec);
protected native void release(long pointer);
......@@ -73,9 +73,9 @@ public class OCRPredictorNative {
while (begin < raw.length) {
int point_num = Math.round(raw[begin]);
int word_num = Math.round(raw[begin + 1]);
OcrResultModel model = parse(raw, begin + 2, point_num, word_num);
begin += 2 + 1 + point_num * 2 + word_num;
results.add(model);
OcrResultModel res = parse(raw, begin + 2, point_num, word_num);
begin += 2 + 1 + point_num * 2 + word_num + 2;
results.add(res);
}
return results;
......@@ -83,19 +83,22 @@ public class OCRPredictorNative {
private OcrResultModel parse(float[] raw, int begin, int pointNum, int wordNum) {
int current = begin;
OcrResultModel model = new OcrResultModel();
model.setConfidence(raw[current]);
OcrResultModel res = new OcrResultModel();
res.setConfidence(raw[current]);
current++;
for (int i = 0; i < pointNum; i++) {
model.addPoints(Math.round(raw[current + i * 2]), Math.round(raw[current + i * 2 + 1]));
res.addPoints(Math.round(raw[current + i * 2]), Math.round(raw[current + i * 2 + 1]));
}
current += (pointNum * 2);
for (int i = 0; i < wordNum; i++) {
int index = Math.round(raw[current + i]);
model.addWordIndex(index);
res.addWordIndex(index);
}
current += wordNum;
res.setClsIdx(raw[current]);
res.setClsConfidence(raw[current + 1]);
Log.i("OCRPredictorNative", "word finished " + wordNum);
return model;
return res;
}
......
deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/OcrResultModel.java
浏览文件 @ 4b214948
......@@ -10,6 +10,9 @@ public class OcrResultModel {
private List<Integer> wordIndex;
private String label;
private float confidence;
private float cls_idx;
private String cls_label;
private float cls_confidence;
public OcrResultModel() {
super();
......@@ -49,4 +52,28 @@ public class OcrResultModel {
public void setConfidence(float confidence) {
this.confidence = confidence;
}
public float getClsIdx() {
return cls_idx;
}
public void setClsIdx(float idx) {
this.cls_idx = idx;
}
public String getClsLabel() {
return cls_label;
}
public void setClsLabel(String label) {
this.cls_label = label;
}
public float getClsConfidence() {
return cls_confidence;
}
public void setClsConfidence(float confidence) {
this.cls_confidence = confidence;
}
}
deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/Predictor.java
浏览文件 @ 4b214948
......@@ -31,23 +31,19 @@ public class Predictor {
protected float inferenceTime = 0;
// Only for object detection
protected Vector<String> wordLabels = new Vector<String>();
protected String inputColorFormat = "BGR";
protected long[] inputShape = new long[]{1, 3, 960};
protected float[] inputMean = new float[]{0.485f, 0.456f, 0.406f};
protected float[] inputStd = new float[]{1.0f / 0.229f, 1.0f / 0.224f, 1.0f / 0.225f};
protected int detLongSize = 960;
protected float scoreThreshold = 0.1f;
protected Bitmap inputImage = null;
protected Bitmap outputImage = null;
protected volatile String outputResult = "";
protected float preprocessTime = 0;
protected float postprocessTime = 0;
public Predictor() {
}
public boolean init(Context appCtx, String modelPath, String labelPath) {
isLoaded = loadModel(appCtx, modelPath, cpuThreadNum, cpuPowerMode);
public boolean init(Context appCtx, String modelPath, String labelPath, int useOpencl, int cpuThreadNum, String cpuPowerMode) {
isLoaded = loadModel(appCtx, modelPath, useOpencl, cpuThreadNum, cpuPowerMode);
if (!isLoaded) {
return false;
}
......@@ -56,49 +52,18 @@ public class Predictor {
}
public boolean init(Context appCtx, String modelPath, String labelPath, int cpuThreadNum, String cpuPowerMode,
String inputColorFormat,
long[] inputShape, float[] inputMean,
float[] inputStd, float scoreThreshold) {
if (inputShape.length != 3) {
Log.e(TAG, "Size of input shape should be: 3");
return false;
}
if (inputMean.length != inputShape[1]) {
Log.e(TAG, "Size of input mean should be: " + Long.toString(inputShape[1]));
return false;
}
if (inputStd.length != inputShape[1]) {
Log.e(TAG, "Size of input std should be: " + Long.toString(inputShape[1]));
return false;
}
if (inputShape[0] != 1) {
Log.e(TAG, "Only one batch is supported in the image classification demo, you can use any batch size in " +
"your Apps!");
return false;
}
if (inputShape[1] != 1 && inputShape[1] != 3) {
Log.e(TAG, "Only one/three channels are supported in the image classification demo, you can use any " +
"channel size in your Apps!");
return false;
}
if (!inputColorFormat.equalsIgnoreCase("BGR")) {
Log.e(TAG, "Only BGR color format is supported.");
return false;
}
boolean isLoaded = init(appCtx, modelPath, labelPath);
public boolean init(Context appCtx, String modelPath, String labelPath, int useOpencl, int cpuThreadNum, String cpuPowerMode,
int detLongSize, float scoreThreshold) {
boolean isLoaded = init(appCtx, modelPath, labelPath, useOpencl, cpuThreadNum, cpuPowerMode);
if (!isLoaded) {
return false;
}
this.inputColorFormat = inputColorFormat;
this.inputShape = inputShape;
this.inputMean = inputMean;
this.inputStd = inputStd;
this.detLongSize = detLongSize;
this.scoreThreshold = scoreThreshold;
return true;
}
protected boolean loadModel(Context appCtx, String modelPath, int cpuThreadNum, String cpuPowerMode) {
protected boolean loadModel(Context appCtx, String modelPath, int useOpencl, int cpuThreadNum, String cpuPowerMode) {
// Release model if exists
releaseModel();
......@@ -118,12 +83,13 @@ public class Predictor {
}
OCRPredictorNative.Config config = new OCRPredictorNative.Config();
config.useOpencl = useOpencl;
config.cpuThreadNum = cpuThreadNum;
config.detModelFilename = realPath + File.separator + "ch_ppocr_mobile_v2.0_det_opt.nb";
config.recModelFilename = realPath + File.separator + "ch_ppocr_mobile_v2.0_rec_opt.nb";
config.clsModelFilename = realPath + File.separator + "ch_ppocr_mobile_v2.0_cls_opt.nb";
Log.e("Predictor", "model path" + config.detModelFilename + " ; " + config.recModelFilename + ";" + config.clsModelFilename);
config.cpuPower = cpuPowerMode;
config.detModelFilename = realPath + File.separator + "det_db.nb";
config.recModelFilename = realPath + File.separator + "rec_crnn.nb";
config.clsModelFilename = realPath + File.separator + "cls.nb";
Log.i("Predictor", "model path" + config.detModelFilename + " ; " + config.recModelFilename + ";" + config.clsModelFilename);
paddlePredictor = new OCRPredictorNative(config);
this.cpuThreadNum = cpuThreadNum;
......@@ -170,82 +136,29 @@ public class Predictor {
}
public boolean runModel() {
public boolean runModel(int run_det, int run_cls, int run_rec) {
if (inputImage == null || !isLoaded()) {
return false;
}
// Pre-process image, and feed input tensor with pre-processed data
Bitmap scaleImage = Utils.resizeWithStep(inputImage, Long.valueOf(inputShape[2]).intValue(), 32);
Date start = new Date();
int channels = (int) inputShape[1];
int width = scaleImage.getWidth();
int height = scaleImage.getHeight();
float[] inputData = new float[channels * width * height];
if (channels == 3) {
int[] channelIdx = null;
if (inputColorFormat.equalsIgnoreCase("RGB")) {
channelIdx = new int[]{0, 1, 2};
} else if (inputColorFormat.equalsIgnoreCase("BGR")) {
channelIdx = new int[]{2, 1, 0};
} else {
Log.i(TAG, "Unknown color format " + inputColorFormat + ", only RGB and BGR color format is " +
"supported!");
return false;
}
int[] channelStride = new int[]{width * height, width * height * 2};
int[] pixels=new int[width*height];
scaleImage.getPixels(pixels,0,scaleImage.getWidth(),0,0,scaleImage.getWidth(),scaleImage.getHeight());
for (int i = 0; i < pixels.length; i++) {
int color = pixels[i];
float[] rgb = new float[]{(float) red(color) / 255.0f, (float) green(color) / 255.0f,
(float) blue(color) / 255.0f};
inputData[i] = (rgb[channelIdx[0]] - inputMean[0]) / inputStd[0];
inputData[i + channelStride[0]] = (rgb[channelIdx[1]] - inputMean[1]) / inputStd[1];
inputData[i+ channelStride[1]] = (rgb[channelIdx[2]] - inputMean[2]) / inputStd[2];
}
} else if (channels == 1) {
int[] pixels=new int[width*height];
scaleImage.getPixels(pixels,0,scaleImage.getWidth(),0,0,scaleImage.getWidth(),scaleImage.getHeight());
for (int i = 0; i < pixels.length; i++) {
int color = pixels[i];
float gray = (float) (red(color) + green(color) + blue(color)) / 3.0f / 255.0f;
inputData[i] = (gray - inputMean[0]) / inputStd[0];
}
} else {
Log.i(TAG, "Unsupported channel size " + Integer.toString(channels) + ", only channel 1 and 3 is " +
"supported!");
return false;
}
float[] pixels = inputData;
Log.i(TAG, "pixels " + pixels[0] + " " + pixels[1] + " " + pixels[2] + " " + pixels[3]
+ " " + pixels[pixels.length / 2] + " " + pixels[pixels.length / 2 + 1] + " " + pixels[pixels.length - 2] + " " + pixels[pixels.length - 1]);
Date end = new Date();
preprocessTime = (float) (end.getTime() - start.getTime());
// Warm up
for (int i = 0; i < warmupIterNum; i++) {
paddlePredictor.runImage(inputData, width, height, channels, inputImage);
paddlePredictor.runImage(inputImage, detLongSize, run_det, run_cls, run_rec);
}
warmupIterNum = 0; // do not need warm
// Run inference
start = new Date();
ArrayList<OcrResultModel> results = paddlePredictor.runImage(inputData, width, height, channels, inputImage);
end = new Date();
Date start = new Date();
ArrayList<OcrResultModel> results = paddlePredictor.runImage(inputImage, detLongSize, run_det, run_cls, run_rec);
Date end = new Date();
inferenceTime = (end.getTime() - start.getTime()) / (float) inferIterNum;
results = postprocess(results);
Log.i(TAG, "[stat] Preprocess Time: " + preprocessTime
+ " ; Inference Time: " + inferenceTime + " ;Box Size " + results.size());
Log.i(TAG, "[stat] Inference Time: " + inferenceTime + " ;Box Size " + results.size());
drawResults(results);
return true;
}
public boolean isLoaded() {
return paddlePredictor != null && isLoaded;
}
......@@ -282,10 +195,6 @@ public class Predictor {
return outputResult;
}
public float preprocessTime() {
return preprocessTime;
}
public float postprocessTime() {
return postprocessTime;
}
......@@ -310,6 +219,7 @@ public class Predictor {
}
}
r.setLabel(word.toString());
r.setClsLabel(r.getClsIdx() == 1 ? "180" : "0");
}
return results;
}
......@@ -319,14 +229,22 @@ public class Predictor {
for (int i = 0; i < results.size(); i++) {
OcrResultModel result = results.get(i);
StringBuilder sb = new StringBuilder("");
sb.append(result.getLabel());
sb.append(" ").append(result.getConfidence());
sb.append("; Points: ");
if(result.getPoints().size()>0){
sb.append("Det: ");
for (Point p : result.getPoints()) {
sb.append("(").append(p.x).append(",").append(p.y).append(") ");
}
}
if(result.getLabel().length() > 0){
sb.append("\n Rec: ").append(result.getLabel());
sb.append(",").append(result.getConfidence());
}
if(result.getClsIdx()!=-1){
sb.append(" Cls: ").append(result.getClsLabel());
sb.append(",").append(result.getClsConfidence());
}
Log.i(TAG, sb.toString()); // show LOG in Logcat panel
outputResultSb.append(i + 1).append(": ").append(result.getLabel()).append("\n");
outputResultSb.append(i + 1).append(": ").append(sb.toString()).append("\n");
}
outputResult = outputResultSb.toString();
outputImage = inputImage;
......@@ -344,6 +262,9 @@ public class Predictor {
for (OcrResultModel result : results) {
Path path = new Path();
List<Point> points = result.getPoints();
if(points.size()==0){
continue;
}
path.moveTo(points.get(0).x, points.get(0).y);
for (int i = points.size() - 1; i >= 0; i--) {
Point p = points.get(i);
......
deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/SettingsActivity.java
浏览文件 @ 4b214948
......@@ -20,16 +20,13 @@ public class SettingsActivity extends AppCompatPreferenceActivity implements Sha
ListPreference etImagePath = null;
ListPreference lpCPUThreadNum = null;
ListPreference lpCPUPowerMode = null;
ListPreference lpInputColorFormat = null;
EditTextPreference etInputShape = null;
EditTextPreference etInputMean = null;
EditTextPreference etInputStd = null;
EditTextPreference etDetLongSize = null;
EditTextPreference etScoreThreshold = null;
List<String> preInstalledModelPaths = null;
List<String> preInstalledLabelPaths = null;
List<String> preInstalledImagePaths = null;
List<String> preInstalledInputShapes = null;
List<String> preInstalledDetLongSizes = null;
List<String> preInstalledCPUThreadNums = null;
List<String> preInstalledCPUPowerModes = null;
List<String> preInstalledInputColorFormats = null;
......@@ -50,7 +47,7 @@ public class SettingsActivity extends AppCompatPreferenceActivity implements Sha
preInstalledModelPaths = new ArrayList<String>();
preInstalledLabelPaths = new ArrayList<String>();
preInstalledImagePaths = new ArrayList<String>();
preInstalledInputShapes = new ArrayList<String>();
preInstalledDetLongSizes = new ArrayList<String>();
preInstalledCPUThreadNums = new ArrayList<String>();
preInstalledCPUPowerModes = new ArrayList<String>();
preInstalledInputColorFormats = new ArrayList<String>();
......@@ -63,10 +60,7 @@ public class SettingsActivity extends AppCompatPreferenceActivity implements Sha
preInstalledImagePaths.add(getString(R.string.IMAGE_PATH_DEFAULT));
preInstalledCPUThreadNums.add(getString(R.string.CPU_THREAD_NUM_DEFAULT));
preInstalledCPUPowerModes.add(getString(R.string.CPU_POWER_MODE_DEFAULT));
preInstalledInputColorFormats.add(getString(R.string.INPUT_COLOR_FORMAT_DEFAULT));
preInstalledInputShapes.add(getString(R.string.INPUT_SHAPE_DEFAULT));
preInstalledInputMeans.add(getString(R.string.INPUT_MEAN_DEFAULT));
preInstalledInputStds.add(getString(R.string.INPUT_STD_DEFAULT));
preInstalledDetLongSizes.add(getString(R.string.DET_LONG_SIZE_DEFAULT));
preInstalledScoreThresholds.add(getString(R.string.SCORE_THRESHOLD_DEFAULT));
// Setup UI components
......@@ -89,11 +83,7 @@ public class SettingsActivity extends AppCompatPreferenceActivity implements Sha
(ListPreference) findPreference(getString(R.string.CPU_THREAD_NUM_KEY));
lpCPUPowerMode =
(ListPreference) findPreference(getString(R.string.CPU_POWER_MODE_KEY));
lpInputColorFormat =
(ListPreference) findPreference(getString(R.string.INPUT_COLOR_FORMAT_KEY));
etInputShape = (EditTextPreference) findPreference(getString(R.string.INPUT_SHAPE_KEY));
etInputMean = (EditTextPreference) findPreference(getString(R.string.INPUT_MEAN_KEY));
etInputStd = (EditTextPreference) findPreference(getString(R.string.INPUT_STD_KEY));
etDetLongSize = (EditTextPreference) findPreference(getString(R.string.DET_LONG_SIZE_KEY));
etScoreThreshold = (EditTextPreference) findPreference(getString(R.string.SCORE_THRESHOLD_KEY));
}
......@@ -112,11 +102,7 @@ public class SettingsActivity extends AppCompatPreferenceActivity implements Sha
editor.putString(getString(R.string.IMAGE_PATH_KEY), preInstalledImagePaths.get(modelIdx));
editor.putString(getString(R.string.CPU_THREAD_NUM_KEY), preInstalledCPUThreadNums.get(modelIdx));
editor.putString(getString(R.string.CPU_POWER_MODE_KEY), preInstalledCPUPowerModes.get(modelIdx));
editor.putString(getString(R.string.INPUT_COLOR_FORMAT_KEY),
preInstalledInputColorFormats.get(modelIdx));
editor.putString(getString(R.string.INPUT_SHAPE_KEY), preInstalledInputShapes.get(modelIdx));
editor.putString(getString(R.string.INPUT_MEAN_KEY), preInstalledInputMeans.get(modelIdx));
editor.putString(getString(R.string.INPUT_STD_KEY), preInstalledInputStds.get(modelIdx));
editor.putString(getString(R.string.DET_LONG_SIZE_KEY), preInstalledDetLongSizes.get(modelIdx));
editor.putString(getString(R.string.SCORE_THRESHOLD_KEY),
preInstalledScoreThresholds.get(modelIdx));
editor.apply();
......@@ -129,10 +115,7 @@ public class SettingsActivity extends AppCompatPreferenceActivity implements Sha
etImagePath.setEnabled(enableCustomSettings);
lpCPUThreadNum.setEnabled(enableCustomSettings);
lpCPUPowerMode.setEnabled(enableCustomSettings);
lpInputColorFormat.setEnabled(enableCustomSettings);
etInputShape.setEnabled(enableCustomSettings);
etInputMean.setEnabled(enableCustomSettings);
etInputStd.setEnabled(enableCustomSettings);
etDetLongSize.setEnabled(enableCustomSettings);
etScoreThreshold.setEnabled(enableCustomSettings);
modelPath = sharedPreferences.getString(getString(R.string.MODEL_PATH_KEY),
getString(R.string.MODEL_PATH_DEFAULT));
......@@ -144,14 +127,8 @@ public class SettingsActivity extends AppCompatPreferenceActivity implements Sha
getString(R.string.CPU_THREAD_NUM_DEFAULT));
String cpuPowerMode = sharedPreferences.getString(getString(R.string.CPU_POWER_MODE_KEY),
getString(R.string.CPU_POWER_MODE_DEFAULT));
String inputColorFormat = sharedPreferences.getString(getString(R.string.INPUT_COLOR_FORMAT_KEY),
getString(R.string.INPUT_COLOR_FORMAT_DEFAULT));
String inputShape = sharedPreferences.getString(getString(R.string.INPUT_SHAPE_KEY),
getString(R.string.INPUT_SHAPE_DEFAULT));
String inputMean = sharedPreferences.getString(getString(R.string.INPUT_MEAN_KEY),
getString(R.string.INPUT_MEAN_DEFAULT));
String inputStd = sharedPreferences.getString(getString(R.string.INPUT_STD_KEY),
getString(R.string.INPUT_STD_DEFAULT));
String detLongSize = sharedPreferences.getString(getString(R.string.DET_LONG_SIZE_KEY),
getString(R.string.DET_LONG_SIZE_DEFAULT));
String scoreThreshold = sharedPreferences.getString(getString(R.string.SCORE_THRESHOLD_KEY),
getString(R.string.SCORE_THRESHOLD_DEFAULT));
etModelPath.setSummary(modelPath);
......@@ -164,14 +141,8 @@ public class SettingsActivity extends AppCompatPreferenceActivity implements Sha
lpCPUThreadNum.setSummary(cpuThreadNum);
lpCPUPowerMode.setValue(cpuPowerMode);
lpCPUPowerMode.setSummary(cpuPowerMode);
lpInputColorFormat.setValue(inputColorFormat);
lpInputColorFormat.setSummary(inputColorFormat);
etInputShape.setSummary(inputShape);
etInputShape.setText(inputShape);
etInputMean.setSummary(inputMean);
etInputMean.setText(inputMean);
etInputStd.setSummary(inputStd);
etInputStd.setText(inputStd);
etDetLongSize.setSummary(detLongSize);
etDetLongSize.setText(detLongSize);
etScoreThreshold.setText(scoreThreshold);
etScoreThreshold.setSummary(scoreThreshold);
}
......
deploy/android_demo/app/src/main/res/layout/activity_main.xml
浏览文件 @ 4b214948
......@@ -23,13 +23,7 @@
android:layout_height="wrap_content"
android:orientation="horizontal">
<Button
android:id="@+id/btn_load_model"
android:layout_width="0dp"
android:layout_height="wrap_content"
android:layout_weight="1"
android:onClick="btn_load_model_click"
android:text="加载模型" />
<Button
android:id="@+id/btn_run_model"
android:layout_width="0dp"
......@@ -52,7 +46,45 @@
android:onClick="btn_choice_img_click"
android:text="选取图片" />
<Button
android:id="@+id/btn_reset_img"
android:layout_width="0dp"
android:layout_height="wrap_content"
android:layout_weight="1"
android:onClick="btn_reset_img_click"
android:text="清空绘图" />
</LinearLayout>
<LinearLayout
android:id="@+id/run_mode_layout"
android:layout_width="fill_parent"
android:layout_height="wrap_content"
android:orientation="horizontal">
<CheckBox
android:id="@+id/cb_opencl"
android:layout_width="0dp"
android:layout_weight="1"
android:layout_height="wrap_content"
android:text="开启OPENCL"
android:onClick="cb_opencl_click"
android:visibility="gone"/>
<TextView
android:layout_width="0dp"
android:layout_weight="0.5"
android:layout_height="wrap_content"
android:text="运行模式:"/>
<Spinner
android:id="@+id/sp_run_mode"
android:layout_width="0dp"
android:layout_weight="1.5"
android:layout_height="wrap_content"
android:entries="@array/run_Model"
/>
</LinearLayout>
<TextView
android:id="@+id/tv_input_setting"
android:layout_width="wrap_content"
......@@ -60,7 +92,7 @@
android:scrollbars="vertical"
android:layout_marginLeft="12dp"
android:layout_marginRight="12dp"
android:layout_marginTop="10dp"
android:layout_marginTop="5dp"
android:layout_marginBottom="5dp"
android:lineSpacingExtra="4dp"
android:singleLine="false"
......
deploy/android_demo/app/src/main/res/layout/activity_mini.xml 已删除 100644 → 0
浏览文件 @ 917606b4
<?xml version="1.0" encoding="utf-8"?>
<!-- for MiniActivity Use Only -->
<androidx.constraintlayout.widget.ConstraintLayout xmlns:android="http://schemas.android.com/apk/res/android"
xmlns:app="http://schemas.android.com/apk/res-auto"
xmlns:tools="http://schemas.android.com/tools"
android:layout_width="match_parent"
android:layout_height="match_parent"
app:layout_constraintLeft_toLeftOf="parent"
app:layout_constraintLeft_toRightOf="parent"
tools:context=".MainActivity">
<TextView
android:id="@+id/sample_text"
android:layout_width="0dp"
android:layout_height="wrap_content"
android:text="Hello World!"
app:layout_constraintLeft_toLeftOf="parent"
app:layout_constraintRight_toRightOf="parent"
app:layout_constraintTop_toBottomOf="@id/imageView"
android:scrollbars="vertical"
/>
<ImageView
android:id="@+id/imageView"
android:layout_width="wrap_content"
android:layout_height="wrap_content"
android:paddingTop="20dp"
android:paddingBottom="20dp"
app:layout_constraintBottom_toTopOf="@id/imageView"
app:layout_constraintLeft_toLeftOf="parent"
app:layout_constraintRight_toRightOf="parent"
app:layout_constraintTop_toTopOf="parent"
tools:srcCompat="@tools:sample/avatars" />
<Button
android:id="@+id/button"
android:layout_width="wrap_content"
android:layout_height="wrap_content"
android:layout_marginBottom="4dp"
android:text="Button"
app:layout_constraintBottom_toBottomOf="parent"
app:layout_constraintLeft_toLeftOf="parent"
app:layout_constraintRight_toRightOf="parent"
tools:layout_editor_absoluteX="161dp" />
</androidx.constraintlayout.widget.ConstraintLayout>
\ No newline at end of file
deploy/android_demo/app/src/main/res/values/arrays.xml
浏览文件 @ 4b214948
<?xml version="1.0" encoding="utf-8"?>
<resources>
<string-array name="image_name_entries">
<item>0.jpg</item>
<item>90.jpg</item>
<item>180.jpg</item>
<item>270.jpg</item>
<item>det_0.jpg</item>
<item>det_90.jpg</item>
<item>det_180.jpg</item>
<item>det_270.jpg</item>
<item>rec_0.jpg</item>
<item>rec_0_180.jpg</item>
<item>rec_1.jpg</item>
<item>rec_1_180.jpg</item>
</string-array>
<string-array name="image_name_values">
<item>images/0.jpg</item>
<item>images/90.jpg</item>
<item>images/180.jpg</item>
<item>images/270.jpg</item>
<item>images/det_0.jpg</item>
<item>images/det_90.jpg</item>
<item>images/det_180.jpg</item>
<item>images/det_270.jpg</item>
<item>images/rec_0.jpg</item>
<item>images/rec_0_180.jpg</item>
<item>images/rec_1.jpg</item>
<item>images/rec_1_180.jpg</item>
</string-array>
<string-array name="cpu_thread_num_entries">
<item>1 threads</item>
......@@ -48,4 +56,12 @@
<item>BGR</item>
<item>RGB</item>
</string-array>
<string-array name="run_Model">
<item>检测+分类+识别</item>
<item>检测+识别</item>
<item>分类+识别</item>
<item>检测</item>
<item>识别</item>
<item>分类</item>
</string-array>
</resources>
\ No newline at end of file
deploy/android_demo/app/src/main/res/values/strings.xml
浏览文件 @ 4b214948
<resources>
<string name="app_name">OCR Chinese</string>
<string name="app_name">PaddleOCR</string>
<string name="CHOOSE_PRE_INSTALLED_MODEL_KEY">CHOOSE_PRE_INSTALLED_MODEL_KEY</string>
<string name="ENABLE_CUSTOM_SETTINGS_KEY">ENABLE_CUSTOM_SETTINGS_KEY</string>
<string name="MODEL_PATH_KEY">MODEL_PATH_KEY</string>
......@@ -7,20 +7,14 @@
<string name="IMAGE_PATH_KEY">IMAGE_PATH_KEY</string>
<string name="CPU_THREAD_NUM_KEY">CPU_THREAD_NUM_KEY</string>
<string name="CPU_POWER_MODE_KEY">CPU_POWER_MODE_KEY</string>
<string name="INPUT_COLOR_FORMAT_KEY">INPUT_COLOR_FORMAT_KEY</string>
<string name="INPUT_SHAPE_KEY">INPUT_SHAPE_KEY</string>
<string name="INPUT_MEAN_KEY">INPUT_MEAN_KEY</string>
<string name="INPUT_STD_KEY">INPUT_STD_KEY</string>
<string name="DET_LONG_SIZE_KEY">DET_LONG_SIZE_KEY</string>
<string name="SCORE_THRESHOLD_KEY">SCORE_THRESHOLD_KEY</string>
<string name="MODEL_PATH_DEFAULT">models/ocr_v2_for_cpu</string>
<string name="MODEL_PATH_DEFAULT">models/ch_PP-OCRv2</string>
<string name="LABEL_PATH_DEFAULT">labels/ppocr_keys_v1.txt</string>
<string name="IMAGE_PATH_DEFAULT">images/0.jpg</string>
<string name="IMAGE_PATH_DEFAULT">images/det_0.jpg</string>
<string name="CPU_THREAD_NUM_DEFAULT">4</string>
<string name="CPU_POWER_MODE_DEFAULT">LITE_POWER_HIGH</string>
<string name="INPUT_COLOR_FORMAT_DEFAULT">BGR</string>
<string name="INPUT_SHAPE_DEFAULT">1,3,960</string>
<string name="INPUT_MEAN_DEFAULT">0.485, 0.456, 0.406</string>
<string name="INPUT_STD_DEFAULT">0.229,0.224,0.225</string>
<string name="DET_LONG_SIZE_DEFAULT">960</string>
<string name="SCORE_THRESHOLD_DEFAULT">0.1</string>
</resources>
deploy/android_demo/app/src/main/res/xml/settings.xml
浏览文件 @ 4b214948
......@@ -47,26 +47,10 @@
android:entryValues="@array/cpu_power_mode_values"/>
</PreferenceCategory>
<PreferenceCategory android:title="Input Settings">
<ListPreference
android:defaultValue="@string/INPUT_COLOR_FORMAT_DEFAULT"
android:key="@string/INPUT_COLOR_FORMAT_KEY"
android:negativeButtonText="@null"
android:positiveButtonText="@null"
android:title="Input Color Format: BGR or RGB"
android:entries="@array/input_color_format_entries"
android:entryValues="@array/input_color_format_values"/>
<EditTextPreference
android:key="@string/INPUT_SHAPE_KEY"
android:defaultValue="@string/INPUT_SHAPE_DEFAULT"
android:title="Input Shape: (1,1,max_width_height) or (1,3,max_width_height)" />
<EditTextPreference
android:key="@string/INPUT_MEAN_KEY"
android:defaultValue="@string/INPUT_MEAN_DEFAULT"
android:title="Input Mean: (channel/255-mean)/std" />
<EditTextPreference
android:key="@string/INPUT_STD_KEY"
android:defaultValue="@string/INPUT_STD_DEFAULT"
android:title="Input Std: (channel/255-mean)/std" />
android:key="@string/DET_LONG_SIZE_KEY"
android:defaultValue="@string/DET_LONG_SIZE_DEFAULT"
android:title="det long size" />
</PreferenceCategory>
<PreferenceCategory android:title="Output Settings">
<EditTextPreference
......
deploy/lite/readme.md
浏览文件 @ 4b214948
- [端侧部署](#端侧部署)
- [1. 准备环境](#1-准备环境)
- [运行准备](#运行准备)
- [1.1 准备交叉编译环境](#11-准备交叉编译环境)
- [1.2 准备预测库](#12-准备预测库)
- [2 开始运行](#2-开始运行)
- [2.1 模型优化](#21-模型优化)
- [2.2 与手机联调](#22-与手机联调)
- [注意：](#注意)
- [FAQ](#faq)
# 端侧部署
本教程将介绍基于[Paddle Lite](https://github.com/PaddlePaddle/Paddle-Lite) 在移动端部署PaddleOCR超轻量中文检测、识别模型的详细步骤。
......@@ -26,23 +37,23 @@ Paddle Lite是飞桨轻量化推理引擎，为手机、IOT端提供高效推理
| 平台 | 预测库下载链接 |
|---|---|
|Android|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.android.armv7.gcc.c++_shared.with_extra.with_cv.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.android.armv8.gcc.c++_shared.with_extra.with_cv.tar.gz)|
|IOS|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.ios.armv7.with_cv.with_extra.with_log.tiny_publish.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.ios.armv8.with_cv.with_extra.with_log.tiny_publish.tar.gz)|
|Android|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.10/inference_lite_lib.android.armv7.gcc.c++_shared.with_extra.with_cv.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.10/inference_lite_lib.android.armv8.gcc.c++_shared.with_extra.with_cv.tar.gz)|
|IOS|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.10/inference_lite_lib.ios.armv7.with_cv.with_extra.with_log.tiny_publish.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.10/inference_lite_lib.ios.armv8.with_cv.with_extra.with_log.tiny_publish.tar.gz)|
注：1. 上述预测库为PaddleLite 2.9分支编译得到，有关PaddleLite 2.9 详细信息可参考 [链接](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.9) 。
注：1. 上述预测库为PaddleLite 2.10分支编译得到，有关PaddleLite 2.10 详细信息可参考 [链接](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.10) 。
- 2. [推荐]编译Paddle-Lite得到预测库，Paddle-Lite的编译方式如下：
```
git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite
# 切换到Paddle-Lite release/v2.9 稳定分支
git checkout release/v2.9
# 切换到Paddle-Lite release/v2.10 稳定分支
git checkout release/v2.10
./lite/tools/build_android.sh --arch=armv8 --with_cv=ON --with_extra=ON
```
注意：编译Paddle-Lite获得预测库时，需要打开`--with_cv=ON --with_extra=ON`两个选项，`--arch`表示`arm`版本，这里指定为armv8，
更多编译命令
介绍请参考 [链接](https://paddle-lite.readthedocs.io/zh/latest/source_compile/compile_andriod.html) 。
介绍请参考 [链接](https://paddle-lite.readthedocs.io/zh/release-v2.10_a/source_compile/linux_x86_compile_android.html) 。
直接下载预测库并解压后，可以得到`inference_lite_lib.android.armv8/`文件夹，通过编译Paddle-Lite得到的预测库位于
`Paddle-Lite/build.lite.android.armv8.gcc/inference_lite_lib.android.armv8/`文件夹下。
......@@ -85,8 +96,8 @@ Paddle-Lite 提供了多种策略来自动优化原始的模型，其中包括
|模型版本|模型简介|模型大小|检测模型|文本方向分类模型|识别模型|Paddle-Lite版本|
|---|---|---|---|---|---|---|
|V2.0|超轻量中文OCR 移动端模型|7.8M|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_det_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_rec_opt.nb)|v2.9|
|V2.0(slim)|超轻量中文OCR 移动端模型|3.3M|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_det_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_rec_slim_opt.nb)|v2.9|
|PP-OCRv2|蒸馏版超轻量中文OCR移动端模型|11M|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_infer_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_infer_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_infer_opt.nb)|v2.10|
|PP-OCRv2(slim)|蒸馏版超轻量中文OCR移动端模型|4.6M|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_slim_opt.nb)|v2.10|
如果直接使用上述表格中的模型进行部署，可略过下述步骤，直接阅读 [2.2节](#2.2与手机联调)。
......@@ -97,7 +108,7 @@ Paddle-Lite 提供了多种策略来自动优化原始的模型，其中包括
# 如果准备环境时已经clone了Paddle-Lite，则不用重新clone Paddle-Lite
git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite
git checkout release/v2.9
git checkout release/v2.10
# 启动编译
./lite/tools/build.sh build_optimize_tool
```
......@@ -123,15 +134,15 @@ cd build.opt/lite/api/
下面以PaddleOCR的超轻量中文模型为例，介绍使用编译好的opt文件完成inference模型到Paddle-Lite优化模型的转换。
```
# 【推荐】 下载PaddleOCR V2.0版本的中英文 inference模型
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_det_slim_infer.tar
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_rec_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_rec_slim_infer.tar
# 【推荐】 下载 PP-OCRv2版本的中英文 inference模型
wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar && tar xf ch_PP-OCRv2_det_slim_quant_infer.tar
wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar && tar xf ch_PP-OCRv2_rec_slim_quant_infer.tar
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_cls_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_cls_slim_infer.tar
# 转换V2.0检测模型
./opt --model_file=./ch_ppocr_mobile_v2.0_det_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_det_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_det_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# 转换V2.0识别模型
./opt --model_file=./ch_ppocr_mobile_v2.0_rec_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_rec_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_rec_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# 转换V2.0方向分类器模型
# 转换检测模型
./opt --model_file=./ch_PP-OCRv2_det_slim_quant_infer/inference.pdmodel --param_file=./ch_PP-OCRv2_det_slim_quant_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv2_det_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# 转换识别模型
./opt --model_file=./ch_PP-OCRv2_rec_slim_quant_infer/inference.pdmodel --param_file=./ch_PP-OCRv2_rec_slim_quant_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv2_rec_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# 转换方向分类器模型
./opt --model_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_cls_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
```
......@@ -186,15 +197,15 @@ wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_cls
```
准备测试图像，以`PaddleOCR/doc/imgs/11.jpg`为例，将测试的图像复制到`demo/cxx/ocr/debug/`文件夹下。
准备lite opt工具优化后的模型文件，比如使用`ch_ppocr_mobile_v2.0_det_slim_opt.nb，ch_ppocr_mobile_v2.0_rec_slim_opt.nb, ch_ppocr_mobile_v2.0_cls_slim_opt.nb`，模型文件放置在`demo/cxx/ocr/debug/`文件夹下。
准备lite opt工具优化后的模型文件，比如使用`ch_PP-OCRv2_det_slim_opt.ch_PP-OCRv2_rec_slim_rec.nb, ch_ppocr_mobile_v2.0_cls_slim_opt.nb`，模型文件放置在`demo/cxx/ocr/debug/`文件夹下。
执行完成后，ocr文件夹下将有如下文件格式：
```
demo/cxx/ocr/
|-- debug/
| |--ch_ppocr_mobile_v2.0_det_slim_opt.nb 优化后的检测模型文件
| |--ch_ppocr_mobile_v2.0_rec_slim_opt.nb 优化后的识别模型文件
| |--ch_PP-OCRv2_det_slim_opt.nb 优化后的检测模型文件
| |--ch_PP-OCRv2_rec_slim_opt.nb 优化后的识别模型文件
| |--ch_ppocr_mobile_v2.0_cls_slim_opt.nb 优化后的文字方向分类器模型文件
| |--11.jpg 待测试图像
| |--ppocr_keys_v1.txt 中文字典文件
......@@ -250,7 +261,7 @@ use_direction_classify 0 # 是否使用方向分类器，0表示不使用，1
export LD_LIBRARY_PATH=${PWD}:$LD_LIBRARY_PATH
# 开始使用，ocr_db_crnn可执行文件的使用方式为:
# ./ocr_db_crnn 检测模型文件 方向分类器模型文件 识别模型文件 测试图像路径 字典文件路径
./ocr_db_crnn ch_ppocr_mobile_v2.0_det_slim_opt.nb ch_ppocr_mobile_v2.0_rec_slim_opt.nb ch_ppocr_mobile_v2.0_cls_slim_opt.nb ./11.jpg ppocr_keys_v1.txt
./ocr_db_crnn ch_PP-OCRv2_det_slim_opt.nb ch_PP-OCRv2_rec_slim_opt.nb ch_ppocr_mobile_v2.0_cls_slim_opt.nb ./11.jpg ppocr_keys_v1.txt
```
如果对代码做了修改，则需要重新编译并push到手机上。
......
deploy/lite/readme_en.md
浏览文件 @ 4b214948
- [Tutorial of PaddleOCR Mobile deployment](#tutorial-of-paddleocr-mobile-deployment)
- [1. Preparation](#1-preparation)
- [Preparation environment](#preparation-environment)
- [1.1 Prepare the cross-compilation environment](#11-prepare-the-cross-compilation-environment)
- [1.2 Prepare Paddle-Lite library](#12-prepare-paddle-lite-library)
- [2 Run](#2-run)
- [2.1 Inference Model Optimization](#21-inference-model-optimization)
- [2.2 Run optimized model on Phone](#22-run-optimized-model-on-phone)
- [注意：](#注意)
- [FAQ](#faq)
# Tutorial of PaddleOCR Mobile deployment
This tutorial will introduce how to use [Paddle Lite](https://github.com/PaddlePaddle/Paddle-Lite) to deploy PaddleOCR ultra-lightweight Chinese and English detection models on mobile phones.
......@@ -28,23 +39,23 @@ There are two ways to obtain the Paddle-Lite library：
| Platform | Paddle-Lite library download link |
|---|---|
|Android|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.android.armv7.gcc.c++_shared.with_extra.with_cv.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.android.armv8.gcc.c++_shared.with_extra.with_cv.tar.gz)|
|IOS|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.ios.armv7.with_cv.with_extra.with_log.tiny_publish.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.ios.armv8.with_cv.with_extra.with_log.tiny_publish.tar.gz)|
|Android|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.10/inference_lite_lib.android.armv7.gcc.c++_shared.with_extra.with_cv.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.10/inference_lite_lib.android.armv8.gcc.c++_shared.with_extra.with_cv.tar.gz)|
|IOS|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.10/inference_lite_lib.ios.armv7.with_cv.with_extra.with_log.tiny_publish.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.10/inference_lite_lib.ios.armv8.with_cv.with_extra.with_log.tiny_publish.tar.gz)|
Note: 1. The above Paddle-Lite library is compiled from the Paddle-Lite 2.9 branch. For more information about Paddle-Lite 2.9, please refer to [link](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.9).
Note: 1. The above Paddle-Lite library is compiled from the Paddle-Lite 2.10 branch. For more information about Paddle-Lite 2.10, please refer to [link](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.10).
- 2. [Recommended] Compile Paddle-Lite to get the prediction library. The compilation method of Paddle-Lite is as follows：
```
git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite
# Switch to Paddle-Lite release/v2.8 stable branch
git checkout release/v2.8
# Switch to Paddle-Lite release/v2.10 stable branch
git checkout release/v2.10
./lite/tools/build_android.sh --arch=armv8 --with_cv=ON --with_extra=ON
```
Note: When compiling Paddle-Lite to obtain the Paddle-Lite library, you need to turn on the two options `--with_cv=ON --with_extra=ON`, `--arch` means the `arm` version, here is designated as armv8,
More compilation commands refer to the introduction [link](https://paddle-lite.readthedocs.io/zh/latest/source_compile/compile_andriod.html) 。
More compilation commands refer to the introduction [link](https://paddle-lite.readthedocs.io/zh/release-v2.10_a/source_compile/linux_x86_compile_android.html) 。
After directly downloading the Paddle-Lite library and decompressing it, you can get the `inference_lite_lib.android.armv8/` folder, and the Paddle-Lite library obtained by compiling Paddle-Lite is located
`Paddle-Lite/build.lite.android.armv8.gcc/inference_lite_lib.android.armv8/` folder.
......@@ -87,10 +98,10 @@ The following table also provides a series of models that can be deployed on mob
|Version|Introduction|Model size|Detection model|Text Direction model|Recognition model|Paddle-Lite branch|
|---|---|---|---|---|---|---|
|V2.0|extra-lightweight chinese OCR optimized model|7.8M|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_det_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_rec_opt.nb)|v2.9|
|V2.0(slim)|extra-lightweight chinese OCR optimized model|3.3M|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_det_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_rec_slim_opt.nb)|v2.9|
|PP-OCRv2|extra-lightweight chinese OCR optimized model|11M|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_infer_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_infer_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_infer_opt.nb)|v2.10|
|PP-OCRv2(slim)|extra-lightweight chinese OCR optimized model|4.6M|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_slim_opt.nb)|v2.10|
If you directly use the model in the above table for deployment, you can skip the following steps and directly read [Section 2.2](#2.2 Run optimized model on Phone).
If you directly use the model in the above table for deployment, you can skip the following steps and directly read [Section 2.2](#2.2-Run-optimized-model-on-Phone).
If the model to be deployed is not in the above table, you need to follow the steps below to obtain the optimized model.
......@@ -98,7 +109,7 @@ The `opt` tool can be obtained by compiling Paddle Lite.
```
git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite
git checkout release/v2.9
git checkout release/v2.10
./lite/tools/build.sh build_optimize_tool
```
......@@ -124,22 +135,22 @@ cd build.opt/lite/api/
The following takes the ultra-lightweight Chinese model of PaddleOCR as an example to introduce the use of the compiled opt file to complete the conversion of the inference model to the Paddle-Lite optimized model
```
# [Recommendation] Download the Chinese and English inference model of PaddleOCR V2.0
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_det_slim_infer.tar
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_rec_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_rec_slim_infer.tar
# 【[Recommendation] Download the Chinese and English inference model of PP-OCRv2
wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar && tar xf ch_PP-OCRv2_det_slim_quant_infer.tar
wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar && tar xf ch_PP-OCRv2_rec_slim_quant_infer.tar
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_cls_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_cls_slim_infer.tar
# Convert V2.0 detection model
./opt --model_file=./ch_ppocr_mobile_v2.0_det_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_det_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_det_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# Convert V2.0 recognition model
./opt --model_file=./ch_ppocr_mobile_v2.0_rec_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_rec_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_rec_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# Convert V2.0 angle classifier model
# Convert detection model
./opt --model_file=./ch_PP-OCRv2_det_slim_quant_infer/inference.pdmodel --param_file=./ch_PP-OCRv2_det_slim_quant_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv2_det_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# Convert recognition model
./opt --model_file=./ch_PP-OCRv2_rec_slim_quant_infer/inference.pdmodel --param_file=./ch_PP-OCRv2_rec_slim_quant_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv2_rec_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# Convert angle classifier model
./opt --model_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_cls_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
```
After the conversion is successful, there will be more files ending with `.nb` in the inference model directory, which is the successfully converted model file.
<a name="2.2 Run optimized model on Phone"></a>
<a name="2.2-Run-optimized-model-on-Phone"></a>
### 2.2 Run optimized model on Phone
Some preparatory work is required first.
......@@ -194,8 +205,8 @@ The structure of the OCR demo is as follows after the above command is executed:
```
demo/cxx/ocr/
|-- debug/
| |--ch_ppocr_mobile_v2.0_det_slim_opt.nb Detection model
| |--ch_ppocr_mobile_v2.0_rec_slim_opt.nb Recognition model
| |--ch_PP-OCRv2_det_slim_opt.nb Detection model
| |--ch_PP-OCRv2_rec_slim_opt.nb Recognition model
| |--ch_ppocr_mobile_v2.0_cls_slim_opt.nb Text direction classification model
| |--11.jpg Image for OCR
| |--ppocr_keys_v1.txt Dictionary file
......@@ -249,7 +260,7 @@ After the above steps are completed, you can use adb to push the file to the pho
export LD_LIBRARY_PATH=${PWD}:$LD_LIBRARY_PATH
# The use of ocr_db_crnn is:
# ./ocr_db_crnn Detection model file Orientation classifier model file Recognition model file Test image path Dictionary file path
./ocr_db_crnn ch_ppocr_mobile_v2.0_det_opt.nb ch_ppocr_mobile_v2.0_rec_opt.nb ch_ppocr_mobile_v2.0_cls_opt.nb ./11.jpg ppocr_keys_v1.txt
./ocr_db_crnn ch_PP-OCRv2_det_slim_opt.nb ch_PP-OCRv2_rec_slim_opt.nb ch_ppocr_mobile_v2.0_cls_opt.nb ./11.jpg ppocr_keys_v1.txt
```
If you modify the code, you need to recompile and push to the phone.
......
deploy/pdserving/README.md
浏览文件 @ 4b214948
......@@ -19,10 +19,14 @@ The introduction and tutorial of Paddle Serving service deployment framework ref
## Contents
- [Environmental preparation](#environmental-preparation)
- [Model conversion](#model-conversion)
- [Paddle Serving pipeline deployment](#paddle-serving-pipeline-deployment)
- [FAQ](#faq)
- [OCR Pipeline WebService](#ocr-pipeline-webservice)
- [Service deployment based on PaddleServing](#service-deployment-based-on-paddleserving)
- [Contents](#contents)
- [Environmental preparation](#environmental-preparation)
- [Model conversion](#model-conversion)
- [Paddle Serving pipeline deployment](#paddle-serving-pipeline-deployment)
- [WINDOWS Users](#windows-users)
- [FAQ](#faq)
<a name="environmental-preparation"></a>
## Environmental preparation
......@@ -201,7 +205,7 @@ The recognition model is the same.
## WINDOWS Users
Windows does not support Pipeline Serving, if we want to lauch paddle serving on Windows, we should use Web Service, for more infomation please refer to [Paddle Serving for Windows Users](https://github.com/PaddlePaddle/Serving/blob/develop/doc/WINDOWS_TUTORIAL.md)
Windows does not support Pipeline Serving, if we want to lauch paddle serving on Windows, we should use Web Service, for more infomation please refer to [Paddle Serving for Windows Users](https://github.com/PaddlePaddle/Serving/blob/develop/doc/Windows_Tutorial_EN.md)
**WINDOWS user can only use version 0.5.0 CPU Mode**
......
deploy/slim/prune/README.md
浏览文件 @ 4b214948
......@@ -45,7 +45,7 @@ python3 setup.py install
'conv10_expand_weights': {0.1: 0.006509952684312718, 0.2: 0.01827734339798862, 0.3: 0.014528405644659832, 0.6: 0.06536008804270439, 0.8: 0.11798612250664964, 0.7: 0.12391408417493704, 0.4: 0.030615754498018757, 0.5: 0.047105205602406594}
'conv10_linear_weights': {0.1: 0.05113190831455035, 0.2: 0.07705573833558801, 0.3: 0.12096721757739311, 0.6: 0.5135061352930738, 0.8: 0.7908166677143281, 0.7: 0.7272187676899062, 0.4: 0.1819252083008504, 0.5: 0.3728054727792405}
}
加载敏感度文件后会返回一个字典，字典中的keys为网络模型参数模型的名字，values为一个字典，里面保存了相应网络层的裁剪敏感度信息。例如在例子中，conv10_expand_weights所对应的网络层在裁掉10%的卷积核后模型性能相较原模型会下降0.65%，详细信息可见[PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/algo/algo.md#2-%E5%8D%B7%E7%A7%AF%E6%A0%B8%E5%89%AA%E8%A3%81%E5%8E%9F%E7%90%86)
加载敏感度文件后会返回一个字典，字典中的keys为网络模型参数模型的名字，values为一个字典，里面保存了相应网络层的裁剪敏感度信息。例如在例子中，conv10_expand_weights所对应的网络层在裁掉10%的卷积核后模型性能相较原模型会下降0.65%，详细信息可见[PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim/blob/release/2.0-alpha/docs/zh_cn/algo/algo.md)
进入PaddleOCR根目录，通过以下命令对模型进行敏感度分析训练：
```bash
......
deploy/slim/prune/README_en.md
浏览文件 @ 4b214948
......@@ -3,7 +3,7 @@
Generally, a more complex model would achive better performance in the task, but it also leads to some redundancy in the model. Model Pruning is a technique that reduces this redundancy by removing the sub-models in the neural network model, so as to reduce model calculation complexity and improve model inference performance.
This example uses PaddleSlim provided[APIs of Pruning](https://paddlepaddle.github.io/PaddleSlim/api/prune_api/) to compress the OCR model.
This example uses PaddleSlim provided[APIs of Pruning](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/docs/zh_cn/api_cn/dygraph/pruners) to compress the OCR model.
[PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim), an open source library which integrates model pruning, quantization (including quantization training and offline quantization), distillation, neural network architecture search, and many other commonly used and leading model compression technique in the industry.
It is recommended that you could understand following pages before reading this example：
......@@ -35,7 +35,7 @@ PaddleOCR also provides a series of [models](../../../doc/doc_en/models_list_en.
### 3. Pruning sensitivity analysis
After the pre-trained model is loaded, sensitivity analysis is performed on each network layer of the model to understand the redundancy of each network layer, and save a sensitivity file which named: sen.pickle. After that, user could load the sensitivity file via the [methods provided by PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/paddleslim/prune/sensitive.py#L221) and determining the pruning ratio of each network layer automatically. For specific details of sensitivity analysis, see：[Sensitivity analysis](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/tutorials/image_classification_sensitivity_analysis_tutorial.md)
After the pre-trained model is loaded, sensitivity analysis is performed on each network layer of the model to understand the redundancy of each network layer, and save a sensitivity file which named: sen.pickle. After that, user could load the sensitivity file via the [methods provided by PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/paddleslim/prune/sensitive.py#L221) and determining the pruning ratio of each network layer automatically. For specific details of sensitivity analysis, see：[Sensitivity analysis](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/en/tutorials/image_classification_sensitivity_analysis_tutorial_en.md)
The data format of sensitivity file：
sen.pickle(Dict){
'layer_weight_name_0': sens_of_each_ratio(Dict){'pruning_ratio_0': acc_loss, 'pruning_ratio_1': acc_loss}
......@@ -47,7 +47,7 @@ PaddleOCR also provides a series of [models](../../../doc/doc_en/models_list_en.
'conv10_expand_weights': {0.1: 0.006509952684312718, 0.2: 0.01827734339798862, 0.3: 0.014528405644659832, 0.6: 0.06536008804270439, 0.8: 0.11798612250664964, 0.7: 0.12391408417493704, 0.4: 0.030615754498018757, 0.5: 0.047105205602406594}
'conv10_linear_weights': {0.1: 0.05113190831455035, 0.2: 0.07705573833558801, 0.3: 0.12096721757739311, 0.6: 0.5135061352930738, 0.8: 0.7908166677143281, 0.7: 0.7272187676899062, 0.4: 0.1819252083008504, 0.5: 0.3728054727792405}
}
The function would return a dict after loading the sensitivity file. The keys of the dict are name of parameters in each layer. And the value of key is the information about pruning sensitivity of corresponding layer. In example, pruning 10% filter of the layer corresponding to conv10_expand_weights would lead to 0.65% degradation of model performance. The details could be seen at: [Sensitivity analysis](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/algo/algo.md#2-%E5%8D%B7%E7%A7%AF%E6%A0%B8%E5%89%AA%E8%A3%81%E5%8E%9F%E7%90%86)
The function would return a dict after loading the sensitivity file. The keys of the dict are name of parameters in each layer. And the value of key is the information about pruning sensitivity of corresponding layer. In example, pruning 10% filter of the layer corresponding to conv10_expand_weights would lead to 0.65% degradation of model performance. The details could be seen at: [Sensitivity analysis](https://github.com/PaddlePaddle/PaddleSlim/blob/release/2.0-alpha/docs/zh_cn/algo/algo.md)
Enter the PaddleOCR root directory，perform sensitivity analysis on the model with the following command：
......
deploy/slim/quantization/README_en.md
浏览文件 @ 4b214948
......@@ -5,11 +5,11 @@ Generally, a more complex model would achieve better performance in the task, bu
Quantization is a technique that reduces this redundancy by reducing the full precision data to a fixed number,
so as to reduce model calculation complexity and improve model inference performance.
This example uses PaddleSlim provided [APIs of Quantization](https://paddlepaddle.github.io/PaddleSlim/api/quantization_api/) to compress the OCR model.
This example uses PaddleSlim provided [APIs of Quantization](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/api_cn/dygraph/quanter/qat.rst) to compress the OCR model.
It is recommended that you could understand following pages before reading this example：
- [The training strategy of OCR model](../../../doc/doc_en/quickstart_en.md)
- [PaddleSlim Document](https://paddlepaddle.github.io/PaddleSlim/api/quantization_api/)
- [PaddleSlim Document](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/api_cn/dygraph/quanter/qat.rst)
## Quick Start
Quantization is mostly suitable for the deployment of lightweight models on mobile terminals.
......
doc/doc_ch/FAQ.md
浏览文件 @ 4b214948
......@@ -349,7 +349,7 @@ A：PaddleOCR已完成Windows和Mac系统适配，运行时注意两点：
#### Q：训练文字识别模型，真实数据有30w，合成数据有500w，需要做样本均衡吗？
A：需要，一般需要保证一个batch中真实数据样本和合成数据样本的比例是1：1~1：3左右效果比较理想。如果合成数据过大，会过拟合到合成数据，预测效果往往不佳。还有一种启发性的尝试是可以先用大量合成数据训练一个base模型，然后再用真实数据微调，在一些简单场景效果也是会有提升的。
A：需要，一般需要保证一个batch中真实数据样本和合成数据样本的比例是5：1~10：1左右效果比较理想。如果合成数据过大，会过拟合到合成数据，预测效果往往不佳。还有一种启发性的尝试是可以先用大量合成数据训练一个base模型，然后再用真实数据微调，在一些简单场景效果也是会有提升的。
#### Q: 当训练数据量少时，如何获取更多的数据？
......@@ -734,7 +734,7 @@ C++TensorRT预测需要使用支持TRT的预测库并在编译时打开[-DWITH_T
#### Q：PaddleOCR中，对于模型预测加速，CPU加速的途径有哪些？基于TenorRT加速GPU对输入有什么要求？
**A**：（1）CPU可以使用mkldnn进行加速；对于python inference的话，可以把enable_mkldnn改为true，[参考代码](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/tools/infer/utility.py#L99)，对于cpp inference的话，在配置文件里面配置use_mkldnn 1即可，[参考代码](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/deploy/cpp_infer/tools/config.txt#L6)
**A**：（1）CPU可以使用mkldnn进行加速；对于python inference的话，可以把enable_mkldnn改为true，[参考代码](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/tools/infer/utility.py#L99)，对于cpp inference的话，可参考[文档](https://github.com/PaddlePaddle/PaddleOCR/tree/dygraph/deploy/cpp_infer)
（2）GPU需要注意变长输入问题等，TRT6 之后才支持变长输入
......
doc/doc_ch/config.md
浏览文件 @ 4b214948
......@@ -66,7 +66,7 @@
| :---------------------: | :---------------------: | :--------------: | :--------------------: |
| model_type | 网络类型 | rec | 目前支持`rec`,`det`,`cls` |
| algorithm | 模型名称 | CRNN | 支持列表见[algorithm_overview](./algorithm_overview.md) |
| **Transform** | 设置变换方式 | - | 目前仅rec类型的算法支持, 具体见[ppocr/modeling/transform](../../ppocr/modeling/transform) |
| **Transform** | 设置变换方式 | - | 目前仅rec类型的算法支持, 具体见[ppocr/modeling/transforms](../../ppocr/modeling/transforms) |
| name | 变换方式类名 | TPS | 目前支持`TPS` |
| num_fiducial | TPS控制点数 | 20 | 上下边各十个 |
| loc_lr | 定位网络学习率 | 0.1 | \ |
......
doc/doc_ch/models_list.md
浏览文件 @ 4b214948
......@@ -6,13 +6,14 @@
> 3. 本文档提供的是PPOCR自研模型列表，更多基于公开数据集的算法介绍与预训练模型可以参考：[算法概览文档](./algorithm_overview.md)。
- [1. 文本检测模型](#文本检测模型)
- [2. 文本识别模型](#文本识别模型)
- [2.1 中文识别模型](#中文识别模型)
- [2.2 英文识别模型](#英文识别模型)
- [2.3 多语言识别模型](#多语言识别模型)
- [3. 文本方向分类模型](#文本方向分类模型)
- [4. Paddle-Lite 模型](#Paddle-Lite模型)
- [PP-OCR系列模型列表（V2.1，2021年9月6日更新）](#pp-ocr系列模型列表v212021年9月6日更新)
- [1. 文本检测模型](#1-文本检测模型)
- [2. 文本识别模型](#2-文本识别模型)
- [2.1 中文识别模型](#21-中文识别模型)
- [2.2 英文识别模型](#22-英文识别模型)
- [2.3 多语言识别模型（更多语言持续更新中...）](#23-多语言识别模型更多语言持续更新中)
- [3. 文本方向分类模型](#3-文本方向分类模型)
- [4. Paddle-Lite 模型](#4-paddle-lite-模型)
PaddleOCR提供的可下载模型包括`推理模型`、`训练模型`、`预训练模型`、`slim模型`，模型区别说明如下：
......@@ -50,7 +51,7 @@ PaddleOCR提供的可下载模型包括`推理模型`、`训练模型`、`预训
|模型名称|模型简介|配置文件|推理模型大小|下载地址|
| --- | --- | --- | --- | --- |
|ch_PP-OCRv2_rec_slim|【最新】slim量化版超轻量模型，支持中英文、数字识别|[ch_PP-OCRv2_rec.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml)| 9M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_train.tar) |
|ch_PP-OCRv2_rec|【最新】原始超轻量模型，支持中英文、数字识别|[ch_PP-OCRv2_rec.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml)|8.5M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar) |
|ch_PP-OCRv2_rec|【最新】原始超轻量模型，支持中英文、数字识别|[ch_PP-OCRv2_rec_distillation.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec_distillation.yml)|8.5M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar) |
|ch_ppocr_mobile_slim_v2.0_rec|slim裁剪量化版超轻量模型，支持中英文、数字识别|[rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml)| 6M |[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_train.tar) |
|ch_ppocr_mobile_v2.0_rec|原始超轻量模型，支持中英文、数字识别|[rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml)|5.2M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_train.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_pre.tar) |
|ch_ppocr_server_v2.0_rec|通用模型，支持中英文、数字识别|[rec_chinese_common_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_common_train_v2.0.yml)|94.8M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_train.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_pre.tar) |
......@@ -100,6 +101,8 @@ PaddleOCR提供的可下载模型包括`推理模型`、`训练模型`、`预训
|模型版本|模型简介|模型大小|检测模型|文本方向分类模型|识别模型|Paddle-Lite版本|
|---|---|---|---|---|---|---|
|PP-OCRv2|蒸馏版超轻量中文OCR移动端模型|11M|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_infer_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_infer_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_infer_opt.nb)|v2.10|
|PP-OCRv2(slim)|蒸馏版超轻量中文OCR移动端模型|4.6M|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_slim_opt.nb)|v2.10|
|PP-OCRv2|蒸馏版超轻量中文OCR移动端模型|11M|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer_opt.nb)|v2.9|
|PP-OCRv2(slim)|蒸馏版超轻量中文OCR移动端模型|4.9M|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_opt.nb)|v2.9|
|V2.0|ppocr_v2.0超轻量中文OCR移动端模型|7.8M|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_det_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_rec_opt.nb)|v2.9|
......
doc/doc_ch/quickstart.md
浏览文件 @ 4b214948
# PaddleOCR快速开始
- [PaddleOCR快速开始](#paddleocr)
+ [1. 安装PaddleOCR whl包](#1)
* [2. 便捷使用](#2)
+ [2.1 命令行使用](#21)
- [1. 安装PaddleOCR whl包](#1)
- [2. 便捷使用](#2)
- [2.1 命令行使用](#21)
- [2.1.1 中英文模型](#211)
- [2.1.2 多语言模型](#212)
- [2.1.3 版面分析](#213)
+ [2.2 Python脚本使用](#22)
- [2.2 Python脚本使用](#22)
- [2.2.1 中英文与多语言使用](#221)
- [2.2.2 版面分析](#222)
......
doc/doc_ch/serving_inference.md
浏览文件 @ 4b214948
......@@ -20,7 +20,7 @@
**Python操作指南：**
目前Serving用于OCR的部分功能还在测试当中，因此在这里我们给出[Servnig latest package](https://github.com/PaddlePaddle/Serving/blob/develop/doc/LATEST_PACKAGES.md)
目前Serving用于OCR的部分功能还在测试当中，因此在这里我们给出[Servnig latest package](https://github.com/PaddlePaddle/Serving/blob/develop/doc/Latest_Packages_CN.md)
大家根据自己的环境选择需要安装的whl包即可，例如以Python 3.5为例，执行下列命令
```
#CPU/GPU版本选择一个
......
doc/doc_ch/thirdparty.md
浏览文件 @ 4b214948
......@@ -21,10 +21,26 @@ PaddleOCR希望可以通过AI的力量助力任何一位有梦想的开发者实
| 通用工具 | [FastOCRLabel](https://gitee.com/BaoJianQiang/FastOCRLabel) | 完整的C#版本标注GUI | [包建强](https://gitee.com/BaoJianQiang) |
| 通用工具 | [DangoOCR离线版](https://github.com/PantsuDango/DangoOCR) | 通用型桌面级即时翻译GUI | [PantsuDango](https://github.com/PantsuDango) |
| 通用工具 | [scr2txt](https://github.com/lstwzd/scr2txt) | 截屏转文字GUI | [lstwzd](https://github.com/lstwzd) |
| 通用工具 | [ocr_sdk](https://github.com/mymagicpower/AIAS/blob/main/1_image_sdks/text_recognition/ocr_sdk) | OCR java SDK工具箱 | [Calvin](https://github.com/mymagicpower) |
| 通用工具 | [iocr](https://github.com/mymagicpower/AIAS/blob/main/8_suite_hub/iocr) | IOCR 自定义模板识别(支持表格识别) | [Calvin](https://github.com/mymagicpower) |
| 通用工具 | [Lmdb Dataset Format Conversion Tool](https://github.com/OneYearIsEnough/PaddleOCR-Recog-LmdbDataset-Conversion) | 文本识别任务中lmdb数据格式转换工具 | [OneYearIsEnough](https://github.com/OneYearIsEnough) |
| 通用工具 | [用paddleocr打造一款“盗幕笔记”](https://github.com/kjf4096/paddleocr_dmbj) | 用PaddleOCR记笔记 | [kjf4096](https://github.com/kjf4096) |
| 垂类工具 | [AI Studio项目](https://aistudio.baidu.com/aistudio/projectdetail/1054614?channelType=0&channel=0) | 英文视频自动生成字幕 | [叶月水狐](https://aistudio.baidu.com/aistudio/personalcenter/thirdview/322052) |
| 垂类工具 | [id_card_ocr](https://github.com/baseli/id_card_ocr) | 身份证复印件识别 | [baseli](https://github.com/baseli) |
| 垂类工具 | [Paddle_Table_Image_Reader](https://github.com/thunder95/Paddle_Table_Image_Reader) | 能看懂表格图片的数据助手 | [thunder95](https://github.com/thunder95]) |
| 垂类工具 | [AI Studio项目](https://aistudio.baidu.com/aistudio/projectdetail/3382897) | OCR流程中对手写体进行过滤 | [daassh](https://github.com/daassh) |
| 垂类场景调优 | [AI Studio项目](https://aistudio.baidu.com/aistudio/projectdetail/2803693) | 电表读数和编号识别 | [深渊上的坑](https://github.com/edencfc) |
| 垂类场景调优 | [AI Studio项目](https://aistudio.baidu.com/aistudio/projectdetail/3284199) | LCD液晶字符检测 | [Dream拒杰](https://github.com/zhangyingying520) |
| 前后处理 | [paddleOCRCorrectOutputs](https://github.com/yuranusduke/paddleOCRCorrectOutputs) | 获取OCR识别结果的key-value | [yuranusduke](https://github.com/yuranusduke) |
|前处理| [optlab](https://github.com/GreatV/optlab) |OCR前处理工具箱，基于Qt和Leptonica。|[GreatV](https://github.com/GreatV)|
|应用部署| [PaddleOCRSharp](https://github.com/raoyutian/PaddleOCRSharp) |PaddleOCR的.NET封装与应用部署。|[raoyutian](https://github.com/raoyutian/PaddleOCRSharp)|
|应用部署| [PaddleSharp](https://github.com/sdcb/PaddleSharp) |PaddleOCR的.NET封装与应用部署，支持跨平台、GPU|[sdcb](https://github.com/sdcb)|
| 应用部署 | [PaddleOCR-Streamlit-Demo](https://github.com/Lovely-Pig/PaddleOCR-Streamlit-Demo) | 使用Streamlit部署PaddleOCR | [Lovely-Pig](https://github.com/Lovely-Pig) |
| 应用部署 | [PaddleOCR-PyWebIO-Demo](https://github.com/Lovely-Pig/PaddleOCR-PyWebIO-Demo) | 使用PyWebIO部署PaddleOCR | [Lovely-Pig](https://github.com/Lovely-Pig) |
| 应用部署 | [PaddleOCR-Paddlejs-Vue-Demo](https://github.com/Lovely-Pig/PaddleOCR-Paddlejs-Vue-Demo) | 使用Paddle.js和Vue部署PaddleOCR | [Lovely-Pig](https://github.com/Lovely-Pig) |
| 应用部署 | [PaddleOCR-Paddlejs-React-Demo](https://github.com/Lovely-Pig/PaddleOCR-Paddlejs-React-Demo) | 使用Paddle.js和React部署PaddleOCR | [Lovely-Pig](https://github.com/Lovely-Pig) |
| 学术前沿模型训练与推理 | [AI Studio项目](https://aistudio.baidu.com/aistudio/projectdetail/3397137) | StarNet-MobileNetV3算法–中文训练 | [xiaoyangyang2](https://github.com/xiaoyangyang2) |
| 学术前沿模型训练与推理 | [ABINet-paddle](https://github.com/Huntersdeng/abinet-paddle) | ABINet算法前向运算的paddle实现以及模型各部分的实现细节分析 | [Huntersdeng](https://github.com/Huntersdeng) |
### 1.2 为PaddleOCR新增功能
......@@ -32,15 +48,22 @@ PaddleOCR希望可以通过AI的力量助力任何一位有梦想的开发者实
- 非常感谢 [tangmq](https://gitee.com/tangmq) 给PaddleOCR增加Docker化部署服务，支持快速发布可调用的Restful API服务([#507](https://github.com/PaddlePaddle/PaddleOCR/pull/507))。
- 非常感谢 [lijinhan](https://github.com/lijinhan) 给PaddleOCR增加java SpringBoot 调用OCR Hubserving接口完成对OCR服务化部署的使用([#1027](https://github.com/PaddlePaddle/PaddleOCR/pull/1027))。
- 非常感谢 [Evezerest](https://github.com/Evezerest)， [ninetailskim](https://github.com/ninetailskim)， [edencfc](https://github.com/edencfc)， [BeyondYourself](https://github.com/BeyondYourself)， [1084667371](https://github.com/1084667371) 贡献了[PPOCRLabel](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/PPOCRLabel/README_ch.md) 的完整代码。
- 非常感谢 [bupt906](https://github.com/bupt906) 贡献MicroNet结构代码([#5251](https://github.com/PaddlePaddle/PaddleOCR/pull/5251))和贡献OneCycle学习率策略代码([#5252](https://github.com/PaddlePaddle/PaddleOCR/pull/5252))
### 1.3 代码与文档优化
### 1.3 代码修复
- 非常感谢 [zhangxin](https://github.com/ZhangXinNan)([Blog](https://blog.csdn.net/sdlypyzq)) 贡献新的可视化方式、添加.gitgnore、处理手动设置PYTHONPATH环境变量的问题([#210](https://github.com/PaddlePaddle/PaddleOCR/pull/210))。
- 非常感谢 [lyl120117](https://github.com/lyl120117) 贡献打印网络结构的代码([#304](https://github.com/PaddlePaddle/PaddleOCR/pull/304))。
- 非常感谢 [BeyondYourself](https://github.com/BeyondYourself) 给PaddleOCR提了很多非常棒的建议，并简化了PaddleOCR的部分代码风格([so many commits)](https://github.com/PaddlePaddle/PaddleOCR/commits?author=BeyondYourself)。
### 1.4 文档优化与翻译
- 非常感谢 **[RangeKing](https://github.com/RangeKing)，[HustBestCat](https://github.com/HustBestCat)，[v3fc](https://github.com/v3fc)，[1084667371](https://github.com/1084667371)** 贡献翻译《动手学OCR》notebook[电子书英文版](https://github.com/PaddlePaddle/PaddleOCR/tree/dygraph/notebook/notebook_en)。
- 非常感谢 [thunderstudying](https://github.com/thunderstudying)，[RangeKing](https://github.com/RangeKing)，[livingbody](https://github.com/livingbody)， [WZMIAOMIAO](https://github.com/WZMIAOMIAO)，[haigang1975](https://github.com/haigang1975) 补充多个英文markdown文档。
- 非常感谢 **[fanruinet](https://github.com/fanruinet)** 润色和修复35篇英文文档([#5205](https://github.com/PaddlePaddle/PaddleOCR/pull/5205))。
- 非常感谢 [Khanh Tran](https://github.com/xxxpsyduck) 和 [Karl Horky](https://github.com/karlhorky) 贡献修改英文文档。
### 1.4 多语言语料
### 1.5 多语言语料
- 非常感谢 [xiangyubo](https://github.com/xiangyubo) 贡献手写中文OCR数据集([#321](https://github.com/PaddlePaddle/PaddleOCR/pull/321))。
- 非常感谢 [Mejans](https://github.com/Mejans) 给PaddleOCR增加新语言奥克西坦语Occitan的字典和语料([#954](https://github.com/PaddlePaddle/PaddleOCR/pull/954))。
......@@ -60,9 +83,9 @@ PaddleOCR非常欢迎社区贡献以PaddleOCR为核心的各种服务、部署
如果您在使用PaddleOCR时遇到了代码bug、功能不符合预期等问题，可以为PaddleOCR贡献您的修改，其中：
- Python代码规范可参考[附录1：Python代码规范](./code_and_doc.md#附录1)。
- Python代码规范可参考[附录1：Python代码规范](./code_and_doc.md/#附录1)。
- 提交代码前请再三确认不会引入新的bug，并在PR中描述优化点。如果该PR解决了某个issue，请在PR中连接到该issue。所有的PR都应该遵守附录3中的[3.2.10 提交代码的一些约定。](./code_and_doc.md#提交代码的一些约定)
- 提交代码前请再三确认不会引入新的bug，并在PR中描述优化点。如果该PR解决了某个issue，请在PR中连接到该issue。所有的PR都应该遵守附录3中的[3.2.10 提交代码的一些约定。](./code_and_doc.md/#提交代码的一些约定)
- 请在提交之前参考下方的[附录3：Pull Request说明](./code_and_doc.md#附录3)。如果您对git的提交流程不熟悉，同样可以参考附录3的3.2节。
......@@ -70,7 +93,7 @@ PaddleOCR非常欢迎社区贡献以PaddleOCR为核心的各种服务、部署
### 2.3 文档优化
如果您在使用PaddleOCR时遇到了文档表述不清楚、描述缺失、链接失效等问题，可以为PaddleOCR贡献您的修改。文档书写规范请参考[附录2：文档规范](./code_and_doc.md#附录2)。**最后请在PR的题目中加上标签`【third-party】` , 在说明中@Evezerest，拥有此标签的PR将会被高优处理。**
如果您在使用PaddleOCR时遇到了文档表述不清楚、描述缺失、链接失效等问题，可以为PaddleOCR贡献您的修改。文档书写规范请参考[附录2：文档规范](./code_and_doc.md/#附录2)。**最后请在PR的题目中加上标签`【third-party】` , 在说明中@Evezerest，拥有此标签的PR将会被高优处理。**
## 3. 更多贡献机会
......@@ -91,3 +114,30 @@ PaddleOCR非常欢迎社区贡献以PaddleOCR为核心的各种服务、部署
- 合入代码之后会在本文档第一节中更新信息，默认链接为github名字及主页，如果有需要更换主页，也可以联系我们。
- 新增重要功能类，会在用户群广而告之，享受开源社区荣誉时刻。
- **如果您有基于PaddleOCR的项目，但未出现在上述列表中，请按照 `4. 联系我们` 的步骤与我们联系。**
## 附录：社区常规赛积分榜
| 开发者 | 总积分 | 开发者 | 总积分 |
| ------------------------------------------------------- | ------ | ----------------------------------------------------- | ------ |
| [RangeKing](https://github.com/RangeKing) | 220 | [WZMIAOMIAO](https://github.com/WZMIAOMIAO) | 36 |
| [hao6699](https://github.com/hao6699) | 145 | [v3fc](https://github.com/v3fc) | 35 |
| [mymagicpower](https://github.com/mymagicpower) | 140 | [imiyu](https://github.com/imiyu) | 30 |
| [raoyutian](https://github.com/raoyutian) | 90 | [haigang1975](https://github.com/haigang1975) | 29 |
| [sdcb](https://github.com/sdcb) | 80 | [daassh](https://github.com/daassh) | 23 |
| [zhiminzhang0830](https://github.com/zhiminzhang0830) | 70 | [xiaoyangyang2](https://github.com/xiaoyangyang2) | 20 |
| [Lovely-Pig](https://github.com/Lovely-Pig) | 70 | [prettyocean85](https://github.com/prettyocean85) | 20 |
| [livingbody](https://github.com/livingbody) | 70 | [nmusik](https://github.com/nmusik) | 20 |
| [fanruinet](https://github.com/fanruinet) | 70 | [kjf4096](https://github.com/kjf4096) | 20 |
| [bupt906](https://github.com/bupt906) | 60 | [chccc1994](https://github.com/chccc1994) | 20 |
| [edencfc](https://github.com/edencfc) | 57 | [BeyondYourself ](https://github.com/BeyondYourself) | 20 |
| [zhangyingying520](https://github.com/zhangyingying520) | 57 | chenguoqi08161 | 18 |
| [ITerydh](https://github.com/ITerydh) | 55 | [weiwenlan](https://github.com/weiwenlan) | 10 |
| [telppa](https://github.com/telppa) | 40 | [shaoshenchen thinc](https://github.com/shaoshenchen) | 10 |
| sosojust1984 | 40 | [jordan2013](https://github.com/jordan2013) | 10 |
| [redearly123](https://github.com/redearly123) | 40 | [JimEverest](https://github.com/JimEverest) | 10 |
| [OneYearIsEnough](https://github.com/OneYearIsEnough) | 40 | [HustBestCat](https://github.com/HustBestCat) | 10 |
| [Huntersdeng](https://github.com/Huntersdeng) | 40 | | |
| [GreatV](https://github.com/GreatV) | 40 | | |
| CLXK294 | 40 | | |
doc/doc_en/config_en.md
浏览文件 @ 4b214948
......@@ -66,7 +66,7 @@ In PaddleOCR, the network is divided into four stages: Transform, Backbone, Neck
| :---------------------: | :---------------------: | :--------------: | :--------------------: |
| model_type | Network Type | rec | Currently support`rec`,`det`,`cls` |
| algorithm | Model name | CRNN | See [algorithm_overview](./algorithm_overview_en.md) for the support list |
| **Transform** | Set the transformation method | - | Currently only recognition algorithms are supported, see [ppocr/modeling/transform](../../ppocr/modeling/transform) for details |
| **Transform** | Set the transformation method | - | Currently only recognition algorithms are supported, see [ppocr/modeling/transforms](../../ppocr/modeling/transforms) for details |
| name | Transformation class name | TPS | Currently supports `TPS` |
| num_fiducial | Number of TPS control points | 20 | Ten on the top and bottom |
| loc_lr | Localization network learning rate | 0.1 | \ |
......
doc/doc_en/models_list_en.md
浏览文件 @ 4b214948
......@@ -4,13 +4,14 @@
> 2. Compared with [models 1.1](https://github.com/PaddlePaddle/PaddleOCR/blob/develop/doc/doc_en/models_list_en.md), which are trained with static graph programming paradigm, models 2.0 are the dynamic graph trained version and achieve close performance.
> 3. All models in this tutorial are all ppocr-series models, for more introduction of algorithms and models based on public dataset, you can refer to [algorithm overview tutorial](./algorithm_overview_en.md).
- [1. Text Detection Model](#Detection)
- [2. Text Recognition Model](#Recognition)
- [2.1 Chinese Recognition Model](#Chinese)
- [2.2 English Recognition Model](#English)
- [2.3 Multilingual Recognition Model](#Multilingual)
- [3. Text Angle Classification Model](#Angle)
- [4. Paddle-Lite Model](#Paddle-Lite)
- [OCR Model List（V2.1, updated on 2021.9.6）](#ocr-model-listv21-updated-on-202196)
- [1. Text Detection Model](#1-text-detection-model)
- [2. Text Recognition Model](#2-text-recognition-model)
- [2.1 Chinese Recognition Model](#21-chinese-recognition-model)
- [2.2 English Recognition Model](#22-english-recognition-model)
- [2.3 Multilingual Recognition Model（Updating...）](#23-multilingual-recognition-modelupdating)
- [3. Text Angle Classification Model](#3-text-angle-classification-model)
- [4. Paddle-Lite Model](#4-paddle-lite-model)
The downloadable models provided by PaddleOCR include `inference model`, `trained model`, `pre-trained model` and `slim model`. The differences between the models are as follows:
......@@ -43,8 +44,8 @@ Relationship of the above models is as follows.
|model name|description|config|model size|download|
| --- | --- | --- | --- | --- |
|ch_PP-OCRv2_rec_slim|[New] Slim qunatization with distillation lightweight model, supporting Chinese, English, multilingual text detection|[ch_PP-OCRv2_rec.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml)| 9M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_train.tar) |
|ch_PP-OCRv2_rec|[New] Original lightweight model, supporting Chinese, English, multilingual text detection|[ch_PP-OCRv2_rec.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml)|8.5M|[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar) |
|ch_PP-OCRv2_rec_slim|[New] Slim qunatization with distillation lightweight model, supporting Chinese, English, multilingual text recognition|[ch_PP-OCRv2_rec.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml)| 9M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_train.tar) |
|ch_PP-OCRv2_rec|[New] Original lightweight model, supporting Chinese, English, multilingual text recognition|[ch_PP-OCRv2_rec_distillation.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec_distillation.yml)|8.5M|[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar) |
|ch_ppocr_mobile_slim_v2.0_rec|Slim pruned and quantized lightweight model, supporting Chinese, English and number recognition|[rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml)| 6M | [inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_train.tar) |
|ch_ppocr_mobile_v2.0_rec|Original lightweight model, supporting Chinese, English and number recognition|[rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml)|5.2M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_train.tar) / [pre-trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_pre.tar) |
|ch_ppocr_server_v2.0_rec|General model, supporting Chinese, English and number recognition|[rec_chinese_common_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_common_train_v2.0.yml)|94.8M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_train.tar) / [pre-trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_pre.tar) |
......@@ -93,6 +94,8 @@ For more supported languages, please refer to : [Multi-language model](./multi_l
## 4. Paddle-Lite Model
|Version|Introduction|Model size|Detection model|Text Direction model|Recognition model|Paddle-Lite branch|
|---|---|---|---|---|---|---|
|PP-OCRv2|extra-lightweight chinese OCR optimized model|11M|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_infer_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_infer_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_infer_opt.nb)|v2.10|
|PP-OCRv2(slim)|extra-lightweight chinese OCR optimized model|4.6M|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_slim_opt.nb)|v2.10|
|PP-OCRv2|extra-lightweight chinese OCR optimized model|11M|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer_opt.nb)|v2.9|
|PP-OCRv2(slim)|extra-lightweight chinese OCR optimized model|4.9M|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_opt.nb)|v2.9|
|V2.0|ppocr_v2.0 extra-lightweight chinese OCR optimized model|7.8M|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_det_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_rec_opt.nb)|v2.9|
......
doc/doc_en/quickstart_en.md
浏览文件 @ 4b214948
# PaddleOCR Quick Start
[PaddleOCR Quick Start](#paddleocr-quick-start)
+ [1. Install PaddleOCR Whl Package](#1-install-paddleocr-whl-package)
* [2. Easy-to-Use](#2-easy-to-use)
+ [2.1 Use by Command Line](#21-use-by-command-line)
......
doc/doc_en/training_en.md
浏览文件 @ 4b214948
......@@ -94,7 +94,7 @@ The current open source models, data sets and magnitudes are as follows:
- Chinese data set, LSVT street view data set crops the image according to the truth value, and performs position calibration, a total of 30w images. In addition, based on the LSVT corpus, 500w of synthesized data.
- Small language data set, using different corpora and fonts, respectively generated 100w synthetic data set, and using ICDAR-MLT as the verification set.
Among them, the public data sets are all open source, users can search and download by themselves, or refer to [Chinese data set](../doc_ch/datasets.md), synthetic data is not open source, users can use open source synthesis tools to synthesize by themselves. Synthesis tools include [text_renderer](https://github.com/Sanster/text_renderer), [SynthText](https://github.com/ankush-me/SynthText), [TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator) etc.
Among them, the public data sets are all open source, users can search and download by themselves, or refer to [Chinese data set](./datasets_en.md), synthetic data is not open source, users can use open source synthesis tools to synthesize by themselves. Synthesis tools include [text_renderer](https://github.com/Sanster/text_renderer), [SynthText](https://github.com/ankush-me/SynthText), [TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator) etc.
<a name="22-vertical-scene"></a>
......
doc/doc_en/tricks_en.md
浏览文件 @ 4b214948
......@@ -12,7 +12,7 @@ Here we have sorted out some Chinese OCR training and prediction tricks, which a
At present, ResNet_vd series and MobileNetV3 series are the backbone networks used in PaddleOCR, whether replacing the other backbone networks will help to improve the accuracy? What should be paid attention to when replacing?
- **Tips**
- Whether text detection or text recognition, the choice of backbone network is a trade-off between prediction effect and prediction efficiency. Generally, a larger backbone network is selected, e.g. ResNet101_vd, then the performance of the detection or recognition is more accurate, but the time cost will increase accordingly. And a smaller backbone network is selected, e.g. MobileNetV3_small_x0_35, the prediction speed is faster, but the accuracy of detection or recognition will be reduced. Fortunately, the detection or recognition effect of different backbone networks is positively correlated with the performance of ImageNet 1000 classification task. [**PaddleClas**](https://github.com/PaddlePaddle/PaddleClas/blob/master/README_en.md) have sorted out the 23 series of classification network structures, such as ResNet_vd、Res2Net、HRNet、MobileNetV3、GhostNet. It provides the top1 accuracy of classification, the time cost of GPU(V100 and T4) and CPU(SD 855), and the 117 pretrained models [**download addresses**](https://paddleclas-en.readthedocs.io/en/latest/models/models_intro_en.html).
- Whether text detection or text recognition, the choice of backbone network is a trade-off between prediction effect and prediction efficiency. Generally, a larger backbone network is selected, e.g. ResNet101_vd, then the performance of the detection or recognition is more accurate, but the time cost will increase accordingly. And a smaller backbone network is selected, e.g. MobileNetV3_small_x0_35, the prediction speed is faster, but the accuracy of detection or recognition will be reduced. Fortunately, the detection or recognition effect of different backbone networks is positively correlated with the performance of ImageNet 1000 classification task. [**PaddleClas**](https://github.com/PaddlePaddle/PaddleClas/blob/release/2.3/docs/en/models/models_intro_en.md) have sorted out the 23 series of classification network structures, such as ResNet_vd、Res2Net、HRNet、MobileNetV3、GhostNet. It provides the top1 accuracy of classification, the time cost of GPU(V100 and T4) and CPU(SD 855), and the 117 pretrained models [**download addresses**](https://paddleclas-en.readthedocs.io/en/latest/models/models_intro_en.html).
- Similar as the 4 stages of ResNet, the replacement of text detection backbone network is to determine those four stages to facilitate the integration of FPN like the object detection heads. In addition, for the text detection problem, the pre trained model in ImageNet1000 can accelerate the convergence and improve the accuracy.
......
doc/joinus.PNG
查看替换文件 @ 917606b4 浏览文件 @ 4b214948
doc/joinus.PNG

199.5 KB | W: | H:

doc/joinus.PNG

198.7 KB | W: | H:

doc/joinus.PNG
doc/joinus.PNG
doc/joinus.PNG
doc/joinus.PNG
  • 2-up
  • Swipe
  • Onion skin
notebook/notebook_ch/2.text_detection/文本检测FAQ.ipynb
浏览文件 @ 4b214948
......@@ -81,7 +81,7 @@
"\n",
"如果对某些层使用更小的学习率学习，静态图里还不是很方便，一个方法是在参数初始化的时候，给权重的属性设置固定的学习率，参考：https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/api/paddle/fluid/param_attr/ParamAttr_cn.html#paramattr\n",
"\n",
"实际上我们实验发现，直接加载模型去fine-tune，不设置某些层不同学习率，效果也都不错\n",
"实际上我们实验发现，直接加载模型去fine-tune，不设置某些层不同学习率，效果也都不错。\n",
"\n",
"**1.11 DB的预处理部分，图片的长和宽为什么要处理成32的倍数？**\n",
"\n",
......@@ -95,7 +95,7 @@
"\n",
"**1.13 PP-OCR检测效果不好，该如何优化？**\n",
"\n",
"A： 具体问题具体分析:\n",
"**A**： 具体问题具体分析:\n",
"- 如果在你的场景上检测效果不可用，首选是在你的数据上做finetune训练；\n",
"- 如果图像过大，文字过于密集，建议不要过度压缩图像，可以尝试修改检测预处理的resize逻辑，防止图像被过度压缩；\n",
"- 检测框大小过于紧贴文字或检测框过大，可以调整db_unclip_ratio这个参数，加大参数可以扩大检测框，减小参数可以减小检测框大小；\n",
......@@ -123,8 +123,8 @@
"\n",
"**A**：GPU加速预测推荐使用TensorRT。\n",
"- 1. 从[链接](https://paddleinference.paddlepaddle.org.cn/master/user_guides/download_lib.html)下载带TensorRT的Paddle安装包或者预测库。\n",
"- 2. 从Nvidia官网下载TensorRT版本，注意下载的TensorRT版本与paddle安装包中编译的TensorRT版本一致。\n",
"- 3. 设置环境变量LD_LIBRARY_PATH，指向TensorRT的lib文件夹\n",
"- 2. 从Nvidia官网下载[TensorRT](https://developer.nvidia.com/tensorrt)，注意下载的TensorRT版本与paddle安装包中编译的TensorRT版本一致。\n",
"- 3. 设置环境变量`LD_LIBRARY_PATH`，指向TensorRT的lib文件夹\n",
"```\n",
"export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<TensorRT-${version}/lib>\n",
"```\n",
......
notebook/notebook_ch/2.text_detection/文本检测理论篇.ipynb
浏览文件 @ 4b214948
......@@ -6,7 +6,7 @@
"collapsed": false
},
"source": [
"# OCR七日课之文本检测综述\n"
"# 文本检测算法理论\n"
]
},
{
......@@ -15,11 +15,11 @@
"collapsed": false
},
"source": [
"## 1. 文本检测\n",
"## 1 文本检测\n",
"\n",
"文本检测任务是找出图像或视频中的文字位置。不同于目标检测任务，目标检测不仅要解决定位问题，还要解决目标分类问题。\n",
"\n",
"文本在图像中的表现形式可以视为一种‘目标‘，通用的目标检测的方法也适用于文本检测，从任务本身上来看：\n",
"文本在图像中的表现形式可以视为一种‘目标’，通用的目标检测的方法也适用于文本检测，从任务本身上来看：\n",
"\n",
"- 目标检测：给定图像或者视频，找出目标的位置（box），并给出目标的类别；\n",
"- 文本检测：给定输入图像或者视频，找出文本的区域，可以是单字符位置或者整个文本行位置；\n",
......@@ -41,14 +41,14 @@
"1. 自然场景中文本具有多样性：文本检测受到文字颜色、大小、字体、形状、方向、语言、以及文本长度的影响；\n",
"2. 复杂的背景和干扰；文本检测受到图像失真，模糊，低分辨率，阴影，亮度等因素的影响；\n",
"3. 文本密集甚至重叠会影响文字的检测；\n",
"4. 文字存在局部一致性，文本行的一小部分，也可视为是独立的文本；\n",
"4. 文字存在局部一致性：文本行的一小部分，也可视为是独立的文本。\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/072f208f2aff47e886cf2cf1378e23c648356686cf1349c799b42f662d8ced00\"\n",
"width=\"1000\" ></center>\n",
"\n",
"<br><center>图3 文本检测场景</center>\n",
"\n",
"针对以上问题，衍生了很多基于深度学习的文本检测算法，解决自然场景文字检测问题，这些方法可以分为基于回归和基于分割的文本检测方法。\n",
"针对以上问题，衍生出了很多基于深度学习的文本检测算法，用于解决自然场景文字检测问题。这些方法可以分为基于回归和基于分割的文本检测方法。\n",
"\n",
"下一节将简要介绍基于深度学习技术的经典文字检测算法。"
]
......@@ -59,7 +59,7 @@
"collapsed": false
},
"source": [
"## 2. 文本检测方法介绍\n",
"## 2 文本检测方法介绍\n",
"\n",
"\n",
"近些年来基于深度学习的文本检测算法层出不穷，这些方法大致可以分为两类：\n",
......@@ -134,7 +134,7 @@
"\n",
"\n",
"\n",
"LOMO[19]针对长文本和弯曲文本问题，提出迭代的优化文本定位特征获取更精细的文本定位，该方法包括三个部分，坐标回归模块DR，迭代优化模块IRM以及任意形状表达模块SEM。分别用于生成文本大致区域，迭代优化文本定位特征，预测文本区域、文本中心线以及文本边界。迭代的优化文本特征可以更好的解决长文本定位问题以及获得更精确的文本区域定位。\n",
"LOMO[19]针对长文本和弯曲文本问题，提出迭代的优化文本定位特征获取更精细的文本定位。该方法包括三个部分：坐标回归模块DR，迭代优化模块IRM以及任意形状表达模块SEM。它们分别用于生成文本大致区域，迭代优化文本定位特征，预测文本区域、文本中心线以及文本边界。迭代的优化文本特征可以更好的解决长文本定位问题以及获得更精确的文本区域定位。\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/e90adf3ca25a45a0af0b84a181fbe2c4954be1fcca8f4049957128548b7131ef\"\n",
"width=\"1000\" ></center>\n",
"<br><center>图11 LOMO框架图</center>\n",
......@@ -228,7 +228,7 @@
"collapsed": false
},
"source": [
"## 3. 总结\n",
"## 3 总结\n",
"\n",
"本节介绍了近几年来文本检测领域的发展，包括基于回归、分割的文本检测方法，并分别列举并介绍了一些经典论文的方法思路。下一节以PaddleOCR开源库为例，详细介绍DBNet的算法原理以及核心代码实现。"
]
......
notebook/notebook_ch/3.text_recognition/文本识别实践部分.ipynb
浏览文件 @ 4b214948
......@@ -13,13 +13,15 @@
"\n",
"通过本章的学习，你可以掌握：\n",
"\n",
"1. 如何使用paddleocr whl 包快速完成文本识别预测\n",
"1. 如何使用PaddleOCR whl包快速完成文本识别预测\n",
"\n",
"2. CRNN的基本原理和网络结构\n",
"\n",
"3. 模型训练的必须步骤和调参方式\n",
"\n",
"4. 使用自定义的数据集训练网络\n"
"4. 使用自定义的数据集训练网络\n",
"\n",
"注：`paddleocr`指代`PaddleOCR whl包`"
]
},
{
......@@ -189,7 +191,7 @@
"source": [
"# 安装 PaddlePaddle GPU 版本\n",
"!pip install paddlepaddle-gpu\n",
"# 安装 paddleocr whl包\n",
"# 安装 PaddleOCR whl包\n",
"! pip install -U pip\n",
"! pip install paddleocr"
]
......@@ -202,7 +204,7 @@
"source": [
"### 1.2 快速预测文字内容\n",
"\n",
"paddleocr whl包会自动下载ppocr轻量级模型作为默认模型\n",
"PaddleOCR whl包会自动下载ppocr轻量级模型作为默认模型\n",
"\n",
"下面展示如何使用whl包进行识别预测：\n",
"\n",
......@@ -326,13 +328,13 @@
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/f6fae3ff66bd413fa182d75782034a2af6aab1994fa148a08e6565f3fb75b18d width=\"600\"></center>\n",
"\n",
"1)backbone：\n",
"1. backbone：\n",
"\n",
"卷积网络作为底层的骨干网络，用于从输入图像中提取特征序列。由于 `conv`、`max-pooling`、`elementwise` 和激活函数都作用在局部区域上，所以它们是平移不变的。因此，特征映射的每一列对应于原始图像的一个矩形区域(称为感受野)，并且这些矩形区域与它们在特征映射上对应的列从左到右的顺序相同。由于CNN需要将输入的图像缩放到固定的尺寸以满足其固定的输入维数，因此它不适合长度变化很大的序列对象。为了更好的支持变长序列，CRNN将backbone最后一层输出的特征向量送到了RNN层，转换为序列特征。\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/6694818123724b0d92d05b63dc9dfb08c7ced6c47c3b4f4d9b110ae9ccfe941d width=\"600\"></center>\n",
"\n",
"2)neck: \n",
"2. neck: \n",
"\n",
"递归层，在卷积网络的基础上，构建递归网络，将图像特征转换为序列特征，预测每个帧的标签分布。\n",
"RNN具有很强的捕获序列上下文信息的能力。使用上下文线索进行基于图像的序列识别比单独处理每个像素更有效。以场景文本识别为例，宽字符可能需要几个连续的帧来充分描述。此外，有些歧义字符在观察其上下文时更容易区分。其次，RNN可以将误差差分反向传播回卷积层，使网络可以统一训练。第三，RNN能够对任意长度的序列进行操作，解决了文本图片变长的问题。CRNN使用双层LSTM作为递归层，解决了长序列训练过程中的梯度消失和梯度爆炸问题。\n",
......@@ -340,7 +342,7 @@
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/41cdb7fb08fb4b55923b0baf66b783e46fd063223d05416fa952369ad20ac83c width=\"600\"></center>\n",
"\n",
"\n",
"3)head: \n",
"3. head: \n",
"\n",
"转录层，通过全连接网络和softmax激活函数，将每帧的预测转换为最终的标签序列。最后使用 CTC Loss 在无需序列对齐的情况下，完成CNN和RNN的联合训练。CTC 有一套特别的合并序列机制，LSTM输出序列后，需要在时序上分类得到预测结果。可能存在多个时间步对应同一个类别，因此需要对相同结果进行合并。为避免合并本身存在的重复字符，CTC 引入了一个 `blank` 字符插入在重复字符之间。\n",
"\n",
......@@ -429,7 +431,7 @@
},
{
"data": {
"image/png": "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\n",
"image/png": "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",
"text/plain": [
"<Figure size 432x288 with 2 Axes>"
]
......@@ -464,7 +466,7 @@
"\n",
"* backbone\n",
"\n",
"PaddleOCR 使用 MobileNetV3 作为骨干网络，组网顺序与网络结构一致，首先定义网络中的公共模块([源码位置](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/ppocr/modeling/backbones/rec_mobilenet_v3.py))：ConvBNLayer、ResidualUnit、make_divisible"
"PaddleOCR 使用 MobileNetV3 作为骨干网络，组网顺序与网络结构一致。首先，定义网络中的公共模块([源码位置](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/ppocr/modeling/backbones/rec_mobilenet_v3.py))：`ConvBNLayer`、`ResidualUnit`、`make_divisible`。"
]
},
{
......@@ -647,7 +649,7 @@
"collapsed": false
},
"source": [
"利用公共模块搭建骨干网络"
"利用公共模块搭建骨干网络："
]
},
{
......@@ -990,7 +992,7 @@
"source": [
"* neck\n",
"\n",
"neck 部分将backbone输出的视觉特征图转换为1维向量输入送到 LSTM 网络中，输出序列特征（ [源码位置](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/ppocr/modeling/necks/rnn.py) ）："
"neck 部分将backbone输出的视觉特征图转换为1维向量输入送到 LSTM 网络中，输出序列特征（[源码位置](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/ppocr/modeling/necks/rnn.py)）："
]
},
{
......@@ -1354,7 +1356,7 @@
"source": [
"确认配置文件中的数据路径是否正确，以 [rec_icdar15_train.yml](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/configs/rec/rec_icdar15_train.yml)为例：\n",
"\n",
"```\n",
"```yaml\n",
"Train:\n",
" dataset:\n",
" name: SimpleDataSet\n",
......@@ -1460,7 +1462,7 @@
"outputs": [
{
"data": {
"image/png": "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\n",
"image/png": "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",
"text/plain": [
"<Figure size 432x288 with 2 Axes>"
]
......@@ -1572,12 +1574,11 @@
"collapsed": false
},
"source": [
"实现完单条数据返回逻辑后，调用 `padde.io.Dataloader` 即可把数据组合成batch，具体可参考 [build_dataloader]()\n",
"\n",
"实现完单条数据返回逻辑后，调用 `padde.io.Dataloader` 即可把数据组合成batch，具体可参考 [build_dataloader](https://github.com/PaddlePaddle/PaddleOCR/blob/95c670faf6cf4551c841764cde43a4f4d9d5e634/ppocr/data/__init__.py#L52)。\n",
"\n",
"* build model\n",
"\n",
" build model 即搭建主要网络结构，具体细节如《2.3 代码实现》所述，本节不做过多介绍，各模块代码可参考[modeling](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.3/ppocr/modeling)\n",
" build model 即搭建主要网络结构，具体细节如《2.3 代码实现》所述，本节不做过多介绍，各模块代码可参考[modeling](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.4/ppocr/modeling)\n",
"\n",
"* build loss\n",
" \n",
......
notebook/notebook_ch/4.ppcor_system_strategy/PP-OCR系统及优化策略.ipynb 0 → 100644
浏览文件 @ 4b214948
因为 它太大了无法显示 source diff 。你可以改为 查看blob。
notebook/notebook_ch/如何使用本书.ipynb
浏览文件 @ 4b214948
......@@ -42,7 +42,7 @@
"\n",
"然后安装第三方库：\n",
"\n",
"```\n",
"```bash\n",
"cd PaddleOCR\n",
"pip3 install -r requirements.txt\n",
"```\n",
......
notebook/notebook_en/1.introduction/introduction_to_OCR_technology.ipynb 0 → 100644
浏览文件 @ 4b214948
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"![](https://ai-studio-static-online.cdn.bcebos.com/72b2077605dd49b78f7f647d6821d10231f6bc52d7ed463da451a6a0bd1fc5ff)\n",
"*Note: The above pictures are from the Internet*\n",
"\n",
"# 1. OCR Technical Background\n",
"## 1.1 Application Scenarios of OCR Technology\n",
"\n",
"* **<font color=red>What is OCR</font>**\n",
"\n",
"OCR (Optical Character Recognition) is one of the key directions in computer vision. The traditional definition of OCR is generally oriented to scanned document objects. Now we often say OCR generally refers to scene text recognition (Scene Text Recognition, STR), mainly for natural scenes, such as plaques and other visible texts in various natural scenes as shown in the figure below.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/c87c0e6f6c0a42cdbc552a4f973c1b0217c369194c1243558753896f3e66032c)\n",
"<center>Figure 1: Document scene text recognition VS. Natural scene text recognition</center>\n",
"\n",
"<br>\n",
"\n",
"* **<font color=red>What are the application scenarios of OCR? </font>**\n",
"\n",
"OCR technology has a wealth of application scenarios. A typical scenario is vertically-oriented structured text recognition widely used in daily life, such as license plate recognition, bank card information recognition, ID card information recognition, train ticket information recognition, and so on. The common feature of these small verticals is that the format is fixed. Therefore, it is very suitable to use OCR technology for automation, greatly reducing labor costs and improving.\n",
"\n",
"This vertically-oriented structured text recognition is currently the most widely used and relatively mature technology scene in OCR.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/56e0df91d0d34443aacb17c9a1c5c186608ee675092648a693503df7fe45e535)\n",
"<center>Figure 2: Application scenarios of OCR technology</center>\n",
"\n",
"In addition to vertically-oriented structured text recognition, general OCR technology also has a wide range of applications and is often combined with other technologies to complete multi-modal tasks. For example, in video scenes, OCR technology is often used for subtitle automatic translation, content security monitoring, etc., Or combined with visual features to complete tasks such as video understanding and video search.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/ca2341a51eb242ee8e1afe121ce3ebbc87a113cef1b643ed9bba92d0c8ee4f0f)\n",
"<center>Figure 3: General OCR in a multi-modal scene</center>\n",
"\n",
"## 1.2 OCR Technical Challenge\n",
"The technical difficulties of OCR can be divided into two aspects: the algorithm layer and the application layer.\n",
"\n",
"* **<font color=red>Algorithm layer</font>**\n",
"\n",
"The rich application scenarios of OCR determine that it will have many technical difficulties. Here are 8 common problems:\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/a56831fbf0c449fe9156a893002cadfe110ccfea835b4d90854a7ce4b1df2a4f)\n",
"<center>Figure 4: Technical difficulties of OCR algorithm layer</center>\n",
"\n",
"These problems bring huge technical challenges to both text detection and text recognition. It can be seen that these challenges are mainly oriented to natural scenes. At present, research in academia mainly focuses on natural scenes, and the commonly used academic datasets in the OCR field are also natural scenes. There are many studies on these issues. Relatively speaking, identification is more challenging than detection.\n",
"\n",
"* **<font color=red>Application layer</font>**\n",
"\n",
"In practical applications, especially in a wide range of general scenarios, in addition to the technical difficulties at the algorithm level such as affine transformation, scale problems, insufficient lighting, and shooting blur summarized in the previous section, OCR technology also faces two major difficulties:\n",
"1. **Massive data requires OCR to be able to process in real time.** OCR applications are often connected to massive data. Real-time processing of the data is required or hoped for. Real-time model speed is a big challenge.\n",
"2. **The end-side application requires that the OCR model is light enough and the recognition speed is fast enough.** OCR applications are often deployed on mobile terminals or embedded hardware. There are generally two modes for terminal-side OCR applications: upload to server vs. terminal-side direct recognition. Considering that the method of uploading to the server has requirements on the network, the real-time performance is low, and the server pressure is high when the request volume is too large, as well as the security of data transmission, we hope to complete the OCR identification directly on the terminal side. However, the storage space and computing power of the terminal side are limited, so there are high requirements for the size and prediction speed of the OCR model.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/5bafdc3da1614c41a95ae39a2c36632f95e2893031a64929b9f49d4a4985cd2d)\n",
"<center>Figure 5: Technical difficulties of OCR application layer</center>"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 2. OCR Cutting-edge Algorithm\n",
"\n",
"Although OCR is a relatively specific task, it involves many aspects of technology, including text detection, text recognition, end-to-end text recognition, document analysis, and so on. Academic research on various related technologies of OCR emerges endlessly. The following will briefly introduce the related work of several key technologies in the OCR task.\n",
"\n",
"## 2.1 Text Detection\n",
"\n",
"The task of text detection is to locate text regions in the input image. In recent years, research on text detection in academia has been very rich. A class of methods regard text detection as a specific scene in target detection, and improve and adapt based on general target detection algorithms. For example, TextBoxes[1] is based on one-stage target detector SSD. The algorithm [2] adjusts the target frame to fit text lines with extreme aspect ratios, while CTPN [3] is improved based on the Faster RCNN [4] architecture. However, there are still some differences between text detection and target detection in the target information and the task itself. For example, the text is generally larger in length and width, often in the shape of \"stripes\", and the text lines may be denser, curved text, etc. Therefore, many algorithms dedicated to text detection have been derived, such as EAST[5], PSENet[6], DBNet[7] and so on.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/548b50212935402abb2e671c158c204737c2c64b9464442a8f65192c8a31b44d\" width=\"500\"></center>\n",
"<center>Figure 6: Example of text detection task</center>\n",
"\n",
"<br>\n",
"\n",
"At present, the more popular text detection algorithms can be roughly divided into two categories: **based on regression** and **based on segmentation**. There are also some algorithms that combine the two. Algorithms based on regression draw on general object detection algorithms, by setting the anchor regression detection frame, or directly doing pixel regression. This type of method has a better detection effect on regular-shaped text, but the detection effect on irregularly-shaped text will be relatively poor. For example, CTPN [3] has better detection effect on horizontal text, but poor detection effect on oblique and curved text. SegLink [8] is more effective for long text, but has limited effect on sparsely distributed text; algorithm based on segmentation Introduced Mask-RCNN [9], this type of algorithm can reach a higher level in various scenes and texts of various shapes, but the disadvantage is that the post-processing is generally more complicated, so there are often speed problems. And it cannot solve the problem of detecting overlapping text.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/4f4ea65578384900909efff93d0b7386e86ece144d8c4677b7bc94b4f0337cfb\" width=\"800\"></center>\n",
"<center>Figure 7: Overview of text detection algorithms</center>\n",
"\n",
"<br>\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/473ba28cd0274d568f90eb8ca9e78864d994f3ebffe6419cb638e193c607b7b3)|![](https://ai-studio-static-online.cdn.bceb8dc)|![](https://ai-studio-static-online.cdn.bcebos.com/53b9e85ce46645c08481d7d7377720f5eea5ac30e37e4e9c9930e1f26b02e278)\n",
"|---|---|---|\n",
"<center>Figure 8: (left) CTPN[3] algorithm optimization based on regression anchor (middle) DB[7] algorithm optimization post-processing based on segmentation (right) SAST[10] algorithm of regression + segmentation</center>\n",
"\n",
"<br>\n",
"\n",
"The related technology of text detection will be interpreted and actual combat in detail in Chapter 2.\n",
"\n",
"## 2.2 Text Recognition\n",
"\n",
"The task of text recognition is to recognize the text content in the image, and the input generally comes from the text area of the image cut out by the text box obtained by text detection. Text recognition can generally be divided into two categories: **Regular Text Recognition** and **Irregular Text Recognition** according to the shape of the text to be recognized. Regular text mainly refers to printed fonts, scanned text, etc., and the text is roughly in the horizontal line position. Irregular text is often not in a horizontal position, and has problems such as bending, occlusion, and blurring. Irregular text scenes are very challenging, and it is also the main research direction in the field of text recognition.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/b292f21e50c94debab7496d4ced96a93774a8525c12346f49cb151bde2a58fe8)\n",
"<center>Figure 9: (Left) Regular text VS. (Right) Irregular text</center>\n",
"\n",
"<br>\n",
"\n",
"The algorithm of regular text recognition can be roughly divided into two types based on CTC and Sequence2Sequence according to the different decoding methods. The processing methods of converting the sequence features learned by the network into the final recognition result are different. The algorithm based on CTC is represented by the classic CRNN [11].\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/403ca85c59d344f88d3b1229ca14b1e90c5c73c9f1d248b7aa94103f9d0af597)\n",
"<center>Figure 10: CTC-based recognition algorithm VS. Attention-based recognition algorithm</center>\n",
"\n",
"The recognition algorithms for irregular texts are more abundant. Methods such as STAR-Net [12] correct the irregular texts into regular rectangles by adding correction modules such as TPS before recognition. Attention-based methods such as RARE [13] enhance the attention to the correlation of parts between sequences. The segmentation-based method treats each character of a text line as an independent individual, and it is easier to recognize a single segmented character than to recognize the entire text line after correction. In addition, with the rapid development of Transformer [14] and its effectiveness in various tasks in recent years, a number of Transformer-based text recognition algorithms have also appeared. These methods use the transformer structure to solve the long-dependency modeling of CNN. The limitations of the problem, but also achieved good results.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/0fa30c3789424473ad9be1c87a4f742c1db69e3defb64651906e5334ed9571a8)\n",
"<center>Figure 11: Recognition algorithm based on character segmentation [15]</center>\n",
"\n",
"<br>\n",
"\n",
"The related technologies of text recognition will be interpreted and actual combat in detail in Chapter 3.\n",
"\n",
"## 2.3 Document Structure Recognition\n",
"\n",
"OCR technology in the traditional sense can solve the detection and recognition needs of text. However, in practical application scenarios, structured information is often needed in the end, such as information formatting and extraction of ID cards and invoices, structured identification of tables, and so on. The application scenarios of OCR technology are mostly express document extraction, contract content comparison, financial factoring document information comparison, and logistics document identification. OCR result + post-processing is a commonly used structuring scheme, but the process is often complicated, and post-processing requires fine design and poor generalization. Under the background of the gradual maturity of OCR technology and the growing demand for structured information extraction, various technologies related to intelligent document analysis, such as layout analysis, table recognition, and key information extraction, have received more and more attention and research.\n",
"\n",
"* **Layout Analysis**\n",
"\n",
"Layout Analysis is mainly used to classify the content of document images. The categories can generally be divided into plain text, titles, tables, pictures, etc. Existing methods generally regard different plates in the document as different targets for detection or segmentation. For example, Soto Carlos [16], based on the target detection algorithm Faster R-CNN, combines context information and uses the inherent position information of the document content to improve the performance. Region detection performance. Sarkar Mausoom et al.[17] proposed a priori-based segmentation mechanism to train a document segmentation model on very high-resolution images, solving the problem that different structures in dense regions cannot be distinguished and merged due to excessive reduction of the original image.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/dedb212e8972497998685ff51af7bfe03fdea57f6acd450281ad100807086e1a)\n",
"<center>Figure 12: Schematic diagram of layout analysis tasks</center>\n",
"\n",
"<br>\n",
"\n",
"* **Table Recognition**\n",
"\n",
"The task of table recognition is to identify and convert the table information in the document into an excel file. The types and styles of tables in text images are complex and diverse, such as different row and column combinations, different content text types, etc. In addition, the style of the document and the lighting environment during shooting have brought great challenges to table recognition. These challenges make table recognition always a research difficulty in the field of document understanding.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/47119a2a2f9a45788390d6506f90d5de7449738008aa4c0ab619b18f37bd8d57)\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/22ca5749441441e69dc0eaeb670832a5d0ae0ce522f34731be7d609a2d36e8c1)\n",
"<center>Figure 13: Schematic diagram of form recognition task</center>\n",
"\n",
"<br>\n",
"\n",
"There are many types of table recognition methods. The early traditional algorithms based on heuristic rules, such as the T-Rect algorithm proposed by Kieninger [18] and others, generally use manual design rules and connected domain detection and analysis. In recent years, with the development of deep learning, some CNN-based table structure recognition algorithms have emerged, such as DeepTabStR proposed by Siddiqui Shoaib Ahmed [19] and others, and TabStruct-Net proposed by Raja Sachin [20] and others. In addition, with the rise of *Graph Neural Network*, some researchers try to apply *Graph Neural Network* to the problem of table structure recognition. Based on the *Graph Neural Network*, table recognition is regarded as a graph reconstruction problem, such as Xue Wenyuan [21] TGRNet proposed by et al. The end-to-end method directly uses the network to complete the HTML representation output of the table structure. Most of the end-to-end methods use the Seq2Seq method to complete the prediction of the table structure, such as some methods based on Attention or Transformer, such as TableMaster [22].\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/a9a3c91898c84f03b382583859526c4b451ace862dbc4a15838f5dde4d0ea657)\n",
"<center>Figure 14: Schematic diagram of form identification method</center>\n",
"\n",
"<br>\n",
"\n",
"* **Key Information Extraction**\n",
"\n",
"Key Information Extraction (KIE) is an important task in Document VQA. It mainly extracts the key information needed from images, such as extracting name and citizen ID number information from ID cards. The types of such information are often It is fixed under a specific task, but is different between different tasks.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/8af011647bb4464f80d07f3efeac469baed27c8185ef4c4883a19f40e8ba91f5)\n",
"<center>Figure 15: Schematic diagram of DocVQA tasks</center>\n",
"\n",
"<br>\n",
"\n",
"KIE is usually divided into two sub-tasks for research:\n",
"\n",
"- SER: Semantic Entity Recognition, to classify each detected text, such as dividing it into name and ID. As shown in the black box and red box in the figure below.\n",
"- RE: Relation Extraction, which classifies each detected text, such as dividing it into questions and answers. Then find the corresponding answer to each question. As shown in the figure below, the red and black boxes represent the question and the answer, respectively, and the yellow line represents the correspondence between the question and the answer.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/2f1bc1a3e4a341ab9552bbf5f6c2be71ba78d7d65da64818b776efe0691e310b)\n",
"<center>Figure 16: ser and re tasks</center>\n",
"\n",
"<br>\n",
"\n",
"The general KIE method is researched based on Named Entity Recognition (NER) [4], but this type of method only uses the text information in the image and lacks the use of visual and structural information, so the accuracy is not high. On this basis, the methods in recent years have begun to merge visual and structural information with text information. According to the principles used when fusing multi-modal information, these methods can be divided into the following four types:\n",
"\n",
"- Grid-based method\n",
"- Token-based method\n",
"- GCN-based method\n",
"- Based on End to End method\n",
"\n",
"<br>\n",
"\n",
"Document analysis related technologies will be explained and actual combat in detail in Chapter 6.\n",
"\n",
"## 2.4 Other Related Technologies\n",
"\n",
"The previous mainly introduced three key technologies in the OCR field: text detection, text recognition, document structured recognition, and more other cutting-edge technologies related to OCR, including end-to-end text recognition, image preprocessing technology in OCR, and OCR data synthesis Etc., please refer to Chapter 7 and Chapter 8 of the tutorial.\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 3. Industrial Practice of OCR Technology\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/3d5f18f7598f405884fa2fab041c95ce415af40712e9489996747f9d122c3d90)\n",
"\n",
"> You are Xiao Wang, what should I do?\n",
"> 1. I won't, I can't, I won't do it 😭\n",
"> 2. It is recommended that the boss find an outsourcing company or commercialization plan, anyway, spend the boss's money 😊\n",
"> 3. Find similar projects online, programming for Github😏\n",
"\n",
"<br>\n",
"\n",
"OCR technology will eventually fall into industrial practice. Although there is a lot of academic research on OCR technology, and the commercial application of OCR technology is relatively mature compared with other AI technologies, there are still some difficulties and challenges in actual industrial applications. The following will analyze from two perspectives of technology and industrial practice.\n",
"\n",
"\n",
"## 3.1 Difficulties in Industrial Practice\n",
"\n",
"In actual industrial practice, developers often need to rely on open source community resources to start or promote projects, and developers using open source models often face three major problems:\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/7e5e79240b9c4f13b675b56bc12edf540f159c922bf24e3cbc4a0635a356c7f9)\n",
"<center>Figure 17: Three major problems in the practice of OCR technology industry</center>\n",
"\n",
"**1. Can't find & can't choose**\n",
"\n",
"The open source community is rich in resources, but information asymmetry makes developers unable to solve pain points efficiently. On the one hand, the open source community resources are too rich. Faced with a requirement, developers cannot quickly find a project that matches the business requirement from the massive code repository, that is, there is a problem of \"can't find\". On the other hand, when selecting algorithms, the indicators on the English public dataset cannot provide a direct reference for the Chinese scenarios that developers often face. Algorithm-by-algorithm verification takes a lot of time and manpower, and there is no guarantee that the most suitable algorithm will be selected, that is, \"can't choose\".\n",
"\n",
"**2. Not applicable to industry scenarios**\n",
"\n",
"The work in the open source community tends to focus more on effect optimization, such as open source or reproduction of academic paper codes, and generally focus more on algorithm effects. Compared with the work that balances the size and speed of the model, it is much less, and the model size and prediction are time-consuming Two indicators that cannot be ignored in industrial practice are as important as the model effect. Whether it is on the mobile side or the server side, the number of images to be recognized is often very large, and it is hoped that the model will be smaller, more accurate, and faster in prediction. GPU is too expensive, it is better to use CPU to run more economically. On the premise of meeting business needs, the lighter the model, the less resources it takes.\n",
"\n",
"**3. Difficult optimization and many training deployment problems**\n",
"\n",
"The direct use of open source algorithms or models generally cannot directly meet business needs. In actual business scenarios, OCR faces a variety of problems. The personalization of business scenarios often requires retraining of custom data sets. On existing open source projects, various optimizations are experimented. The cost of the method is higher. In addition, OCR application scenarios are very rich. There are a wide range of application requirements on the server and various mobile devices. The diversification of the hardware environment needs to support rich deployment methods. The open source community’s projects focus more on algorithms and models, and predict deployment. This part is obviously under-supported. To apply OCR technology from the algorithm in the paper to the application of technology, it has high requirements for the algorithm and engineering ability of the developer.\n",
"\n",
"## 3.2 Industrial OCR Development Kit PaddleOCR\n",
"\n",
"OCR industry practice requires a complete set of full-process solutions to speed up the research and development progress and save valuable research and development time. In other words, the ultra-lightweight model and its full-process solutions, especially for mobile terminals and embedded devices with limited computing power and storage space, can be said to be a rigid demand.\n",
"\n",
"In this context, the industrial-grade OCR development kit [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR) came into being.\n",
"\n",
"The construction idea of ​​PaddleOCR starts from user portraits and needs, and relies on the core framework of flying oars, selects and reproduces a wealth of cutting-edge algorithms, and develops PP characteristic models that are more suitable for industrial landing based on recurring algorithms, and integrates training and promotion to provide A variety of predictive deployment methods to meet different demand scenarios of actual applications.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/e09929b4a31e44f9b5e3d542d12411332669d2e1a21d45ad88b1dd91142ec86c)\n",
"<center>Figure 18: Panorama of PaddleOCR development kit</center>\n",
"\n",
"<br>\n",
"\n",
"It can be seen from the panorama that PaddleOCR relies on the core framework of the flying paddle, and provides a wealth of solutions in the model algorithm, pre-training model library, industrial-grade deployment, etc., and provides data synthesis and semi-automatic data annotation tools to meet the needs of development Data production needs of the author.\n",
"\n",
"**At the model algorithm level**, PaddleOCR provides solutions for the two tasks of **text detection and recognition** and **document structure analysis** respectively. In terms of text detection and recognition, PaddleOCR has reproduced or open sourced 4 text detection algorithms, 8 text recognition algorithms, and 1 end-to-end text recognition algorithm. On this basis, a general text detection and recognition solution of the PP-OCR series is developed. In terms of document structure analysis, PaddleOCR provides algorithms such as layout analysis, table recognition, key information extraction, and named entity recognition, and based on this, it proposes a PP-Structure document analysis solution. A rich selection of algorithms can meet the needs of developers in different business scenarios. The unification of the code framework also facilitates the optimization and performance comparison of different algorithms for developers.\n",
"\n",
"**At the level of pre-training model library**, based on PP-OCR and PP-Structure solutions, PaddleOCR has developed and open-sourced PP series characteristic models suitable for industrial practice, including general-purpose, ultra-lightweight and multi-language text detection and recognition Models, and complex document analysis models. The PP series characteristic models are deeply optimized on the original algorithm, so that they can reach the practical level of the industry in terms of effect and performance. Developers can either directly apply to business scenarios or use business data for simple finetune. Easily develop a \"practical model\" suitable for your business needs.\n",
"\n",
"**At the industrial level of deployment**, PaddleOCR provides a server-side prediction solution based on Paddle Inference, a service-based deployment solution based on Paddle Serving, and an end-side deployment solution based on Paddle-Lite to meet the deployment needs of different hardware environments , At the same time, it provides a model compression scheme based on PaddleSlim, which can further compress the model size. The above deployment methods have completed the whole process of training and pushing to ensure that developers can deploy efficiently, stably and reliably.\n",
"\n",
"**At the data tool level**, PaddleOCR provides a semi-automatic data annotation tool PPOCRLabel and a data synthesis tool Style-Text to help developers more conveniently produce the data sets and annotation information required for model training. PPOCRLabel, as the industry's first open source semi-automatic OCR data annotation tool, is aimed at the tedious and tedious process of labeling, high mechanicality, manual labeling of a large amount of training data, and expensive time and money. The built-in PP-OCR model realizes pre-labeling + manual verification. The labeling mode can greatly improve labeling efficiency and save labor costs. The data synthesis tool Style-Text mainly solves the serious shortage of real data in actual scenes. Traditional synthesis algorithms cannot synthesize text styles (fonts, colors, spacing, background). Only a few target scene images are needed to synthesize a large number of target scene styles in batches. Similar text images.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/90a358d6a62c49b7b8db47e18c77878c60f80cf9c81541bfa3befea68d9dbc0f)\n",
"<center>Figure 19: Schematic diagram of PPOCRLabel usage</center>\n",
"<br>\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/b63b10bc530c42bea3d3b923da6000f1cfef006d7eec4ff3bdc0439bd9c333c9)\n",
"<center>Figure 20: Example of Style-Text synthesis effect</center>\n",
"\n",
"<br>\n",
"\n",
"### 3.2.1 PP-OCR and PP-Structrue\n",
"\n",
"The PP series characteristic model is a model that is deeply optimized for the practical needs of the industry by various visual development kits of the flying propeller, striving for a balance between speed and accuracy. The PP series featured models in PaddleOCR include PP-OCR series models for text detection and recognition tasks and PP-Structure series models for document analysis.\n",
"\n",
"**(1) PP-OCR Chinese and English model**\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/3372558042044d43983b815069e1e43cb84432b993ed400f946976e75bd51f38)\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/f0a0b936382c42dd8809e98759b4c84434d79386606b4d5b8a86416db6dbaeee)\n",
"<center>Figure 21: Example of PP-OCR model recognition results in Chinese and English</center>\n",
"\n",
"<br>\n",
"\n",
"The typical two-stage OCR algorithm adopted by the Chinese and English models of PP-OCR, that is, the composition method of detection model + recognition model, the specific algorithm framework is as follows:\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/8af1371b5e3c486bb90a041903200c7c666c8bbc98c245dc802ff8c4da98617e)\n",
"<center>Figure 22: Schematic diagram of PP-OCR system pipeline</center>\n",
"\n",
"<br>\n",
"\n",
"It can be seen that in addition to input and output, the core framework of PP-OCR contains 3 modules, namely: text detection module, detection frame correction module, and text recognition module.\n",
"- Text detection module: The core is a text detection model trained on the [DB](https://arxiv.org/abs/1911.08947) detection algorithm to detect the text area in the image;\n",
"- Detection frame correction module: Input the detected text box into the detection frame correction module. At this stage, the text box represented by the four points is corrected into a rectangular frame, which is convenient for subsequent text recognition. On the other hand, the text direction will be judged and corrected. For example, if the text line is judged to be upside down, it will be corrected. This function is realized by training a text direction classifier;\n",
"- Text recognition module: Finally, the text recognition module performs text recognition on the corrected detection box to obtain the text content in each text box. The classic text recognition algorithm used in PP-OCR [CRNN](https://arxiv.org/abs/1507.05717).\n",
"\n",
"PaddleOCR has successively introduced PP-OCR[23] and PP-OCRv2[24] models.\n",
"\n",
"PP-OCR model is divided into mobile version (lightweight version) and server version (universal version). The mobile version model is mainly optimized based on the lightweight backbone network MobileNetV3. The optimized model (detection model + text direction classification model + recognition model) ) The size is only 8.1M, the average single image prediction on the CPU takes 350ms, and the T4 GPU is about 110ms. After cropping and quantization, it can be further compressed to 3.5M without changing the accuracy, which is convenient for end-side deployment. The previous test predicts that it will only take 260ms. For more PP-OCR evaluation data, please refer to [benchmark](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.2/doc/doc_ch/benchmark.md).\n",
"\n",
"PP-OCRv2 maintains the overall framework of PP-OCR, mainly for further strategic optimization of effects. The improvement includes 3 aspects:\n",
"- Compared with the PP-OCR mobile version, the model effect is improved by over 7%;\n",
"- In terms of speed, compared to the PP-OCR server version, it has increased by more than 220%;\n",
"- In terms of model size, with a total size of 11.6M, both server and mobile terminals can be easily deployed.\n",
"\n",
"The specific optimization strategies of PP-OCR and PP-OCRv2 will be explained in detail in Chapter 4.\n",
"\n",
"In addition to the Chinese and English models, PaddleOCR also trained and open-sourced English digital models and multi-language recognition models based on different data sets. All of the above are ultra-lightweight models and are suitable for different language scenarios.\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/5978652a826647b98344cf61aa1c2027662af989b73e4a0e917d83718422eeb0)\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/1a8a8e24b5a440d388dae767adf0ea9c049335b04e964abbb176f58c5b028d7e)\n",
"<center>Figure 23: Schematic diagram of the recognition effect of the English digital model and multilingual model of PP-OCR</center>\n",
"\n",
"<br>\n",
"\n",
"**(2) PP-Structure document analysis model**\n",
"\n",
"PP-Structure supports three subtasks: layout analysis, table recognition, and DocVQA.\n",
"\n",
"The core functions of PP-Structure are as follows:\n",
"- Support layout analysis of documents in the form of pictures, which can be divided into 5 types of areas: text, title, table, picture and list (used in conjunction with Layout-Parser)\n",
"- Support text, title, picture and list area to be extracted as text fields (used in conjunction with PP-OCR)\n",
"- Supports structured analysis in the table area, and the final result is output to an Excel file\n",
"- Support Python whl package and command line two ways, simple and easy to use\n",
"- Support custom training for two types of tasks: layout analysis and table structuring\n",
"- Support VQA tasks-SER and RE\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/129708c265644dbc90d6c8f7db224b3a6f11f37bb586463a82e7ccb50bcc2e76)\n",
"<center>Figure 24: Schematic diagram of PP-Structure system (this figure only contains layout analysis + table identification)</center>\n",
"\n",
"<br>\n",
"\n",
"The specific plan of PP-Structure will be explained in detail in Chapter 6.\n",
"\n",
"### 3.2.2 Industrial-grade Deployment Plan\n",
"\n",
"The flying paddle supports full-process and full-scene inference deployment. There are three main sources of models. The first one uses PaddlePaddle API to build a network structure for training. The second is based on the flying paddle kit series. The flying paddle kit provides a wealth of models. Library, simple and easy-to-use API, with out-of-the-box use, including visual model library PaddleCV, intelligent speech library PaddleSpeech and natural language processing library PaddleNLP, etc. The third type uses X2Paddle tools from third-party frameworks (PyTorh, ONNX, TensorFlow, etc.) The output model.\n",
"\n",
"The paddle model can be compressed, quantified, and distilled using PaddleSlim tools. It supports five deployment schemes, namely, servicing Paddle Serving, server/cloud Paddle Inference, mobile/edge Paddle Lite, web front end Paddle.js, and for Paddle. Unsupported hardware, such as MCU, Horizon, Kunyun and other domestic chips, can be converted into a third-party framework that supports ONNX with the help of Paddle2ONNX.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/c9ffe78e7db14e4eb103e7f393a16fbf2ab438540250474a8e0e7adc4aeb7ee0)\n",
"<center>Figure 25: Flying propeller support deployment method</center>\n",
"\n",
"<br>\n",
"\n",
"Paddle Inference supports server-side and cloud deployment, with high performance and versatility. It is deeply adapted and optimized for different platforms and different application scenarios. Paddle Inference is the native reasoning library for flying paddles, ensuring that the model can be trained on the server side. Use, rapid deployment, suitable for high-performance hardware using multiple application language environments to deploy models with complex algorithms. The hardware covers x86 CPUs, Nvidia GPUs, and AI accelerators such as Baidu Kunlun XPU and Huawei Shengteng.\n",
"\n",
"Paddle Lite is an end-side inference engine with lightweight and high-performance features. It has been configured and optimized in-depth for end-side devices and various application scenarios. Currently supports multiple platforms such as Android, IOS, embedded Linux devices, macOS, etc. The hardware covers ARM CPU and GPU, X86 CPU and new hardware such as Baidu Kunlun, Huawei Ascend and Kirin, Rockchip, etc.\n",
"\n",
"Paddle Serving is a high-performance service framework designed to help users quickly deploy models in cloud services in a few steps. At present, Paddle Serving supports functions such as custom pre-processing, model combination, model hot load update, multi-machine multi-card multi-model, distributed reasoning, K8S deployment, security gateway and model encryption deployment, and support for multi-language and multi-client access. Paddle Serving The official also provides deployment examples of more than 40 models, including PaddleOCR, to help users get started faster.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/4d8063d74194434ea9b7c9f81c7fbdfd2131e13770124d2e99c1b9670f12e019)\n",
"<center>Figure 26: Support deployment mode of flying propeller</center>\n",
"\n",
"<br>\n",
"\n",
"The above deployment plan will be explained in detail and actual combat based on the PP-OCRv2 model in Chapter 5."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 4. Summary\n",
"\n",
"This section first introduces the application scenarios and cutting-edge algorithms of OCR technology, and then analyzes the difficulties and three major challenges of OCR technology in industrial practice.\n",
"\n",
"The contents of the subsequent chapters of this tutorial are arranged as follows:\n",
"\n",
"* The second and third chapters introduce detection and identification technology and practice respectively;\n",
"* Chapter 4 introduces PP-OCR optimization strategy;\n",
"* Chapter 5 Predicting and deploying actual combat;\n",
"* Chapter 6 introduces document structuring;\n",
"* Chapter 7 introduces other OCR-related algorithms such as end-to-end, data preprocessing, and data synthesis;\n",
"* Chapter 8 introduces OCR related data sets and data synthesis tools.\n",
"\n",
"# Reference\n",
"\n",
"[1] Liao, Minghui, et al. \"Textboxes: A fast text detector with a single deep neural network.\" Thirty-first AAAI conference on artificial intelligence. 2017.\n",
"\n",
"[2] Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector[C]//European conference on computer vision. Springer, Cham, 2016: 21-37.\n",
"\n",
"[3] Tian, Zhi, et al. \"Detecting text in natural image with connectionist text proposal network.\" European conference on computer vision. Springer, Cham, 2016.\n",
"\n",
"[4] Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-time object detection with region proposal networks[J]. Advances in neural information processing systems, 2015, 28: 91-99.\n",
"\n",
"[5] Zhou, Xinyu, et al. \"East: an efficient and accurate scene text detector.\" Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. 2017.\n",
"\n",
"[6] Wang, Wenhai, et al. \"Shape robust text detection with progressive scale expansion network.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019.\n",
"\n",
"[7] Liao, Minghui, et al. \"Real-time scene text detection with differentiable binarization.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 07. 2020.\n",
"\n",
"[8] Deng, Dan, et al. \"Pixellink: Detecting scene text via instance segmentation.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 32. No. 1. 2018.\n",
"\n",
"[9] He K, Gkioxari G, Dollár P, et al. Mask r-cnn[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2961-2969.\n",
"\n",
"[10] Wang P, Zhang C, Qi F, et al. A single-shot arbitrarily-shaped text detector based on context attended multi-task \n",
"learning[C]//Proceedings of the 27th ACM international conference on multimedia. 2019: 1277-1285.\n",
"\n",
"[11] Shi, B., Bai, X., & Yao, C. (2016). An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE transactions on pattern analysis and machine intelligence, 39(11), 2298-2304.\n",
"\n",
"[12] Star-Net Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spa- tial transformer networks. In Advances in neural information processing systems, pages 2017–2025, 2015.\n",
"\n",
"[13] Shi, B., Wang, X., Lyu, P., Yao, C., & Bai, X. (2016). Robust scene text recognition with automatic rectification. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4168-4176).\n",
"\n",
"[14] Sheng, F., Chen, Z., & Xu, B. (2019, September). NRTR: A no-recurrence sequence-to-sequence model for scene text recognition. In 2019 International Conference on Document Analysis and Recognition (ICDAR) (pp. 781-786). IEEE.\n",
"\n",
"[15] Lyu P, Liao M, Yao C, et al. Mask textspotter: An end-to-end trainable neural network for spotting text with arbitrary shapes[C]//Proceedings of the European Conference on Computer Vision (ECCV). 2018: 67-83.\n",
"\n",
"[16] Soto C, Yoo S. Visual detection with context for document layout analysis[C]//Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019: 3464-3470.\n",
"\n",
"[17] Sarkar M, Aggarwal M, Jain A, et al. Document Structure Extraction using Prior based High Resolution Hierarchical Semantic Segmentation[C]//European Conference on Computer Vision. Springer, Cham, 2020: 649-666.\n",
"\n",
"[18] Kieninger T, Dengel A. A paper-to-HTML table converting system[C]//Proceedings of document analysis systems (DAS). 1998, 98: 356-365.\n",
"\n",
"[19] Siddiqui S A, Fateh I A, Rizvi S T R, et al. Deeptabstr: Deep learning based table structure recognition[C]//2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, 2019: 1403-1409.\n",
"\n",
"[20] Raja S, Mondal A, Jawahar C V. Table structure recognition using top-down and bottom-up cues[C]//European Conference on Computer Vision. Springer, Cham, 2020: 70-86.\n",
"\n",
"[21] Xue W, Yu B, Wang W, et al. TGRNet: A Table Graph Reconstruction Network for Table Structure Recognition[J]. arXiv preprint arXiv:2106.10598, 2021.\n",
"\n",
"[22] Ye J, Qi X, He Y, et al. PingAn-VCGroup's Solution for ICDAR 2021 Competition on Scientific Literature Parsing Task B: Table Recognition to HTML[J]. arXiv preprint arXiv:2105.01848, 2021.\n",
"\n",
"[23] Du Y, Li C, Guo R, et al. PP-OCR: A practical ultra lightweight OCR system[J]. arXiv preprint arXiv:2009.09941, 2020.\n",
"\n",
"[24] Du Y, Li C, Guo R, et al. PP-OCRv2: Bag of Tricks for Ultra Lightweight OCR System[J]. arXiv preprint arXiv:2109.03144, 2021.\n",
"\n"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "py35-paddle1.2.0"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.4"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
notebook/notebook_en/2.text_detection/text_detection_FAQ.ipynb 0 → 100644
浏览文件 @ 4b214948
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Text detection FAQ\n",
"\n",
"This section lists some of the problems that developers often encounter when using PaddleOCR's text detection model, and gives corresponding solutions or suggestions.\n",
"\n",
"The FAQ is introduced in two parts, namely:\n",
" -Text detection training related\n",
" -Text detection and prediction related"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 1. FAQ about Text Detection Training\n",
"\n",
"**1.1 What are the text detection algorithms provided by PaddleOCR?**\n",
"\n",
"**A**: PaddleOCR contains a variety of text detection models, including regression-based text detection methods EAST and SAST, and segmentation-based text detection methods DB, PSENet.\n",
"\n",
"\n",
"**1.2: What data sets are used in the Chinese ultra-lightweight and general models in the PaddleOCR project? How many samples were trained, what configuration of GPUs, how many epochs were run, and how long did they run?**\n",
"\n",
"**A**: For the ultra-lightweight DB detection model, the training data includes open source data sets lsvt, rctw, CASIA, CCPD, MSRA, MLT, BornDigit, iflytek, SROIE and synthetic data sets, etc. The total data volume is 10W, The data set is divided into 5 parts. A random sampling strategy is used during training. The training takes about 500 epochs on a 4-card V100GPU, which takes 3 days.\n",
"\n",
"\n",
"**1.3 Does the text detection training label require specific text labeling? What does the \"###\" in the label mean?**\n",
"\n",
"**A**: Text detection training only needs the coordinates of the text area. The label can be four or fourteen points, arranged in the order of upper left, upper right, lower right, and lower left. The label file provided by PaddleOCR contains text fields. For unclear text in the text area, ### will be used instead. When training the detection model, the text field in the label will not be used.\n",
" \n",
"**1.4 Is the effect of the text detection model trained when the text lines are tight?**\n",
"\n",
"**A**: When using segmentation-based methods, such as DB, to detect dense text lines, it is best to collect a batch of data for training, and during training, a binary image will be generated [shrink_ratio](https://github.com/PaddlePaddle/PaddleOCR/blob/8b656a3e13631dfb1ac21d2095d4d4a4993ef710/ppocr/data/imaug/make_shrink_map.py?_pjax=%23js-repo-pjax-container%2C%20div%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L37)Turn down the parameter. In addition, when forecasting, you can appropriately reduce [unclip_ratio](https://github.com/PaddlePaddle/PaddleOCR/blob/8b656a3e13631dfb1ac21d2095d4d4a4993ef710/configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml?_pjax=%23js-repo-pjax-container%2C%20div%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L59) parameter, the larger the unclip_ratio parameter value, the larger the detection frame.\n",
"\n",
"\n",
"**1.5 For some large-sized document images, DB will have more missed inspections during inspection. How to avoid this kind of missed inspections?**\n",
"\n",
"**A**: First of all, you need to determine whether the model is not well-trained or is the problem handled during prediction. If the model is not well trained, it is recommended to add more data for training, or add more data to enhance it during training.\n",
"If the problem is that the predicted image is too large, you can increase the longest side setting parameter [det_limit_side_len] entered during prediction [det_limit_side_len](https://github.com/PaddlePaddle/PaddleOCR/blob/8b656a3e13631dfb1ac21d2095d4d4a4993ef710/tools/infer/utility.py?_pjax=%23js-repo-pjax-container%2C%20div%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L47), which is 960 by default.\n",
"Secondly, you can observe whether the missed text has segmentation results by visualizing the post-processed segmentation map. If there is no segmentation result, the model is not well trained. If there is a complete segmentation area, it means that it is a problem of post-prediction processing. In this case, it is recommended to adjust [DB post-processing parameters](https://github.com/PaddlePaddle/PaddleOCR/blob/8b656a3e13631dfb1ac21d2095d4d4a4993ef710/tools/infer/utility.py?_pjax=%23js-repo-pjax-container%2C%20div%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L51-L53)。\n",
"\n",
"\n",
"**1.6 The problem of missed detection of DB model bending text (such as a slightly deformed document image)?**\n",
"\n",
"**A**: When calculating the average score of the text box in the DB post-processing, it is the average score of the rectangle area, which is easy to cause the missed detection of the curved text. The average score of the polygon area has been added, which will be more accurate, but the speed is somewhat different. Decrease, can be selected as needed, and you can view the [Visual Contrast Effect] (https://github.com/PaddlePaddle/PaddleOCR/pull/2604) in the relevant pr. This function is selected by the parameter [det_db_score_mode](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.1/tools/infer/utility.py#L51), the parameter value is optional [`fast` (default) , `slow`], `fast` corresponds to the original rectangle mode, and `slow` corresponds to the polygon mode. Thanks to the user [buptlihang](https://github.com/buptlihang) for mentioning [pr](https://github.com/PaddlePaddle/PaddleOCR/pull/2574) to help solve this problem.\n",
"\n",
"\n",
"**1.7 For simple OCR tasks with low accuracy requirements, how many data sets do I need to prepare?**\n",
"\n",
"**A**: (1) The amount of training data is related to the complexity of the problem to be solved. The greater the difficulty and the higher the accuracy requirements, the greater the data set requirements, and in general, the more training data in practice, the better the effect.\n",
"\n",
"(2) For scenes with low accuracy requirements, the amount of data required for detection tasks and recognition tasks is different. For inspection tasks, 500 images can guarantee the basic inspection results. For recognition tasks, it is necessary to ensure that the number of line text images in which each character in the recognition dictionary appears in different scenes needs to be greater than 200 (for example, if there are 5 words in the dictionary, each word needs to appear in more than 200 pictures, then The minimum required number of images should be between 200-1000), so that the basic recognition effect can be guaranteed.\n",
"\n",
"\n",
"**1.8 How to get more data when the amount of training data is small?**\n",
"\n",
"**A**: When the amount of training data is small, you can try the following three ways to get more data: (1) Collect more training data manually, the most direct and effective way. (2) Basic image processing or transformation based on PIL and opencv. For example, the three modules of ImageFont, Image, ImageDraw in PIL write text into the background, opencv's rotating affine transformation, Gaussian filtering and so on. (3) Synthesize data using data generation algorithms, such as algorithms such as pix2pix.\n",
"\n",
"\n",
"**1.9 How to replace the backbone of text detection/recognition?**\n",
"\n",
"A: Whether it is text detection or text recognition, the choice of backbone network is a trade-off between prediction effect and prediction efficiency. Generally, if you choose a larger-scale backbone network, such as ResNet101_vd, the detection or recognition will be more accurate, but the prediction time will increase accordingly. However, choosing a smaller-scale backbone network, such as MobileNetV3_small_x0_35, will predict faster, but the accuracy of detection or recognition will be greatly reduced. Fortunately, the detection or recognition effects of different backbone networks are positively correlated with the image 1000 classification task in the ImageNet dataset. PaddleClas, a flying paddle image classification suite, summarizes 23 series of classification network structures such as ResNet_vd, Res2Net, HRNet, MobileNetV3, GhostNet, etc. The top1 recognition accuracy rate of the above image classification task, GPU (V100 and T4) and CPU (Snapdragon 855) The prediction time-consuming and the corresponding 117 pre-training model download addresses.\n",
"\n",
"(1) The replacement of the text detection backbone network is mainly to determine 4 stages similar to ResNet to facilitate the integration of subsequent detection heads similar to FPN. In addition, for the text detection problem, the classification pre-training model trained by ImageNet can accelerate the convergence and improve the effect.\n",
"\n",
"(2) The replacement of the backbone network for text recognition requires attention to the drop position of the network width and height stride. Since text recognition generally has a large ratio of width to height, the frequency of height reduction is less, and the frequency of width reduction is more. You can refer to [Changes to the MobileNetV3 backbone network in PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.3/ppocr/modeling/backbones/rec_mobilenet_v3.py)。\n",
"\n",
"\n",
"**1.10 How to finetune the detection model, such as freezing the previous layer or learning with a small learning rate for some layers?**\n",
"\n",
"**A**: If you freeze certain layers, you can set the stop_gradient property of the variable to True, so that all the parameters before calculating this variable will not be updated, refer to: https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/faq/train_cn.html#id4\n",
"\n",
"If learning with a smaller learning rate for some layers is not very convenient in the static graph, one method is to set a fixed learning rate for the weight attribute when the parameters are initialized, refer to: https://www.paddlepaddle.org.cn/documentation/docs/en/develop/api/paddle/fluid/param_attr/ParamAttr_cn.html#paramattr\n",
"\n",
"In fact, our experiment found that directly loading the model to fine-tune without setting different learning rates of certain layers, the effect is also good.\n",
"\n",
"**1.11 In the preprocessing part of DB, why should the length and width of the picture be processed into multiples of 32?**\n",
"\n",
"**A**: It is related to the stride of the network downsampling. Take the resnet backbone network under inspection as an example. After the image is input to the network, it needs to be downsampled by 2 times for 5 times, a total of 32 times. Therefore, it is recommended that the input image size be a multiple of 32.\n",
"\n",
"\n",
"**1.12 In the PP-OCR series models, why does the backbone network for text detection not use SEBlock?**\n",
"\n",
"**A**: The SE module is an important module of the MobileNetV3 network. Its purpose is to estimate the importance of each feature channel of the feature map, assign weights to each feature of the feature map, and improve the expressive ability of the network. However, for text detection, the resolution of the input network is relatively large, generally 640\\*640. It is difficult to use the SE module to estimate the importance of each feature channel of the feature map. The network improvement ability is limited, but the module is relatively time-consuming. In the PP-OCR system, the backbone network for text detection does not use the SE module. Experiments also show that when the SE module is removed, the size of the ultra-lightweight model can be reduced by 40%, and the text detection effect is basically not affected. For details, please refer to the PP-OCR technical article, https://arxiv.org/abs/2009.09941.\n",
"\n",
"\n",
"**1.13 The PP-OCR detection effect is not good, how to optimize it?**\n",
"\n",
"**A**: Specific analysis of specific issues:\n",
"- If the detection effect is not available on your scene, the first choice is to do finetune training on your data;\n",
"- If the image is too large and the text is too dense, it is recommended not to over-compress the image. You can try to modify the resize logic of the detection preprocessing to prevent the image from being over-compressed;\n",
"- The size of the detection frame is too close to the text or the detection frame is too large, you can adjust the db_unclip_ratio parameter, increasing the parameter can enlarge the detection frame, and reducing the parameter can reduce the size of the detection frame;\n",
"- There are many missed detection problems in the detection frame, which can reduce the threshold parameter det_db_box_thresh for DB detection to prevent some detection frames from being filtered out. You can also try to set det_db_score_mode to'slow';\n",
"- Other methods can choose use_dilation as True to expand the feature map of the detection output. In general, the effect will be improved.\n",
"\n",
"\n",
"## 2. FAQ about Text Detection and Prediction\n",
"\n",
"**2.1 In DB, some boxes are too pasted with text, but some corners of the text are removed to affect the recognition. Is there any way to alleviate this problem?**\n",
"\n",
"**A**: The post-processing parameter [unclip_ratio](https://github.com/PaddlePaddle/PaddleOCR/blob/d80afce9b51f09fd3d90e539c40eba8eb5e50dd6/tools/infer/utility.py?_pjax=%23js-repo-pjax-container%2C%20div%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L52) can be appropriately increased. the larger the parameter, the larger the text box.\n",
"\n",
"\n",
"**2.2 Why does the PaddleOCR detection prediction only support one image test? That is, test_batch_size_per_card=1**\n",
"\n",
"**A**: When predicting, the image is scaled in equal proportions, the longest side is 960, and the length and width of different images after scaling in equal proportions are inconsistent, and they cannot form a batch, so set test_batch_size to 1.\n",
"\n",
"\n",
"**2.3 Accelerate PaddleOCR's text detection model prediction on the CPU?**\n",
"\n",
"**A**: x86 CPU can use mkldnn (OneDNN) for acceleration; enable [enable_mkldnn](https://github.com/PaddlePaddle/PaddleOCR/blob/8b656a3e13631dfb1ac21d2095d4d4a4993ef710/tools/infer/utility.py#L105) Parameters. In addition, in conjunction with increasing the number of threads used for prediction on the CPU, [num_threads](https://github.com/PaddlePaddle/PaddleOCR/blob/8b656a3e13631dfb1ac21d2095d4d4a4993ef710/tools/infer/utility.py#L106) can effectively speed up the prediction speed on the CPU.\n",
"\n",
"**2.4 Accelerate PaddleOCR's text detection model prediction on GPU?**\n",
"\n",
"**A**: TensorRT is recommended for GPU accelerated prediction.\n",
"- 1. Download the Paddle installation package or prediction library with TensorRT from [link](https://paddleinference.paddlepaddle.org.cn/master/user_guides/download_lib.html).\n",
"- 2. Download the [TensorRT](https://developer.nvidia.com/tensorrt) from the Nvidia official website. Note that the downloaded TensorRT version is consistent with the TensorRT version compiled in the paddle installation package.\n",
"- 3. Set the environment variable `LD_LIBRARY_PATH` to point to the lib folder of TensorRT\n",
"```\n",
"export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<TensorRT-${version}/lib>\n",
"```\n",
"- 4. Enable [tensorrt option](https://github.com/PaddlePaddle/PaddleOCR/blob/8b656a3e13631dfb1ac21d2095d4d4a4993ef710/tools/infer/utility.py?_pjax=%23js-repo-pjax-container%2%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L38).\n",
"\n",
"**2.5 How to deploy PaddleOCR model on the mobile terminal?**\n",
"\n",
"**A**: Flying Oar Paddle has a special tool for mobile deployment [PaddleLite](https://github.com/PaddlePaddle/Paddle-Lite), and PaddleOCR provides DB+CRNN as the demo android arm deployment code , Refer to [link](https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.3/deploy/lite/readme.md).\n",
"\n",
"\n",
"**2.6 How to use PaddleOCR multi-process prediction?**\n",
"\n",
"**A**: PaddleOCR recently added [Multi-Process Predictive Control Parameters](https://github.com/PaddlePaddle/PaddleOCR/blob/8b656a3e13631dfb1ac21d2095d4d4a4993ef710/tools/infer/utility.py?_pjax=%23js-repo-pjax-container%2C%20div%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L111), `use_mp` indicates whether When using multiple processes, `total_process_num` indicates the number of processes when using multiple processes. For specific usage, please refer to [document](https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.3/doc/doc_ch/inference.md#1-%E8%B6%85%E8%BD%BB%E9%87%8F%E4%B8%AD%E6%96%87ocr%E6%A8%A1%E5%9E%8B%E6%8E%A8%E7%90%86).\n",
"\n",
"**2.7 Video memory explosion and memory leak during prediction?**\n",
"\n",
"**A**: If it is the prediction of the training model, the video memory is not enough because the model is too large or the input image is too large, you can refer to the code and add paddle.no_grad() before the main function runs to reduce the video memory usage. If the memory usage of the inference model is too high, you can add [config.enable_memory_optim()](https://github.com/PaddlePaddle/PaddleOCR/blob/8b656a3e13631dfb1ac21d2095d4d4a4993ef710/tools/infer/utility.py?_pjax=%23js-repo-pjax-container%2C%20div%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L267) to reduce the memory usage when configuring Config.\n",
"\n",
"In addition, regarding the memory leak when using Paddle to predict, it is recommended to install the latest version of paddle. The memory leak has been fixed."
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.8.5"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
notebook/notebook_en/2.text_detection/text_detection_practice.ipynb 0 → 100644
浏览文件 @ 4b214948
因为 它太大了无法显示 source diff 。你可以改为 查看blob。
notebook/notebook_en/2.text_detection/text_detection_theory.ipynb 0 → 100644
浏览文件 @ 4b214948
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Text Detection Algorithm Theory\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 1 Text Detection\n",
"\n",
"The task of text detection is to find out the position of text in an image or video. Different from the task of target detection, target detection must not only solve the positioning problem, but also solve the problem of target classification.\n",
"\n",
"The manifestation of text in images can be regarded as a kind of 'target', and general target detection methods are also suitable for text detection. From the perspective of the task itself:\n",
"\n",
"- Target detection: Given an image or video, find out the location (box) of the target, and give the target category;\n",
"- Text detection: Given an input image or video, find out the area of the text, which can be a single character position or a whole text line position;\n",
"\n",
"\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/af2d8eca913a4d5a968945ae6cac180b009c6cc94abc43bfbaf1ba6a3de98125\" width=\"400\" ></center>\n",
"\n",
"<br><center>Figure 1: Schematic diagram of target detection</center>\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/400b9100573b4286b40b0a668358bcab9627f169ab934133a1280361505ddd33\" width=\"1000\" ></center>\n",
"\n",
"<br><center>Figure 2: Schematic diagram of text detection</center>\n",
"\n",
"Object detection and text detection are both \"location\" problems. But text detection does not need to classify the target, and the shape of the text is complex and diverse.\n",
"\n",
"The current text detection is generally natural scene text detection. The difficulty lies in:\n",
"\n",
"1. Text in natural scenes is diverse: text detection is affected by text color, size, font, shape, direction, language, and text length;\n",
"2. Complex background and interference; text detection is affected by image distortion, blur, low resolution, shadow, brightness and other factors;\n",
"3. Dense or overlapping text will affect text detection;\n",
"4. Local consistency of text: a small part of a text line can also be considered as independent text.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/072f208f2aff47e886cf2cf1378e23c648356686cf1349c799b42f662d8ced00\"\n",
"width=\"1000\" ></center>\n",
"\n",
"<br><center>Figure 3: Text detection scene</center>\n",
"\n",
"In response to the above problems, many text detection algorithms based on deep learning have been derived to solve the problem of text detection in natural scenes. These methods can be divided into regression-based and segmentation-based text detection methods.\n",
"\n",
"The next section will briefly introduce the classic text detection algorithm based on deep learning technology."
]
},
{
"cell_type": "markdown",
"metadata": {
"tags": []
},
"source": [
"## 2 Introduction to Text Detection Methods\n",
"\n",
"\n",
"In recent years, deep learning-based text detection algorithms have emerged one after another. These methods can be roughly divided into two categories:\n",
"1. Regression-based text detection method\n",
"2. Text detection method based on segmentation\n",
"\n",
"\n",
"This section screens the commonly used text detection methods from 2017 to 2021, and is classified according to the above two types of methods as shown in the following table:\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/22314238b70b486f942701107ffddca48b87235a473c4d8db05b317f132daea0\"\n",
"width=\"600\" ></center>\n",
"<br><center>Figure 4: Text detection algorithm</center>\n",
"\n",
"\n",
"### 2.1 Regression-based Text Detection\n",
"\n",
"The method based on the regression text detection method is similar to the method of the target detection algorithm. The text detection method has only two categories. The text in the image is regarded as the target to be detected, and the rest is regarded as the background.\n",
"\n",
"#### 2.1.1 Horizontal Text Detection\n",
"\n",
"Early text detection algorithms based on deep learning are improved from the target detection method and support horizontal text detection. For example, the TextBoxes algorithm is improved based on the SSD algorithm, and the CTPN is improved based on the two-stage target detection Fast-RCNN algorithm.\n",
"\n",
"In TextBoxes[1], the algorithm is adjusted according to the one-stage target detector SSD, and the default text box is changed to a quadrilateral that adapts to the specifications of the text direction and aspect ratio, providing an end-to-end training text detection method without complicated Post-processing.\n",
"-Use a pre-selection box with a larger aspect ratio\n",
"-The convolution kernel has been changed from 3x3 to 1x5, which is more suitable for long text detection\n",
"-Adopt multi-scale input\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/3864ccf9d009467cbc04225daef0eb562ac0c8c36f9b4f5eab036c319e5f05e7\" width=\"1000\" ></center>\n",
"<br><center>Figure 5: Textbox frame diagram</center>\n",
"\n",
"CTPN[3] is based on the Fast-RCNN algorithm, expands the RPN module and designs a CRNN-based module to allow the entire network to detect text sequences from convolutional features. The two-stage method obtains more accurate feature positioning through ROI Pooling. But TextBoxes and CTPN only support the detection of horizontal text.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/452833c2016e4cf7b35291efd09740c13c4bfb8f7c56446b8f7a02fc7eb3e901\" width=\"1000\" ></center>\n",
"<br><center>Figure 6: CTPN frame diagram</center>\n",
"\n",
"#### 2.1.2 Any Angle Text Detection\n",
"\n",
"TextBoxes++[2] is improved on the basis of TextBoxes, and supports the detection of text at any angle. Structurally, unlike TextBoxes, TextBoxes++ detects multi-angle text. First, modify the aspect ratio of the preselection box and adjust the aspect ratio to 1, 2, 3, 5, 1/2, 1/3, 1/5. The second is to change the $1*5$ convolution kernel to $3*5$ to better learn the characteristics of the slanted text; finally, the representation information of the output rotating box of TextBoxes++.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/ae96e3acbac04be296b6d54a4d72e5881d592fcc91f44882b24bc7d38b9d2658\"\n",
"width=\"1000\" ></center>\n",
"<br><center>Figure 7: TextBoxes++ frame diagram</center>\n",
"\n",
"\n",
"EAST [4] proposed a two-stage text detection method for the location of slanted text, including FCN feature extraction and NMS part. EAST proposes a new text detection pipline structure, which can be trained end-to-end and supports the detection of text in any orientation, and has the characteristics of simple structure and high performance. FCN supports the output of inclined rectangular frame and horizontal frame, and the output format can be freely selected.\n",
"-If the output detection shape is RBox, output Box rotation angle and AABB text shape information, AABB represents the offset to the top, bottom, left, and right sides of the text box. RBox can rotate rectangular text.\n",
"-If the output detection box is a four-point box, the last dimension of the output is 8 numbers, which represents the position offset from the four corner vertices of the quadrilateral. This output method can predict irregular quadrilateral text.\n",
"\n",
"Considering that the text box output by FCN is relatively redundant, for example, the box generated by the adjacent pixels of a text area has a high degree of coincidence. But it is not the detection frame generated by the same text, and the degree of coincidence is very small. Therefore, EAST proposes to merge the prediction boxes by row first. Finally, filter the remaining quads with the original NMS.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/d7411ada08714adab73fa0edf7555a679327b71e29184446a33d81cdd910e4fc\"\n",
"width=\"1000\" ></center>\n",
"<br><center>Figure 8: EAST frame diagram</center>\n",
"\n",
"\n",
"MOST [15] proposed that the TFAM module dynamically adjusts the receptive field of coarse-grained detection results, and also proposed that PA-NMS combines reliable detection and prediction results based on location information. In addition, the Instance-wise IoU loss function is also proposed during training, which is used to balance training to handle text instances of different scales. This method can be combined with the EAST method, and has better detection effect and performance in detecting texts with extreme aspect ratios and different scales.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/73052d9439714bba86ffe4a959d58c523b07baf3f1d74882b4517e71f5a645fe\"\n",
"width=\"1000\" ></center>\n",
"<br><center>Figure 9: MOST frame diagram</center>\n",
"\n",
"\n",
"#### 2.1.3 Curved Text Detection\n",
"\n",
"Using regression to solve the problem of curved text detection, a simple idea is to describe the boundary polygon of the curved text with multi-point coordinates, and then directly predict the vertex coordinates of the polygon.\n",
"\n",
"CTD [6] proposed to directly predict the boundary polygons of 14 vertices of curved text. The network uses the Bi-LSTM [13] layer to refine the prediction coordinates of the vertices, and realizes the detection of curved text based on the regression method.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/6e33d76ebb814cac9ebb2942b779054af160857125294cd69481680aca2fa98a\"\n",
"width=\"600\" ></center>\n",
"<br><center>Figure 10: CTD frame diagram</center>\n",
"\n",
"\n",
"\n",
"LOMO [19] proposes iterative optimization of text localization features to obtain finer text localization for long text and curved text problems. The method consists of three parts: the coordinate regression module DR, the iterative optimization module IRM and the arbitrary shape expression module SEM. They are used to generate approximate text regions, iteratively optimize text localization features, and predict text regions, text centerlines, and text boundaries. Iteratively optimized text features can better solve the problem of long text localization and obtain more accurate text area localization.\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/e90adf3ca25a45a0af0b84a181fbe2c4954be1fcca8f4049957128548b7131ef\"\n",
"width=\"1000\" ></center>\n",
"<br><center>Figure 11: LOMO frame diagram</center>\n",
"\n",
"\n",
"Contournet [18] is based on the proposed modeling of text contour points to obtain a curved text detection frame. This method first uses Adaptive-RPN to obtain the proposal features of the text area, and then designs a local orthogonal texture perception LOTM module to learn horizontal and vertical textures. The feature is represented by contour points. Finally, by considering the feature responses in two orthogonal directions at the same time, the Point Re-Scoring algorithm can effectively filter out the prediction of strong unidirectional or weak orthogonal activation, and the final text contour can be used as a A group of high-quality contour points are shown.\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/1f59ab5db899412f8c70ba71e8dd31d4ea9480d6511f498ea492c97dd2152384\"\n",
"width=\"600\" ></center>\n",
"<br><center>Figure 12: Contournet frame diagram</center>\n",
"\n",
"\n",
"PCR [14] proposed progressive coordinate regression to deal with curved text detection. The problem is divided into three stages. Firstly, the text area is roughly detected, and the text box is obtained. In addition, the corners of the smallest bounding box of the text are predicted by the designed Contour Localization Mechanism. Coordinates, and then the curved text is predicted by superimposing multiple CLM modules and RCLM modules. This method uses the text contour information aggregation to obtain a rich text contour feature representation, which can not only suppress the influence of redundant noise points on the coordinate regression, but also locate the text area more accurately.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/c677c4602cee44999ae4b38bd780b69795887f2ae10747968bb084db6209b6cc\"\n",
"width=\"600\" ></center>\n",
"<br><center>Figure 13: PCR frame diagram</center>"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"\n",
"### 2.2 Text Detection Based on Segmentation\n",
"\n",
"Although the regression-based method has achieved good results in text detection, it is often difficult to obtain a smooth text surrounding curve for solving curved text, and the model is more complex and does not have performance advantages. Therefore, researchers proposed a text segmentation method based on image segmentation. First, classify at the pixel level, determine whether each pixel belongs to a text target, obtain the probability map of the text area, and obtain the enclosing curve of the text segmentation area through post-processing. .\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/fb9e50c410984c339481869ba11c1f39f80a4d74920b44b084601f2f8a23099f\"\n",
"width=\"600\" ></center>\n",
"<br><center>Figure 14: Schematic diagram of text segmentation algorithm</center>\n",
"\n",
"\n",
"Such methods are usually based on segmentation to achieve text detection, and segmentation-based methods have natural advantages for text detection with irregular shapes. The main idea of ​​the segmentation-based text detection method is to obtain the text area in the image through the segmentation method, and then use opencv, polygon and other post-processing to obtain the minimum enclosing curve of the text area.\n",
"\n",
"\n",
"Pixellink [7] uses a segmentation method to solve the text detection problem. The segmentation object is a text area. The pixels in the same text line (word) are linked together to segment the text, and the text bounding box is directly extracted from the segmentation result without a position. Regression can achieve the effect of text detection based on regression. However, there is a problem with the segmentation-based method. For texts with similar positions, the text segmentation area is prone to \"sticky\" problems. Wu, Yue et al. [8] proposed to separate the text while learning the boundary position of the text to better distinguish the text area. In addition, Tian et al. [9] proposed to map the pixels of the same text to the mapping space. In the mapping space, the distance of the mapping vector of the unified text is close, and the distance of the mapping vector of different texts becomes longer.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/462b5e1472824452a2c530939cda5e59ada226b2d0b745d19dd56068753a7f97\"\n",
"width=\"600\" ></center>\n",
"<br><center>Figure 15: PixelLink frame diagram</center>\n",
"\n",
"For the multi-scale problem of text detection, MSR [20] proposes to extract multiple scale features of the same image, then merge these features and up-sample to the original image size. The network finally predicts the text center area and each point of the text center area. The x-coordinate offset and y-coordinate offset of the nearest boundary point can finally get the contour coordinate set of the text area.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/9597efd68a224d60b74d7c51c99f7ff0ba9939e5cdb84fb79209b7e213f7d039\"\n",
"width=\"600\" ></center>\n",
"<br><center>Figure 16: MSR frame diagram</center>\n",
" \n",
"Aiming at the problem of segmentation-based text algorithms that are difficult to distinguish between adjacent texts, PSENet [10] proposed a progressive scale expansion network to learn text segmentation regions, predict text regions with different shrinkage ratios, and expand the detected text regions one by one. The essence of this method is The above is a variant of the boundary learning method, which can effectively solve the problem of detecting adjacent text of any shape.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/fa870b69a2a5423cad7422f64c32e0645dfc31a4ecc94a52832cf8742cded5ba\"\n",
"width=\"1000\" ></center>\n",
"<br><center>Figure 17: PSENet frame diagram</center>\n",
"\n",
"Assume that PSENet post-processing uses 3 kernels of different scales, as shown in the above figure s1, s2, and s3. First, start from the minimum kernel s1, calculate the connected domain of the text segmentation area, and get (b), and then expand the connected domain along the up, down, left, and right, and classify the pixels that belong to s2 but not to s1 in the expanded area. When encountering conflicts, the principle of \"first come first served\" is adopted, and the scale expansion operation is repeated, and finally independent segmented regions of different text lines can be obtained.\n",
"\n",
"\n",
"Seglink++ [17] proposed a characterization of the attraction and repulsion relationship between text block units for curved text and dense text, and then designed a minimum spanning tree algorithm to combine the units to obtain the final text detection box, and An instance-aware loss function is proposed so that the Seglink++ method can be trained end-to-end.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/1a16568361c0468db537ac25882eed096bca83f9c1544a92aee5239890f9d8d9\"\n",
"width=\"1000\" ></center>\n",
"<br><center>Figure 18: Seglink++ frame diagram</center>\n",
"\n",
"Although the segmentation method solves the problem of curved text detection, complex post-processing logic and prediction speed are also goals that need to be optimized.\n",
"\n",
"PAN [11] aims at the problem of slow text detection and prediction speed, and improves the performance of the algorithm from the aspects of network design and post-processing. First, PAN uses the lightweight ResNet18 as the Backbone, and also designs the lightweight feature enhancement module FPEM and feature fusion module FFM to enhance the features extracted by the Backbone. In terms of post-processing, a pixel clustering method is used to merge pixels whose distance from the kernel is less than the threshold d along the predicted text center (kernel). PAN guarantees high accuracy while having faster prediction speed.\n",
"\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/a76771f91db246ee8be062f96fa2a8abc7598dd87e6d4755b63fac71a4ebc170\"\n",
"width=\"1000\" ></center>\n",
"<br><center>Figure 19: PAN frame diagram</center>\n",
"\n",
"DBNet [12] aimed at the problem of time-consuming post-processing that requires the use of thresholds for binarization based on segmentation methods. It proposed a learnable threshold and cleverly designed a binarization function that approximates the step function to make the segmentation The network can learn the threshold of text segmentation end-to-end during training. The automatic adjustment of the threshold not only improves accuracy, but also simplifies post-processing and improves the performance of text detection.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/0d6423e3c79448f8b09090cf2dcf9d0c7baa0f6856c645808502678ae88d2917\"\n",
"width=\"1000\" ></center>\n",
"<br><center>Figure 20: DB frame diagram</center>\n",
"\n",
"FCENet [16] proposed to express the text enclosing curve with Fourier transform parameters. Since the Fourier coefficient representation can theoretically fit any closed curve, by designing a suitable model to predict an arbitrary shape text enclosing box based on Fourier transform In this way, the detection accuracy of highly curved text instances in natural scene text detection is improved.\n",
"\n",
"<center><img src=\"https://ai-studio-static-online.cdn.bcebos.com/45e9a374d97145689a961977f896c8f9f470a66655234c1498e1c8477e277954\"\n",
"width=\"1000\" ></center>\n",
"<br><center>Figure 21: FCENet frame diagram</center>\n",
"\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 3 Summary\n",
"\n",
"This section introduces the development of the field of text detection in recent years, including text detection methods based on regression and segmentation, and respectively enumerates and introduces the method ideas of some classic papers. The next section takes the PaddleOCR open source library as an example to introduce in detail the algorithm principles and core code implementation of DBNet."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Reference\n",
"1. Liao, Minghui, et al. \"Textboxes: A fast text detector with a single deep neural network.\" Thirty-first AAAI conference on artificial intelligence. 2017.\n",
"2. Liao, Minghui, Baoguang Shi, and Xiang Bai. \"Textboxes++: A single-shot oriented scene text detector.\" IEEE transactions on image processing 27.8 (2018): 3676-3690.\n",
"3. Tian, Zhi, et al. \"Detecting text in natural image with connectionist text proposal network.\" European conference on computer vision. Springer, Cham, 2016.\n",
"4. Zhou, Xinyu, et al. \"East: an efficient and accurate scene text detector.\" Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. 2017.\n",
"5. Wang, Fangfang, et al. \"Geometry-aware scene text detection with instance transformation network.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.\n",
"6. Yuliang, Liu, et al. \"Detecting curve text in the wild: New dataset and new solution.\" arXiv preprint arXiv:1712.02170 (2017).\n",
"7. Deng, Dan, et al. \"Pixellink: Detecting scene text via instance segmentation.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 32. No. 1. 2018.\n",
"8. Wu, Yue, and Prem Natarajan. \"Self-organized text detection with minimal post-processing via border learning.\" Proceedings of the IEEE International Conference on Computer Vision. 2017.\n",
"9. Tian, Zhuotao, et al. \"Learning shape-aware embedding for scene text detection.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019.\n",
"10. Wang, Wenhai, et al. \"Shape robust text detection with progressive scale expansion network.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019.\n",
"11. Wang, Wenhai, et al. \"Efficient and accurate arbitrary-shaped text detection with pixel aggregation network.\" Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019.\n",
"12. Liao, Minghui, et al. \"Real-time scene text detection with differentiable binarization.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 07. 2020.\n",
"13. Hochreiter, Sepp, and Jürgen Schmidhuber. \"Long short-term memory.\" Neural computation 9.8 (1997): 1735-1780.\n",
"14. Dai, Pengwen, et al. \"Progressive Contour Regression for Arbitrary-Shape Scene Text Detection.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.\n",
"15. He, Minghang, et al. \"MOST: A Multi-Oriented Scene Text Detector with Localization Refinement.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.\n",
"16. Zhu, Yiqin, et al. \"Fourier contour embedding for arbitrary-shaped text detection.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.\n",
"17. Tang, Jun, et al. \"Seglink++: Detecting dense and arbitrary-shaped scene text by instance-aware component grouping.\" Pattern recognition 96 (2019): 106954.\n",
"18. Wang, Yuxin, et al. \"Contournet: Taking a further step toward accurate arbitrary-shaped scene text detection.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.\n",
"19. Zhang, Chengquan, et al. \"Look more than once: An accurate detector for text of arbitrary shapes.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019.\n",
"20. Xue C, Lu S, Zhang W. Msr: Multi-scale shape regression for scene text detection[J]. arXiv preprint arXiv:1901.02596, 2019. \n"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "py35-paddle1.2.0"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.4"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
notebook/notebook_en/3.text_recognition/text_recognition_practice.ipynb 0 → 100644
浏览文件 @ 4b214948
因为 它太大了无法显示 source diff 。你可以改为 查看blob。
notebook/notebook_en/3.text_recognition/text_recognition_theory.ipynb 0 → 100644
浏览文件 @ 4b214948
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"\n",
"# Text Recognition Algorithm Theory\n",
"\n",
"This chapter mainly introduces the theoretical knowledge of text recognition algorithms, including background introduction, algorithm classification and some classic paper ideas.\n",
"\n",
"Through the study of this chapter, you can master:\n",
"\n",
"1. The goal of text recognition\n",
"\n",
"2. Classification of text recognition algorithms\n",
"\n",
"3. Typical ideas of various algorithms\n",
"\n",
"\n",
"## 1 Background Introduction\n",
"\n",
"Text recognition is a subtask of OCR (Optical Character Recognition), and its task is to recognize the text content of a fixed area. In the two-stage method of OCR, it is followed by text detection and converts image information into text information.\n",
"\n",
"Specifically, the model inputs a positioned text line, and the model predicts the text content and confidence level in the picture. The visualization results are shown in the following figure:\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/a7c3404f778b489db9c1f686c7d2ff4d63b67c429b454f98b91ade7b89f8e903 width=\"600\"></center>\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/e72b1d6f80c342ac951d092bc8c325149cebb3763ec849ec8a2f54e7c8ad60ca width=\"600\"></center>\n",
"<br><center>Figure 1: Visualization results of model predicttion</center>\n",
"\n",
"There are many application scenarios for text recognition, including document recognition, road sign recognition, license plate recognition, industrial number recognition, etc. According to actual scenarios, text recognition tasks can be divided into two categories: **Regular text recognition** and **Irregular Text recognition**.\n",
"\n",
"* Regular text recognition: mainly refers to printed fonts, scanned text, etc., and the text is considered to be roughly in the horizontal position\n",
"\n",
"* Irregular text recognition: It often appears in natural scenes, and due to the huge differences in text curvature, direction, deformation, etc., the text is often not in the horizontal position, and there are problems such as bending, occlusion, and blurring.\n",
"\n",
"\n",
"The figure below shows the data patterns of IC15 and IC13, which represent irregular text and regular text respectively. It can be seen that irregular text often has problems such as distortion, blurring, and large font differences. It is closer to the real scene and is also more challenging.\n",
"\n",
"Therefore, the current major algorithms are trying to obtain higher indicators on irregular data sets.\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/bae4fce1370b4751a3779542323d0765a02a44eace7b44d2a87a241c13c6f8cf width=\"400\">\n",
"<br><center>Figure 2: IC15 picture sample (irregular text)</center>\n",
"<img src=https://ai-studio-static-online.cdn.bcebos.com/b55800d3276f4f5fad170ea1b567eb770177fce226f945fba5d3247a48c15c34 width=\"400\"></center>\n",
"<br><center>Figure 3: IC13 picture sample (rule text)</center>\n",
"\n",
"\n",
"When comparing the capabilities of different recognition algorithms, they are often compared on these two types of public data sets. Comparing the effects on multiple dimensions, currently the more common English benchmark data sets are classified as follows:\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/4d0aada261064031a16816b39a37f2ff6af70dbb57004cb7a106ae6485f14684 width=\"600\"></center>\n",
"<br><center>Figure 4: Common English benchmark data sets</center>\n",
"\n",
"## 2 Text Recognition Algorithm Classification\n",
"\n",
"In the traditional text recognition method, the task is divided into 3 steps, namely image preprocessing, character segmentation and character recognition. It is necessary to model a specific scene, and it will become invalid once the scene changes. In the face of complex text backgrounds and scene changes, methods based on deep learning have better performance.\n",
"\n",
"Most existing recognition algorithms can be represented by the following unified framework, and the algorithm flow is divided into 4 stages:\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/a2750f4170864f69a3af36fc13db7b606d851f2f467d43cea6fbf3521e65450f)\n",
"\n",
"\n",
"We have sorted out the mainstream algorithm categories and main papers, refer to the following table:\n",
"<center>\n",
" \n",
"| Algorithm category | Main ideas | Main papers |\n",
"| -------- | --------------- | -------- |\n",
"| Traditional algorithm | Sliding window, character extraction, dynamic programming |-|\n",
"| ctc | Based on ctc method, sequence is not aligned, faster recognition | CRNN, Rosetta |\n",
"| Attention | Attention-based method, applied to unconventional text | RARE, DAN, PREN |\n",
"| Transformer | Transformer-based method | SRN, NRTR, Master, ABINet |\n",
"| Correction | The correction module learns the text boundary and corrects it to the horizontal direction | RARE, ASTER, SAR |\n",
"| Segmentation | Based on the method of segmentation, extract the character position and then do classification | Text Scanner, Mask TextSpotter |\n",
" \n",
"</center>\n",
"\n",
"\n",
"### 2.1 Regular Text Recognition\n",
"\n",
"\n",
"There are two mainstream algorithms for text recognition, namely the CTC (Conectionist Temporal Classification)-based algorithm and the Sequence2Sequence algorithm. The difference is mainly in the decoding stage.\n",
"\n",
"The CTC-based algorithm connects the encoded sequence to the CTC for decoding; the Sequence2Sequence-based method connects the sequence to the Recurrent Neural Network (RNN) module for cyclic decoding. Both methods have been verified to be effective and mainstream. Two major practices.\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/f64eee66e4a6426f934c1befc3b138629324cf7360c74f72bd6cf3c0de9d49bd width=\"600\"></center>\n",
"<br><center>Figure 5: Left: CTC-based method, right: Sequece2Sequence-based method </center>\n",
"\n",
"\n",
"#### 2.1.1 Algorithm Based on CTC\n",
"\n",
"The most typical algorithm based on CTC is CRNN (Convolutional Recurrent Neural Network) [1], and its feature extraction part uses mainstream convolutional structures, commonly used ResNet, MobileNet, VGG, etc. Due to the particularity of text recognition tasks, there is a large amount of contextual information in the input data. The convolution kernel characteristics of convolutional neural networks make it more focused on local information and lack long-dependent modeling capabilities, so it is difficult to use only convolutional networks. Dig into the contextual connections between texts. In order to solve this problem, the CRNN text recognition algorithm introduces the bidirectional LSTM (Long Short-Term Memory) to enhance the context modeling. Experiments prove that the bidirectional LSTM module can effectively extract the context information in the picture. Finally, the output feature sequence is input to the CTC module, and the sequence result is directly decoded. This structure has been verified to be effective and widely used in text recognition tasks. Rosetta [2] is a recognition network proposed by FaceBook, which consists of a fully convolutional model and CTC. Gao Y [3] et al. used CNN convolution instead of LSTM, with fewer parameters, and the performance improvement accuracy was the same.\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/d3c96dd9e9794fddb12fa16f926abdd3485194f0a2b749e792e436037490899b width=\"600\"></center>\n",
"<center>Figure 6: CRNN structure diagram </center>\n",
"\n",
"\n",
"#### 2.1.2 Sequence2Sequence algorithm\n",
"\n",
"In the Sequence2Sequence algorithm, the Encoder encodes all input sequences into a unified semantic vector, which is then decoded by the Decoder. In the decoding process of the decoder, the output of the previous moment is continuously used as the input of the next moment, and the decoding is performed in a loop until the stop character is output. The general encoder is an RNN. For each input word, the encoder outputs a vector and hidden state, and uses the hidden state for the next input word to get the semantic vector in a loop; the decoder is another RNN, which receives the encoder Output a vector and output a series of words to create a transformation. Inspired by Sequence2Sequence in the field of translation, Shi [4] proposed an attention-based codec framework to recognize text. In this way, rnn can learn character-level language models hidden in strings from training data.\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/f575333696b7438d919975dc218e61ccda1305b638c5497f92b46a7ec3b85243 width=\"400\" hight=\"500\"></center>\n",
"<center>Figure 7: Sequence2Sequence structure diagram </center>\n",
"\n",
"The above two algorithms have very good effects on regular text, but due to the limitations of network design, this type of method is difficult to solve the task of irregular text recognition of bending and rotation. In order to solve such problems, some algorithm researchers have proposed a series of improved algorithms on the basis of the above two types of algorithms.\n",
"\n",
"### 2.2 Irregular Text Recognition\n",
"\n",
"* Irregular text recognition algorithms can be divided into 4 categories: correction-based methods; Attention-based methods; segmentation-based methods; and Transformer-based methods.\n",
"\n",
"#### 2.2.1 Correction-based Method\n",
"\n",
"The correction-based method uses some visual transformation modules to convert irregular text into regular text as much as possible, and then uses conventional methods for recognition.\n",
"\n",
"The RARE [4] model first proposed a correction scheme for irregular text. The entire network is divided into two main parts: a spatial transformation network STN (Spatial Transformer Network) and a recognition network based on Sequence2Squence. Among them, STN is the correction module. Irregular text images enter STN and are transformed into a horizontal image through TPS (Thin-Plate-Spline). This transformation can correct curved and transmissive text to a certain extent, and send it to sequence recognition after correction. Network for decoding.\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/66406f89507245e8a57969b9bed26bfe0227a8cf17a84873902dd4a464b97bb5 width=\"600\"></center>\n",
"<center>Figure 8: RARE structure diagram </center>\n",
"\n",
"The RARE paper pointed out that this method has greater advantages in irregular text data sets, especially comparing the two data sets CUTE80 and SVTP, which are more than 5 percentage points higher than CRNN, which proves the effectiveness of the correction module. Based on this [6] also combines a text recognition system with a spatial transformation network (STN) and an attention-based sequence recognition network.\n",
"\n",
"Correction-based methods have better migration. In addition to Attention-based methods such as RARE, STAR-Net [5] applies correction modules to CTC-based algorithms, which is also a good improvement compared to traditional CRNN.\n",
"\n",
"#### 2.2.2 Attention-based Method\n",
"\n",
"The Attention-based method mainly focuses on the correlation between the parts of the sequence. This method was first proposed in the field of machine translation. It is believed that the result of the current word in the process of text translation is mainly affected by certain words, so it needs to be The decisive word has greater weight. The same is true in the field of text recognition. When decoding the encoded sequence, each step selects the appropriate context to generate the next state, which is conducive to obtaining more accurate results.\n",
"\n",
"R^2AM [7] first introduced Attention into the field of text recognition. The model first extracts the encoded image features from the input image through a recursive convolutional layer, and then uses the implicitly learned character-level language statistics to decode the output through a recurrent neural network character. In the decoding process, the Attention mechanism is introduced to realize soft feature selection to make better use of image features. This selective processing method is more in line with human intuition.\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/a64ef10d4082422c8ac81dcda4ab75bf1db285d6b5fd462a8f309240445654d5 width=\"600\"></center>\n",
"<center>Figure 9: R^2AM structure drawing </center>\n",
"\n",
"A large number of algorithms will be explored and updated in the field of Attention in the future. For example, SAR[8] extends 1D attention to 2D attention. The RARE mentioned in the correction module is also a method based on Attention. Experiments prove that the Attention-based method has a good accuracy improvement compared with the CTC method.\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/4e2507fb58d94ec7a9b4d17151a986c84c5053114e05440cb1e7df423d32cb02 width=\"600\"></center>\n",
"<center>Figure 10: Attention diagram</center>\n",
"\n",
"\n",
"#### 2.2.3 Method Based on Segmentation\n",
"\n",
"The method based on segmentation is to treat each character of the text line as an independent individual, and it is easier to recognize the segmented individual characters than to recognize the entire text line after correction. It attempts to locate the position of each character in the input text image, and applies a character classifier to obtain these recognition results, simplifying the complex global problem into a local problem solving, and it has a relatively good effect in the irregular text scene. However, this method requires character-level labeling, and there is a certain degree of difficulty in data acquisition. Lyu [9] et al. proposed an instance word segmentation model for word recognition, which uses a method based on FCN (Fully Convolutional Network) in its recognition part. [10] Considering the problem of text recognition from a two-dimensional perspective, a character attention FCN is designed to solve the problem of text recognition. When the text is bent or severely distorted, this method has better positioning results for both regular and irregular text.\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/fd3e8ef0d6ce4249b01c072de31297ca5d02fc84649846388f890163b624ff10 width=\"800\"></center>\n",
"<center>Figure 11: Mask TextSpotter structure diagram </center>\n",
"\n",
"\n",
"\n",
"#### 2.2.4 Transformer-based Method\n",
"\n",
"With the rapid development of Transformer, both classification and detection fields have verified the effectiveness of Transformer in visual tasks. As mentioned in the regular text recognition part, CNN has limitations in long-dependency modeling. The Transformer structure just solves this problem. It can focus on global information in the feature extractor and can replace additional context modeling modules (LSTM ).\n",
"\n",
"Part of the text recognition algorithm uses Transformer's Encoder structure and convolution to extract sequence features. The Encoder is composed of multiple blocks stacked by MultiHeadAttentionLayer and Positionwise Feedforward Layer. The self-attention in MulitHeadAttention uses matrix multiplication to simulate the timing calculation of RNN, breaking the barrier of long-term dependence on timing in RNN. There are also some algorithms that use Transformer's Decoder module to decode, which can obtain stronger semantic information than traditional RNNs, and parallel computing has higher efficiency.\n",
"\n",
"The SRN[11] algorithm connects the Encoder module of Transformer to ResNet50 to enhance the 2D visual features. A parallel attention module is proposed, which uses the reading order as a query, making the calculation independent of time, and finally outputs the aligned visual features of all time steps in parallel. In addition, SRN also uses Transformer's Eecoder as a semantic module to integrate the visual information and semantic information of the picture, which has greater benefits in irregular text such as occlusion and blur.\n",
"\n",
"NRTR [12] uses a complete Transformer structure to encode and decode the input picture, and only uses a few simple convolutional layers for high-level feature extraction, and verifies the effectiveness of the Transformer structure in text recognition.\n",
"\n",
"<center><img src=https://ai-studio-static-online.cdn.bcebos.com/e7859f4469a842f0bd450e7e793a679d6e828007544241d09785c9b4ea2424a2 width=\"800\"></center>\n",
"<center>Figure 12: NRTR structure drawing </center>\n",
"\n",
"SRACN [13] uses Transformer's decoder to replace LSTM, once again verifying the efficiency and accuracy advantages of parallel training.\n",
"\n",
"## 3 Summary\n",
"\n",
"This section mainly introduces the theoretical knowledge and mainstream algorithms related to text recognition, including CTC-based methods, Sequence2Sequence-based methods, and segmentation-based methods. The ideas and contributions of classic papers are listed respectively. The next section will explain the practical course based on the CRNN algorithm, from networking to optimization to complete the entire training process,\n",
"\n",
"## 4 Reference\n",
"\n",
"\n",
"[1]Shi, B., Bai, X., & Yao, C. (2016). An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE transactions on pattern analysis and machine intelligence, 39(11), 2298-2304.\n",
"\n",
"[2]Fedor Borisyuk, Albert Gordo, and Viswanath Sivakumar. Rosetta: Large scale system for text detection and recognition in images. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 71–79. ACM, 2018.\n",
"\n",
"[3]Gao, Y., Chen, Y., Wang, J., & Lu, H. (2017). Reading scene text with attention convolutional sequence modeling. arXiv preprint arXiv:1709.04303.\n",
"\n",
"[4]Shi, B., Wang, X., Lyu, P., Yao, C., & Bai, X. (2016). Robust scene text recognition with automatic rectification. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4168-4176).\n",
"\n",
"[5] Star-Net Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spa- tial transformer networks. In Advances in neural information processing systems, pages 2017–2025, 2015.\n",
"\n",
"[6]Baoguang Shi, Mingkun Yang, XingGang Wang, Pengyuan Lyu, Xiang Bai, and Cong Yao. Aster: An attentional scene text recognizer with flexible rectification. IEEE transactions on pattern analysis and machine intelligence, 31(11):855–868, 2018.\n",
"\n",
"[7] Lee C Y , Osindero S . Recursive Recurrent Nets with Attention Modeling for OCR in the Wild[C]// IEEE Conference on Computer Vision & Pattern Recognition. IEEE, 2016.\n",
"\n",
"[8]Li, H., Wang, P., Shen, C., & Zhang, G. (2019, July). Show, attend and read: A simple and strong baseline for irregular text recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 33, No. 01, pp. 8610-8617).\n",
"\n",
"[9]P. Lyu, C. Yao, W. Wu, S. Yan, and X. Bai. Multi-oriented scene text detection via corner localization and region segmentation. In Proc. CVPR, pages 7553–7563, 2018.\n",
"\n",
"[10] Liao, M., Zhang, J., Wan, Z., Xie, F., Liang, J., Lyu, P., ... & Bai, X. (2019, July). Scene text recognition from two-dimensional perspective. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 33, No. 01, pp. 8714-8721).\n",
"\n",
"[11] Yu, D., Li, X., Zhang, C., Liu, T., Han, J., Liu, J., & Ding, E. (2020). Towards accurate scene text recognition with semantic reasoning networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 12113-12122).\n",
"\n",
"[12] Sheng, F., Chen, Z., & Xu, B. (2019, September). NRTR: A no-recurrence sequence-to-sequence model for scene text recognition. In 2019 International Conference on Document Analysis and Recognition (ICDAR) (pp. 781-786). IEEE.\n",
"\n",
"[13]Yang, L., Wang, P., Li, H., Li, Z., & Zhang, Y. (2020). A holistic representation guided attention network for scene text recognition. Neurocomputing, 414, 67-75.\n",
"\n",
"[14]Wang, T., Zhu, Y., Jin, L., Luo, C., Chen, X., Wu, Y., ... & Cai, M. (2020, April). Decoupled attention network for text recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 34, No. 07, pp. 12216-12224).\n",
"\n",
"[15] Wang, Y., Xie, H., Fang, S., Wang, J., Zhu, S., & Zhang, Y. (2021). From two to one: A new scene text recognizer with visual language modeling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 14194-14203).\n",
"\n",
"[16] Fang, S., Xie, H., Wang, Y., Mao, Z., & Zhang, Y. (2021). Read Like Humans: Autonomous, Bidirectional and Iterative Language Modeling for Scene Text Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 7098-7107).\n",
"\n",
"[17] Yan, R., Peng, L., Xiao, S., & Yao, G. (2021). Primitive Representation Learning for Scene Text Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 284-293)."
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "py35-paddle1.2.0"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.4"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
notebook/notebook_en/4.ppocr_system_strategy/ppocr_system_strategy.ipynb 0 → 100644
浏览文件 @ 4b214948
因为 它太大了无法显示 source diff 。你可以改为 查看blob。
notebook/notebook_en/6.document_analysis/document_analysis_practice-VQA.ipynb 0 → 100644
浏览文件 @ 4b214948
因为 它太大了无法显示 source diff 。你可以改为 查看blob。
notebook/notebook_en/6.document_analysis/document_analysis_theory.ipynb 0 → 100644
浏览文件 @ 4b214948
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Document Analysis Technology\n",
"\n",
"This chapter mainly introduces the theoretical knowledge of document analysis technology, including background introduction, algorithm classification and corresponding ideas.\n",
"\n",
"Through the study of this chapter, you can master:\n",
"\n",
"1. Classification and typical ideas of layout analysis\n",
"2. Classification and typical ideas of table recognition\n",
"3. Classification and typical ideas of information extraction"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"\n",
"Layout analysis is mainly used for document retrieval, key information extraction, content classification, etc. Its task is mainly to classify the content of document images. Content categories can generally be divided into plain text, titles, tables, pictures, and lists. However, the diversity and complexity of document layout, formats, poor document image quality, and the lack of large-scale annotated datasets make layout analysis still a challenging task. Document analysis often includes the following research directions:\n",
"\n",
"1. Layout analysis module: Divide each document page into different content areas. This module can be used not only to delimit relevant and irrelevant areas, but also to classify the types of content it recognizes.\n",
"2. Optical Character Recognition (OCR) module: Locate and recognize all text present in the document.\n",
"3. Form recognition module: Recognize and convert the form information in the document into an excel file.\n",
"4. Information extraction module: Use OCR results and image information to understand and identify the specific information expressed in the document or the relationship between the information.\n",
"\n",
"Since the OCR module has been introduced in detail in the previous chapters, the following three modules will be introduced separately for the above layout analysis, table recognition and information extraction. For each module, the classic or common methods and data sets of the module will be introduced."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 1. Layout Analysis\n",
"\n",
"### 1.1 Background Introduction\n",
"\n",
"Layout analysis is mainly used for document retrieval, key information extraction, content classification, etc. Its task is mainly to classify document images. Content categories can generally be divided into plain text, titles, tables, pictures, and lists. However, the diversity and complexity of document layouts, formats, poor document image quality, and the lack of large-scale annotated data sets make layout analysis still a challenging task.\n",
"The visualization of the layout analysis task is shown in the figure below:\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/2510dc76c66c49b8af079f25d08a9dcba726b2ce53d14c8ba5cd9bd57acecf19\" width=\"1000\"/></center>\n",
"<center>Figure 1: Layout analysis effect diagram</center>\n",
"\n",
"The existing solutions are generally based on target detection or semantic segmentation methods, which are based on detecting or segmenting different patterns in the document as different targets.\n",
"\n",
"Some representative papers are divided into the above two categories, as shown in the following table:\n",
"\n",
"| Category | Main paper |\n",
"| ---------------- | -------- |\n",
"| Method based on target detection | [Visual Detection with Context](https://aclanthology.org/D19-1348.pdf)，[Object Detection](https://arxiv.org/pdf/2003.13197v1.pdf)，[VSR](https://arxiv.org/pdf/2105.06220v1.pdf)\n",
"| Semantic segmentation method |[Semantic Segmentation](https://arxiv.org/pdf/1911.12170v2.pdf) |\n",
"\n",
"\n",
"### 1.2 Method Based on Target Detection \n",
"\n",
"Soto Carlos [1] is based on the target detection algorithm Faster R-CNN, combines context information and uses the inherent location information of the document content to improve the area detection performance. Li Kai [2] et al. also proposed a document analysis method based on object detection, which solved the cross-domain problem by introducing a feature pyramid alignment module, a region alignment module, and a rendering layer alignment module. These three modules complement each other. And adjust the domain from a general image perspective and a specific document image perspective, thus solving the problem of large labeled training datasets being different from the target domain. The following figure is a flow chart of layout analysis based on the target detection Faster R-CNN algorithm. \n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/d396e0d6183243898c0961250ee7a49bc536677079fb4ba2ac87c653f5472f01\" width=\"800\"/></center>\n",
"<center>Figure 2: Flow chart of layout analysis based on Faster R-CNN</center>\n",
"\n",
"### 1.3 Methods Based on Semantic Segmentation \n",
"\n",
"Sarkar Mausoom [3] et al. proposed a priori-based segmentation mechanism to train a document segmentation model on very high-resolution images, which solves the problem of indistinguishable and merging of dense regions caused by excessively shrinking the original image. Zhang Peng [4] et al. proposed a unified framework VSR (Vision, Semantics and Relations) for document layout analysis in combination with the vision, semantics and relations in the document. The framework uses a two-stream network to extract the visual and Semantic features, and adaptively fusion of these features through the adaptive aggregation module, solves the limitations of the existing CV-based methods of low efficiency of different modal fusion and lack of relationship modeling between layout components.\n",
"\n",
"### 1.4 Data Set\n",
"\n",
"Although the existing methods can solve the layout analysis task to a certain extent, these methods rely on a large amount of labeled training data. Recently, many data sets have been proposed for document analysis tasks.\n",
"\n",
"1. PubLayNet[5]: The data set contains 500,000 document images, of which 400,000 are used for training, 50,000 are used for verification, and 50,000 are used for testing. Five forms of table, text, image, title and list are marked.\n",
"2. HJDataset[6]: The data set contains 2271 document images. In addition to the bounding box and mask of the content area, it also includes the hierarchical structure and reading order of layout elements.\n",
"\n",
"A sample of the PubLayNet data set is shown in the figure below:\n",
"<center class=\"two\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/4b153117c9384f98a0ce5a6c6e7c205a4b1c57e95c894ccb9688cbfc94e68a1c\" width=\"400\"/><img src=\"https://ai-studio-static-online.cdn.bcebos.com/efb9faea39554760b280f9e0e70631d2915399fa97774eecaa44ee84411c4994\" width=\"400\"/>\n",
"</center>\n",
"<center>Figure 3: PubLayNet example</center>\n",
"Reference:\n",
"\n",
"[1]：Soto C, Yoo S. Visual detection with context for document layout analysis[C]//Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019: 3464-3470.\n",
"\n",
"[2]：Li K, Wigington C, Tensmeyer C, et al. Cross-domain document object detection: Benchmark suite and method[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 12915-12924.\n",
"\n",
"[3]：Sarkar M, Aggarwal M, Jain A, et al. Document Structure Extraction using Prior based High Resolution Hierarchical Semantic Segmentation[C]//European Conference on Computer Vision. Springer, Cham, 2020: 649-666.\n",
"\n",
"[4]：Zhang P, Li C, Qiao L, et al. VSR: A Unified Framework for Document Layout Analysis combining Vision, Semantics and Relations[J]. arXiv preprint arXiv:2105.06220, 2021.\n",
"\n",
"[5]：Zhong X, Tang J, Yepes A J. Publaynet: largest dataset ever for document layout analysis[C]//2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, 2019: 1015-1022.\n",
"\n",
"[6]：Li M, Xu Y, Cui L, et al. DocBank: A benchmark dataset for document layout analysis[J]. arXiv preprint arXiv:2006.01038, 2020.\n",
"\n",
"[7]：Shen Z, Zhang K, Dell M. A large dataset of historical japanese documents with complex layouts[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2020: 548-549."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 2. Form Recognition\n",
"\n",
"### 2.1 Background Introduction\n",
"\n",
"Tables are common page elements in various types of documents. With the explosive growth of various types of documents, how to efficiently find tables from documents and obtain content and structure information, that is, table identification has become an urgent problem to be solved. The difficulties of form identification are summarized as follows:\n",
"\n",
"1. The types and styles of tables are complex and diverse, such as *different rows and columns combined, different content text types*, etc.\n",
"2. The style of the document itself has various styles.\n",
"3. The lighting environment during shooting, etc.\n",
"\n",
"The task of table recognition is to convert the table information in the document to an excel file. The task visualization is as follows:\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/99faa017e28b4928a408573406870ecaa251b626e0e84ab685e4b6f06f601a5f\" width=\"1600\"/></center>\n",
"\n",
"\n",
"<center>Figure 4: Example image of table recognition, the left side is the original image, and the right side is the result image after table recognition, presented in Excel format</center>\n",
"\n",
"Existing table recognition algorithms can be divided into the following four categories according to the principle of table structure reconstruction:\n",
"1. Method based on heuristic rules\n",
"2. CNN-based method\n",
"3. GCN-based method\n",
"4. Method based on End to End\n",
"\n",
"Some representative papers are divided into the above four categories, as shown in the following table:\n",
"\n",
"| Category | Idea | Main papers |\n",
"| ---------------- | ---- | -------- |\n",
"|Method based on heuristic rules|Manual design rules, connected domain detection analysis and processing|[T-Rect](https://www.researchgate.net/profile/Andreas-Dengel/publication/249657389_A_Paper-to-HTML_Table_Converting_System/links/0c9605322c9a67274d000000/A-Paper-to-HTML-Table-Converting-System.pdf)，[pdf2table](https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.724.7272&rep=rep1&type=pdf)|\n",
"| CNN-based method | target detection, semantic segmentation | [CascadeTabNet](https://arxiv.org/pdf/2004.12629v2.pdf), [Multi-Type-TD-TSR](https://arxiv.org/pdf/2105.11021v1.pdf), [LGPMA](https://arxiv.org/pdf/2105.06224v2.pdf), [tabstruct-net](https://arxiv.org/pdf/2010.04565v1.pdf), [CDeC-Net](https://arxiv.org/pdf/2008.10831v1.pdf), [TableNet](https://arxiv.org/pdf/2001.01469v1.pdf), [TableSense](https://arxiv.org/pdf/2106.13500v1.pdf), [Deepdesrt](https://www.dfki.de/fileadmin/user_upload/import/9672_PID4966073.pdf), [Deeptabstr](https://www.dfki.de/fileadmin/user_upload/import/10649_DeepTabStR.pdf), [GTE](https://arxiv.org/pdf/2005.00589v2.pdf), [Cycle-CenterNet](https://arxiv.org/pdf/2109.02199v1.pdf), [FCN](https://www.researchgate.net/publication/339027294_Rethinking_Semantic_Segmentation_for_Table_Structure_Recognition_in_Documents)|\n",
"| GCN-based method | Based on graph neural network, the table recognition is regarded as a graph reconstruction problem | [GNN](https://arxiv.org/pdf/1905.13391v2.pdf), [TGRNet](https://arxiv.org/pdf/2106.10598v3.pdf), [GraphTSR](https://arxiv.org/pdf/1908.04729v2.pdf)|\n",
"| Method based on End to End | Use attention mechanism | [Table-Master](https://arxiv.org/pdf/2105.01848v1.pdf)|\n",
"\n",
"### 2.2 Traditional Algorithm Based on Heuristic Rules\n",
"Early research on table recognition was mainly based on heuristic rules. For example, the T-Rect system proposed by Kieninger [1] et al. used a bottom-up method to analyze the connected domain of document images, and then merge them according to defined rules to obtain logical text blocks. Then, pdf2table proposed by Yildiz[2] et al. is the first method for table recognition on PDF documents, which utilizes some unique information of PDF files (such as text, drawing paths and other information that are difficult to obtain in image documents) to assist with table recognition. In recent work, Koci[3] et al. expressed the layout area in the page as a graph, and then used the Remove and Conquer (RAC) algorithm to identify the table as a subgraph.\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/66aeedb3f0924d80aee15f185e6799cc687b51fc20b74b98b338ca2ea25be3f3\" width=\"1000\"/></center>\n",
"<center>Figure 5: Schematic diagram of heuristic algorithm</center>\n",
"\n",
"### 2.3 Method Based on Deep Learning CNN\n",
"With the rapid development of deep learning technology in computer vision, natural language processing, speech processing and other fields, researchers have applied deep learning technology to the field of table recognition and achieved good results.\n",
"\n",
"In the DeepTabStR algorithm, Siddiqui Shoaib Ahmed [12] et al. described the table structure recognition problem as an object detection problem, and used deformable convolution to better detect table cells. Raja Sachin[6] et al. proposed that TabStruct-Net combines cell detection and structure recognition visually to perform table structure recognition, which solves the problem of recognition errors due to large changes in the table layout. However, this method cannot Deal with the problem of more empty cells in rows and columns.\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/838be28836444bc1835ac30a25613d8b045a1b5aedd44b258499fe9f93dd298f\" width=\"1600\"/></center>\n",
"<center>Figure 6: Schematic diagram of algorithm based on deep learning CNN</center>\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/4c40dda737bd44b09a533e1b1dd2e4c6a90ceea083bf4238b7f3c7b21087f409\" width=\"1600\"/></center>\n",
"<center>Figure 7: Example of algorithm error based on deep learning CNN</center>\n",
"\n",
"The previous table structure recognition method generally starts from the elements of different granularities (row/column, text area), and it is easy to ignore the problem of merging empty cells. Qiao Liang [10] et al. proposed a new framework LGPMA, which makes full use of the information from local and global features through mask re-scoring strategy, and then can obtain more reliable alignment of the cell area, and finally introduces the inclusion of cell matching, empty The table structure restoration pipeline of cell search and empty cell merging handles the problem of table structure identification.\n",
"\n",
"In addition to the above separate table recognition algorithms, there are also some methods that complete table detection and table recognition in one model. Schreiber Sebastian [11] et al. proposed DeepDeSRT, which uses Faster RCNN for table detection and FCN semantic segmentation model for Table structure row and column detection, but this method uses two independent models to solve these two problems. Prasad Devashish [4] et al. proposed an end-to-end deep learning method CascadeTabNet, which uses the Cascade Mask R-CNN HRNet model to perform table detection and structure recognition at the same time, which solves the problem of using two independent methods to process table recognition in the past. The insufficiency of the problem. Paliwal Shubham [8] et al. proposed a novel end-to-end deep multi-task architecture TableNet, which is used for table detection and structure recognition. At the same time, additional spatial semantic features are added to TableNet during training to further improve the performance of the model. Zheng Xinyi [13] et al. proposed a system framework GTE for table recognition, using a cell detection network to guide the training of the table detection network, and proposed a hierarchical network and a new clustering-based cell structure recognition algorithm, the framework can be connected to the back of any target detection model to facilitate the training of different table recognition algorithms. Previous research mainly focused on parsing from scanned PDF documents with a simple layout and well-aligned table images. However, the tables in real scenes are generally complex and may have serious deformation, bending or occlusion. Therefore, Long Rujiao [14] et al. also constructed a table recognition data set WTW in a realistic complex scene, and proposed a Cycle-CenterNet method, which uses the cyclic pairing module optimization and the proposed new pairing loss to accurately group discrete units into structured In the table, the performance of table recognition is improved.\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/a01f714cbe1f42fc9c45c6658317d9d7da2cec9726844f6b9fa75e30cadc9f76\" width=\"1600\"/></center>\n",
"<center>Figure 8: Schematic diagram of end-to-end algorithm</center>\n",
"\n",
"The CNN-based method cannot handle the cross-row and column tables well, so in the follow-up method, two research methods are divided to solve the cross-row and column problems in the table.\n",
"\n",
"### 2.4 Method Based on Deep Learning GCN\n",
"In recent years, with the rise of Graph Convolutional Networks (Graph Convolutional Network), some researchers have tried to apply graph neural networks to table structure recognition problems. Qasim Shah Rukh [20] et al. converted the table structure recognition problem into a graph problem compatible with graph neural networks, and designed a novel differentiable architecture that can not only take advantage of the advantages of convolutional neural networks to extract features, but also The advantages of the effective interaction between the vertices of the graph neural network can be used, but this method only uses the location features of the cells, and does not use the semantic features. Chi Zewen [19] et al. proposed a novel graph neural network, GraphTSR, for table structure recognition in PDF files. It takes cells in the table as input, and then uses the characteristics of the edges and nodes of the graph to be connected. Predicting the relationship between cells to identify the table structure solves the problem of cell identification across rows or columns to a certain extent. Xue Wenyuan [21] et al. reformulated the problem of table structure recognition as table map reconstruction, and proposed an end-to-end method for table structure recognition, TGRNet, which includes cell detection branch and cell logic location branch. , These two branches jointly predict the spatial and logical positions of different cells, which solves the problem that the previous method did not pay attention to the logical position of cells.\n",
"\n",
"Diagram of GraphTSR table recognition algorithm:\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/8ff89661142045a8aef54f8a7a2c69b1d243f8269034406a9e66bee2149f730f\" width=\"1600\"/></center>\n",
"<center>Figure 9: Diagram of GraphTSR table recognition algorithm</center>\n",
"\n",
"### 2.5 Based on An End-to-End Approach\n",
"\n",
"Different from other post-processing to complete the reconstruction of the table structure, based on the end-to-end method, directly use the network to complete the HTML representation output of the table structure\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/7865e58a83824facacfaa91bec12ccf834217cb706454dc5a0c165c203db79fb) | ![](https://ai-studio-static-online.cdn.bcebos.com/77d913b1b92f4a349b8f448e08ba78458d687eef4af142678a073830999f3edc))\n",
"---|---\n",
"Figure 10: Input and output of the end-to-end method | Figure 11: Image Caption example\n",
"\n",
"Most end-to-end methods use Image Caption's Seq2Seq method to complete the prediction of the table structure, such as some methods based on Attention or Transformer.\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/3571280a9c364d3499a062e3edc724294fb5eaef8b38440991941e87f0af0c3b\" width=\"800\"/></center>\n",
"<center>Figure 12: Schematic diagram of Seq2Seq</center>\n",
"\n",
"Ye Jiaquan [22] obtained the table structure output model in TableMaster by improving the Master text algorithm based on Transformer. In addition, a branch is added for the coordinate regression of the box. The author did not split the model into two branches in the last layer, but decoupled the sequence prediction and the box regression into two after the first Transformer decoding layer. Branches. The comparison between its network structure and the original Master network is shown in the figure below:\n",
"\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/f573709447a848b4ba7c73a2e297f0304caaca57c5c94588aada1f4cd893946c\" width=\"800\"/></center>\n",
"<center>Figure 13: Left: master network diagram, right: TableMaster network diagram</center>\n",
"\n",
"\n",
"### 2.6 Data Set\n",
"\n",
"Since the deep learning method is data-driven, a large amount of labeled data is required to train the model, and the small size of the existing data set is also an important constraint, so some data sets have also been proposed.\n",
"\n",
"1. PubTabNet[16]: Contains 568k table images and corresponding structured HTML representations.\n",
"2. PubMed Tables (PubTables-1M) [17]: Table structure recognition data set, containing highly detailed structural annotations, 460,589 pdf images used for table inspection tasks, and 947,642 table images used for table recognition tasks.\n",
"3. TableBank[18]: Table detection and recognition data set, using Word and Latex documents on the Internet to construct table data containing 417K high-quality annotations.\n",
"4. SciTSR[19]: Table structure recognition data set, most of the images are converted from the paper, which contains 15,000 tables from PDF files and their corresponding structure tags.\n",
"5. TabStructDB[12]: Includes 1081 table areas, which are marked with dense row and column information.\n",
"6. WTW[14]: Large-scale data set scene table detection and recognition data set, this data set contains table data in various deformation, bending and occlusion situations, and contains 14,581 images in total.\n",
"\n",
"Data set example\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/c9763df56e67434f97cd435100d50ded71ba66d9d4f04d7f8f896d613cdf02b0\" /></center>\n",
"<center>Figure 14: Sample diagram of PubTables-1M data set</center>\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/64de203bbe584642a74f844ac4b61d1ec3c5a38cacb84443ac961fbcc54a66ce\" width=\"600\"/></center>\n",
"<center>Figure 15: Sample diagram of WTW data set</center>\n",
"\n",
"\n",
"\n",
"Reference\n",
"\n",
"[1]：Kieninger T, Dengel A. A paper-to-HTML table converting system[C]//Proceedings of document analysis systems (DAS). 1998, 98: 356-365.\n",
"\n",
"[2]：Yildiz B, Kaiser K, Miksch S. pdf2table: A method to extract table information from pdf files[C]//IICAI. 2005: 1773-1785.\n",
"\n",
"[3]：Koci E, Thiele M, Lehner W, et al. Table recognition in spreadsheets via a graph representation[C]//2018 13th IAPR International Workshop on Document Analysis Systems (DAS). IEEE, 2018: 139-144.\n",
"\n",
"[4]：Prasad D, Gadpal A, Kapadni K, et al. CascadeTabNet: An approach for end to end table detection and structure recognition from image-based documents[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2020: 572-573.\n",
"\n",
"[5]：Fischer P, Smajic A, Abrami G, et al. Multi-Type-TD-TSR–Extracting Tables from Document Images Using a Multi-stage Pipeline for Table Detection and Table Structure Recognition: From OCR to Structured Table Representations[C]//German Conference on Artificial Intelligence (Künstliche Intelligenz). Springer, Cham, 2021: 95-108.\n",
"\n",
"[6]：Raja S, Mondal A, Jawahar C V. Table structure recognition using top-down and bottom-up cues[C]//European Conference on Computer Vision. Springer, Cham, 2020: 70-86.\n",
"\n",
"[7]：Agarwal M, Mondal A, Jawahar C V. Cdec-net: Composite deformable cascade network for table detection in document images[C]//2020 25th International Conference on Pattern Recognition (ICPR). IEEE, 2021: 9491-9498.\n",
"\n",
"[8]：Paliwal S S, Vishwanath D, Rahul R, et al. Tablenet: Deep learning model for end-to-end table detection and tabular data extraction from scanned document images[C]//2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, 2019: 128-133.\n",
"\n",
"[9]：Dong H, Liu S, Han S, et al. Tablesense: Spreadsheet table detection with convolutional neural networks[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2019, 33(01): 69-76.\n",
"\n",
"[10]：Qiao L, Li Z, Cheng Z, et al. LGPMA: Complicated Table Structure Recognition with Local and Global Pyramid Mask Alignment[J]. arXiv preprint arXiv:2105.06224, 2021.\n",
"\n",
"[11]：Schreiber S, Agne S, Wolf I, et al. Deepdesrt: Deep learning for detection and structure recognition of tables in document images[C]//2017 14th IAPR international conference on document analysis and recognition (ICDAR). IEEE, 2017, 1: 1162-1167.\n",
"\n",
"[12]：Siddiqui S A, Fateh I A, Rizvi S T R, et al. Deeptabstr: Deep learning based table structure recognition[C]//2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, 2019: 1403-1409.\n",
"\n",
"[13]：Zheng X, Burdick D, Popa L, et al. Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context[C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2021: 697-706.\n",
"\n",
"[14]：Long R, Wang W, Xue N, et al. Parsing Table Structures in the Wild[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021: 944-952.\n",
"\n",
"[15]：Siddiqui S A, Khan P I, Dengel A, et al. Rethinking semantic segmentation for table structure recognition in documents[C]//2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, 2019: 1397-1402.\n",
"\n",
"[16]：Zhong X, ShafieiBavani E, Jimeno Yepes A. Image-based table recognition: data, model, and evaluation[C]//Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16. Springer International Publishing, 2020: 564-580.\n",
"\n",
"[17]：Smock B, Pesala R, Abraham R. PubTables-1M: Towards a universal dataset and metrics for training and evaluating table extraction models[J]. arXiv preprint arXiv:2110.00061, 2021.\n",
"\n",
"[18]：Li M, Cui L, Huang S, et al. Tablebank: Table benchmark for image-based table detection and recognition[C]//Proceedings of the 12th Language Resources and Evaluation Conference. 2020: 1918-1925.\n",
"\n",
"[19]：Chi Z, Huang H, Xu H D, et al. Complicated table structure recognition[J]. arXiv preprint arXiv:1908.04729, 2019.\n",
"\n",
"[20]：Qasim S R, Mahmood H, Shafait F. Rethinking table recognition using graph neural networks[C]//2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, 2019: 142-147.\n",
"\n",
"[21]：Xue W, Yu B, Wang W, et al. TGRNet: A Table Graph Reconstruction Network for Table Structure Recognition[J]. arXiv preprint arXiv:2106.10598, 2021.\n",
"\n",
"[22]：Ye J, Qi X, He Y, et al. PingAn-VCGroup's Solution for ICDAR 2021 Competition on Scientific Literature Parsing Task B: Table Recognition to HTML[J]. arXiv preprint arXiv:2105.01848, 2021.\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 3. Document VQA\n",
"\n",
"The boss sent a task: develop an ID card recognition system\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/63bbe893465e4f98b3aec80a042758b520d43e1a993a47e39bce1123c2d29b3f\" width=\"1600\"/></center>\n",
"\n",
"> How to choose a plan\n",
"> 1. Use rules to extract information after text detection\n",
"> 2. Use scale type to extract information after text detection\n",
"> 3. Outsourcing\n",
"\n",
"\n",
"### 3.1 Background Introduction\n",
"In the VQA (Visual Question Answering) task, questions and answers are mainly aimed at the content of the image, but for text images, the content of interest is the text information in the image, so this type of method can be divided into Text-VQA and text-VQA in natural scenes. DocVQA of the scanned document scene, the relationship between the three is shown in the figure below.\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/a91cfd5152284152b020ca8a396db7a21fd982e3661540d5998cc19c17d84861\" width=\"600\"/></center>\n",
"<center>Figure 16: VQA level</center>\n",
"\n",
"The sample pictures of VQA, Text-VQA and DocVQA are shown in the figure below.\n",
"\n",
"|Task type|VQA | Text-VQA | DocVQA| \n",
"|---|---|---|---|\n",
"|Task description|Ask questions regarding **picture content**|Ask questions regarding **text content on pictures**|Ask questions regarding **text content of document images**|\n",
"|Sample picture|![vqa](https://ai-studio-static-online.cdn.bcebos.com/fc21b593276247249591231b3373608151ed8ae7787f4d6ba39e8779fdd12201)|![textvqa](https://ai-studio-static-online.cdn.bcebos.com/cd2404edf3bf430b89eb9b2509714499380cd02e4aa74ec39ca6d7aebcf9a559)|![docvqa](https://ai-studio-static-online.cdn.bcebos.com/0eec30a6f91b4f949c56729b856f7ff600d06abee0774642801c070303edfe83)|\n",
"\n",
"Because DocVQA is closer to actual application scenarios, a large number of academic and industrial work has emerged. In common scenarios, the questions asked in DocVQA are fixed. For example, the questions in the ID card scenario are generally\n",
"1. What is the citizenship number?\n",
"2. What is your name?\n",
"3. What is a clan?\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/2d2b86468daf47c98be01f44b8d6efa64bc09e43cd764298afb127f19b07aede\" width=\"800\"/></center>\n",
"<center>Figure 17: Example of an ID card</center>\n",
"\n",
"\n",
"Based on this prior knowledge, DocVQA's research began to lean towards the Key Information Extraction (KIE) task. This time we also mainly discuss the KIE-related research. The KIE task mainly extracts the key information needed from the image, such as extracting from the ID card. Name and citizen identification number information.\n",
"\n",
"KIE is usually divided into two sub-tasks for research\n",
"1. SER: Semantic Entity Recognition, to classify each detected text, such as dividing it into name and ID. As shown in the black box and red box in the figure below.\n",
"2. RE: Relation Extraction, which classifies each detected text, such as dividing it into questions and answers. Then find the corresponding answer to each question. As shown in the figure below, the red and black boxes represent the question and the answer, respectively, and the yellow line represents the correspondence between the question and the answer.\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/899470ba601349fbbc402a4c83e6cdaee08aaa10b5004977b1f684f346ebe31f\" width=\"800\"/></center>\n",
"<center>Figure 18: Example of SER, RE task</center>\n",
"\n",
"The general KIE method is researched based on Named Entity Recognition (NER) [4], but this type of method only uses the text information in the image and lacks the use of visual and structural information, so the accuracy is not high. On this basis, the methods in recent years have begun to integrate visual and structural information with text information. According to the principles used when fusing multimodal information, these methods can be divided into the following three types:\n",
"\n",
"1. Grid-based approach\n",
"1. Token-based approach\n",
"2. GCN-based method\n",
"3. Based on the End to End method\n",
"\n",
"Some representative papers are divided into the above three categories, as shown in the following table:\n",
"\n",
"| Category | Ideas | Main Papers |\n",
"| ---------------- | ---- | -------- |\n",
"| Grid-based method | Fusion of multi-modal information on images (text, layout, image) | [Chargrid](https://arxiv.org/pdf/1809.08799) |\n",
"| Token-based method|Using methods such as Bert for multi-modal information fusion|[LayoutLM](https://arxiv.org/pdf/1912.13318), [LayoutLMv2](https://arxiv.org/pdf/2012.14740), [StrucText](https://arxiv.org/pdf/2108.02923), |\n",
"| GCN-based method|Using graph network structure for multi-modal information fusion|[GCN](https://arxiv.org/pdf/1903.11279), [PICK](https://arxiv.org/pdf/2004.07464), [SDMG-R](https://arxiv.org/pdf/2103.14470), [SERA](https://arxiv.org/pdf/2110.09915) |\n",
"| Based on End to End method | Unify OCR and key information extraction into one network | [Trie](https://arxiv.org/pdf/2005.13118) |\n",
"\n",
"### 3.2 Grid-Based Method\n",
"\n",
"The Grid-based method performs multimodal information fusion at the image level. Chargrid[5] firstly performs character-level text detection and recognition on the image, and then completes the construction of the network input by filling the one-hot encoding of the category into the corresponding character area (the non-black part in the right image below) , the input is finally passed through the CNN network of the encoder-decoder structure to perform coordinate detection and category classification of key information.\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/f248841769ec4312a9015b4befda37bf29db66226431420ca1faad517783875e\" width=\"800\"/></center>\n",
"<center>Figure 19: Chargrid data example</center>\n",
"\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/0682e52e275b4187a0e74f54961a50091fd3a0cdff734e17bedcbc993f6e29f9\" width=\"800\"/></center>\n",
"<center>Figure 20: Chargrid network</center>\n",
"\n",
"\n",
"Compared with the traditional method based only on text, this method can use both text information and structural information, so it can achieve a certain accuracy improvement. It's good to combine the two.\n",
"\n",
"### 3.3 Token-Based Method\n",
"LayoutLM[6] encodes 2D position information and text information together into the BERT model, and draws on the pre-training ideas of Bert in NLP, pre-training on large-scale data sets, and in downstream tasks, LayoutLM also adds image information To further improve the performance of the model. Although LayoutLM combines text, location and image information, the image information is fused in the training of downstream tasks, so the multi-modal fusion of the three types of information is not sufficient. Based on LayoutLM, LayoutLMv2 [7] integrates image information with text and layout information in the pre-training stage through transformers, and also adds a spatial perception self-attention mechanism to the Transformer to assist the model to better integrate visual and text features. Although LayoutLMv2 fuses text, location and image information in the pre-training stage, the visual features learned by the model are not fine enough due to the limitation of the pre-training task. StrucTexT [8] based on the previous multi-modal methods, proposed two new tasks, Sentence Length Prediction (SLP) and Paired Boxes Direction (PBD) in the pre-training task to help the network learn fine visual features. Among them, the SLP task makes the model Learn the length of the text segment, the PDB task allows the model to learn the matching relationship between Box directions. Through these two new pre-training tasks, the deep cross-modal fusion between text, visual and layout information can be accelerated.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/17a26ade09ee4311b90e49a1c61d88a72a82104478434f9dabd99c27a65d789b) | ![](https://ai-studio-static-online.cdn.bcebos.com/d75addba67ef4b06a02ae40145e609d3692d613ff9b74cec85123335b465b3cc))\n",
"---|---\n",
"Figure 21: Transformer algorithm flow chart | Figure 22: LayoutLMv2 algorithm flow chart\n",
"\n",
"### 3.4 GCN-Based Method\n",
"\n",
"Although the existing GCN-based methods [10] use text and structure information, they do not make good use of image information. PICK [11] added image information to the GCN network and proposed a graph learning module to automatically learn edge types. SDMG-R [12] encodes the image as a bimodal graph. The nodes of the graph are the visual and textual information of the text area. The edges represent the direct spatial relationship between adjacent texts. By iteratively spreading information along the edges and inferring graph node categories, SDMG -R solves the problem that existing methods are incapable of unseen templates.\n",
"\n",
"\n",
"The PICK flow chart is shown in the figure below:\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/d3282959e6b2448c89b762b3b9bbf6197a0364b101214a1f83cf01a28623c01c\" width=\"800\"/></center>\n",
"<center>Figure 23: PICK algorithm flow chart</center>\n",
"\n",
"SERA[10]The biaffine parser in dependency syntax analysis is introduced into document relation extraction, and GCN is used to fuse text and visual information.\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/a97b7647968a4fa59e7b14b384dd7ffe812f158db8f741459b6e6bb0e8b657c7\" width=\"800\"/></center>\n",
"<center>Figure 24: SERA algorithm flow chart</center>\n",
"\n",
"### 3.5 Method Based on End to End\n",
"\n",
"Existing methods divide KIE into two independent tasks: text reading and information extraction. However, they mainly focus on improving the task of information extraction, ignoring that text reading and information extraction are interrelated. Therefore, Trie [9] Proposed a unified end-to-end network that can learn these two tasks at the same time and reinforce each other in the learning process.\n",
"\n",
"<center class=\"img\">\n",
"<img src=\"https://ai-studio-static-online.cdn.bcebos.com/6e4a3b0f65254f6b9d40cea0875854d4f47e1dca6b1e408cad435b3629600608\" width=\"1300\"/></center>\n",
"<center>Figure 25: Trie algorithm flow chart</center>\n",
"\n",
"\n",
"### 3.6 Data Set\n",
"The data sets used for KIE mainly include the following two:\n",
"1. SROIE: Task 3 of the SROIE data set [2] aims to extract four predefined information from the scanned receipt: company, date, address or total. There are 626 samples in the data set for training and 347 samples for testing.\n",
"2. FUNSD: FUNSD data set [3] is a data set used to extract form information from scanned documents. It contains 199 marked real scan forms. Of the 199 samples, 149 are used for training and 50 are used for testing. The FUNSD data set assigns a semantic entity tag to each word: question, answer, title or other.\n",
"3. XFUN: The XFUN data set is a multilingual data set proposed by Microsoft. It contains 7 languages. Each language contains 149 training sets and 50 test sets.\n",
"\n",
"![](https://ai-studio-static-online.cdn.bcebos.com/dfdf530d79504761919c1f093f9a86dac21e6db3304c4892998ea1823f3187c6) | ![](https://ai-studio-static-online.cdn.bcebos.com/3b2a9f9476be4e7f892b73bd7096ce8d88fe98a70bae47e6ab4c5fcc87e83861))\n",
"---|---\n",
"Figure 26: sroie example image | Figure 27: xfun example image\n",
"\n",
"Reference：\n",
"\n",
"[1]：Mathew M, Karatzas D, Jawahar C V. Docvqa: A dataset for vqa on document images[C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2021: 2200-2209.\n",
"\n",
"[2]：Huang Z, Chen K, He J, et al. Icdar2019 competition on scanned receipt ocr and information extraction[C]//2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, 2019: 1516-1520.\n",
"\n",
"[3]：Jaume G, Ekenel H K, Thiran J P. Funsd: A dataset for form understanding in noisy scanned documents[C]//2019 International Conference on Document Analysis and Recognition Workshops (ICDARW). IEEE, 2019, 2: 1-6.\n",
"\n",
"[4]：Lample G, Ballesteros M, Subramanian S, et al. Neural architectures for named entity recognition[J]. arXiv preprint arXiv:1603.01360, 2016.\n",
"\n",
"[5]：Katti A R, Reisswig C, Guder C, et al. Chargrid: Towards understanding 2d documents[J]. arXiv preprint arXiv:1809.08799, 2018.\n",
"\n",
"[6]：Xu Y, Li M, Cui L, et al. Layoutlm: Pre-training of text and layout for document image understanding[C]//Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020: 1192-1200.\n",
"\n",
"[7]：Xu Y, Xu Y, Lv T, et al. LayoutLMv2: Multi-modal pre-training for visually-rich document understanding[J]. arXiv preprint arXiv:2012.14740, 2020.\n",
"\n",
"[8]：Li Y, Qian Y, Yu Y, et al. StrucTexT: Structured Text Understanding with Multi-Modal Transformers[C]//Proceedings of the 29th ACM International Conference on Multimedia. 2021: 1912-1920.\n",
"\n",
"[9]：Zhang P, Xu Y, Cheng Z, et al. Trie: End-to-end text reading and information extraction for document understanding[C]//Proceedings of the 28th ACM International Conference on Multimedia. 2020: 1413-1422.\n",
"\n",
"[10]：Liu X, Gao F, Zhang Q, et al. Graph convolution for multimodal information extraction from visually rich documents[J]. arXiv preprint arXiv:1903.11279, 2019.\n",
"\n",
"[11]：Yu W, Lu N, Qi X, et al. Pick: Processing key information extraction from documents using improved graph learning-convolutional networks[C]//2020 25th International Conference on Pattern Recognition (ICPR). IEEE, 2021: 4363-4370.\n",
"\n",
"[12]：Sun H, Kuang Z, Yue X, et al. Spatial Dual-Modality Graph Reasoning for Key Information Extraction[J]. arXiv preprint arXiv:2103.14470, 2021."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 4. Summary\n",
"In this section, we mainly introduce the theoretical knowledge of three sub-modules related to document analysis technology: layout analysis, table recognition and information extraction. Below we will explain this form recognition and DOC-VQA practical tutorial based on the PaddleOCR framework."
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "py35-paddle1.2.0"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.4"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
notebook/notebook_en/how_to_use_these_notebooks.ipynb 0 → 100644
浏览文件 @ 4b214948
{
"cells": [
{
"cell_type": "markdown",
"metadata": {
"tags": []
},
"source": [
"# 1. Course Prerequisites\n",
"\n",
"The OCR model involved in this course is based on deep learning, so its related basic knowledge, environment configuration, project engineering and other materials will be introduced in this section, especially for readers who are not familiar with deep learning. content.\n",
"\n",
"### 1.1 Preliminary Knowledge\n",
"\n",
"The \"learning\" of deep learning has been developed from the content of neurons, perceptrons, and multilayer neural networks in machine learning. Therefore, understanding the basic machine learning algorithms is of great help to the understanding and application of deep learning. The \"deepness\" of deep learning is embodied in a series of vector-based mathematical operations such as convolution and pooling used in the process of processing a large amount of information. If you lack the theoretical foundation of the two, you can learn from teacher Li Hongyi's [Linear Algebra](https://aistudio.baidu.com/aistudio/course/introduce/2063) and [Machine Learning](https://aistudio.baidu.com/aistudio/course/introduce/1978) courses.\n",
"\n",
"For the understanding of deep learning itself, you can refer to the zero-based course of Bai Ran, an outstanding architect of Baidu: [Baidu architects take you hands-on with zero-based practice deep learning](https://aistudio.baidu.com/aistudio/course/introduce/1297), which covers the development history of deep learning and introduces the complete components of deep learning through a classic case. It is a set of practice-oriented deep learning courses.\n",
"\n",
"For the practice of theoretical knowledge, [Python basic knowledge](https://aistudio.baidu.com/aistudio/course/introduce/1224) is essential. At the same time, in order to quickly reproduce the deep learning model, the deep learning framework used in this course For: Flying PaddlePaddle. If you have used other frameworks, you can quickly learn how to use flying paddles through [Quick Start Document](https://www.paddlepaddle.org.cn/documentation/docs/zh/practices/quick_start/hello_paddle.html).\n",
"\n",
"### 1.2 Basic Environment Preparation\n",
"\n",
"If you want to run the code of this course in a local environment and have not built a Python environment before, you can follow the [zero-base operating environment preparation](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/doc/doc_ch/environment.md), install Anaconda or docker environment according to your operating system.\n",
"\n",
"If you don't have local resources, you can run the code through the AI Studio training platform. Each item in it is presented in a notebook, which is convenient for developers to learn. If you are not familiar with the related operations of Notebook, you can refer to [AI Studio Project Description](https://ai.baidu.com/ai-doc/AISTUDIO/0k3e2tfzm).\n",
"\n",
"### 1.3 Get and Run the Code\n",
"\n",
"This course relies on the formation of PaddleOCR's code repository. First, clone the complete project of PaddleOCR:\n",
"\n",
"```bash\n",
"# [recommend]\n",
"git clone https://github.com/PaddlePaddle/PaddleOCR\n",
"\n",
"# If you cannot pull successfully due to network problems, you can also choose to use the hosting on Code Cloud:\n",
"git clone https://gitee.com/paddlepaddle/PaddleOCR\n",
"```\n",
"\n",
"> Note: The code cloud hosted code may not be able to synchronize the update of this github project in real time, there is a delay of 3~5 days, please use the recommended method first.\n",
">\n",
"> If you are not familiar with git operations, you can download the compressed package directly from the `Code` on the homepage of PaddleOCR\n",
"\n",
"Then install third-party libraries:\n",
"\n",
"```bash\n",
"cd PaddleOCR\n",
"pip3 install -r requirements.txt\n",
"```\n",
"\n",
"\n",
"\n",
"### 1.4 Access to Information\n",
"\n",
"[PaddleOCR Usage Document](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/README.md) describes in detail how to use PaddleOCR to complete model application, training and deployment. The document is rich in content, most of the user’s questions are described in the document or FAQ, especially in [FAQ](https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.3/doc/doc_en/FAQ_en.md), in accordance with the application process of deep learning, has precipitated the user's common questions, it is recommended that you read it carefully.\n",
"\n",
"### 1.5 Ask for Help\n",
"\n",
"If you encounter BUG, ease of use or documentation related issues while using PaddleOCR, you can contact the official via [Github issue](https://github.com/PaddlePaddle/PaddleOCR/issues), please follow the issue template Provide as much information as possible so that official personnel can quickly locate the problem. At the same time, the WeChat group is the daily communication position for the majority of PaddleOCR users, and it is more suitable for asking some consulting questions. In addition to the PaddleOCR team members, there will also be enthusiastic developers answering your questions."
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "py35-paddle1.2.0"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.4"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
ppocr/data/imaug/__init__.py
浏览文件 @ 4b214948
......@@ -22,7 +22,8 @@ from .make_shrink_map import MakeShrinkMap
from .random_crop_data import EastRandomCropData, RandomCropImgMask
from .make_pse_gt import MakePseGt
from .rec_img_aug import RecAug, RecResizeImg, ClsResizeImg, SRNRecResizeImg, NRTRRecResizeImg, SARRecResizeImg
from .rec_img_aug import RecAug, RecResizeImg, ClsResizeImg, \
SRNRecResizeImg, NRTRRecResizeImg, SARRecResizeImg, PRENResizeImg
from .randaugment import RandAugment
from .copy_paste import CopyPaste
from .ColorJitter import ColorJitter
......
ppocr/data/imaug/label_ops.py
浏览文件 @ 4b214948
......@@ -785,6 +785,53 @@ class SARLabelEncode(BaseRecLabelEncode):
return [self.padding_idx]
class PRENLabelEncode(BaseRecLabelEncode):
def __init__(self,
max_text_length,
character_dict_path,
use_space_char=False,
**kwargs):
super(PRENLabelEncode, self).__init__(
max_text_length, character_dict_path, use_space_char)
def add_special_char(self, dict_character):
padding_str = '<PAD>' # 0
end_str = '<EOS>' # 1
unknown_str = '<UNK>' # 2
dict_character = [padding_str, end_str, unknown_str] + dict_character
self.padding_idx = 0
self.end_idx = 1
self.unknown_idx = 2
return dict_character
def encode(self, text):
if len(text) == 0 or len(text) >= self.max_text_len:
return None
if self.lower:
text = text.lower()
text_list = []
for char in text:
if char not in self.dict:
text_list.append(self.unknown_idx)
else:
text_list.append(self.dict[char])
text_list.append(self.end_idx)
if len(text_list) < self.max_text_len:
text_list += [self.padding_idx] * (
self.max_text_len - len(text_list))
return text_list
def __call__(self, data):
text = data['label']
encoded_text = self.encode(text)
if encoded_text is None:
return None
data['label'] = np.array(encoded_text)
return data
class VQATokenLabelEncode(object):
"""
Label encode for NLP VQA methods
......@@ -799,7 +846,7 @@ class VQATokenLabelEncode(object):
ocr_engine=None,
**kwargs):
super(VQATokenLabelEncode, self).__init__()
from paddlenlp.transformers import LayoutXLMTokenizer, LayoutLMTokenizer
from paddlenlp.transformers import LayoutXLMTokenizer, LayoutLMTokenizer, LayoutLMv2Tokenizer
from ppocr.utils.utility import load_vqa_bio_label_maps
tokenizer_dict = {
'LayoutXLM': {
......@@ -809,6 +856,10 @@ class VQATokenLabelEncode(object):
'LayoutLM': {
'class': LayoutLMTokenizer,
'pretrained_model': 'layoutlm-base-uncased'
},
'LayoutLMv2': {
'class': LayoutLMv2Tokenizer,
'pretrained_model': 'layoutlmv2-base-uncased'
}
}
self.contains_re = contains_re
......
ppocr/data/imaug/rec_img_aug.py
浏览文件 @ 4b214948
......@@ -141,6 +141,25 @@ class SARRecResizeImg(object):
return data
class PRENResizeImg(object):
def __init__(self, image_shape, **kwargs):
"""
Accroding to original paper's realization, it's a hard resize method here.
So maybe you should optimize it to fit for your task better.
"""
self.dst_h, self.dst_w = image_shape
def __call__(self, data):
img = data['image']
resized_img = cv2.resize(
img, (self.dst_w, self.dst_h), interpolation=cv2.INTER_LINEAR)
resized_img = resized_img.transpose((2, 0, 1)) / 255
resized_img -= 0.5
resized_img /= 0.5
data['image'] = resized_img.astype(np.float32)
return data
def resize_norm_img_sar(img, image_shape, width_downsample_ratio=0.25):
imgC, imgH, imgW_min, imgW_max = image_shape
h = img.shape[0]
......
ppocr/data/imaug/vqa/token/vqa_token_chunk.py
浏览文件 @ 4b214948
......@@ -12,6 +12,8 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from collections import defaultdict
class VQASerTokenChunk(object):
def __init__(self, max_seq_len=512, infer_mode=False, **kwargs):
......@@ -39,6 +41,8 @@ class VQASerTokenChunk(object):
encoded_inputs_example[key] = data[key]
encoded_inputs_all.append(encoded_inputs_example)
if len(encoded_inputs_all) == 0:
return None
return encoded_inputs_all[0]
......@@ -101,17 +105,18 @@ class VQAReTokenChunk(object):
"entities": self.reformat(entities_in_this_span),
"relations": self.reformat(relations_in_this_span),
})
if len(item['entities']) > 0:
item['entities']['label'] = [
self.entities_labels[x] for x in item['entities']['label']
]
encoded_inputs_all.append(item)
if len(encoded_inputs_all) == 0:
return None
return encoded_inputs_all[0]
def reformat(self, data):
new_data = {}
new_data = defaultdict(list)
for item in data:
for k, v in item.items():
if k not in new_data:
new_data[k] = []
new_data[k].append(v)
return new_data
ppocr/losses/__init__.py
浏览文件 @ 4b214948
......@@ -33,6 +33,7 @@ from .rec_srn_loss import SRNLoss
from .rec_nrtr_loss import NRTRLoss
from .rec_sar_loss import SARLoss
from .rec_aster_loss import AsterLoss
from .rec_pren_loss import PRENLoss
# cls loss
from .cls_loss import ClsLoss
......@@ -59,7 +60,7 @@ def build_loss(config):
'DBLoss', 'PSELoss', 'EASTLoss', 'SASTLoss', 'FCELoss', 'CTCLoss',
'ClsLoss', 'AttentionLoss', 'SRNLoss', 'PGLoss', 'CombinedLoss',
'NRTRLoss', 'TableAttentionLoss', 'SARLoss', 'AsterLoss', 'SDMGRLoss',
'VQASerTokenLayoutLMLoss', 'LossFromOutput'
'VQASerTokenLayoutLMLoss', 'LossFromOutput', 'PRENLoss'
]
config = copy.deepcopy(config)
module_name = config.pop('name')
......
ppocr/losses/rec_ctc_loss.py
浏览文件 @ 4b214948
......@@ -31,7 +31,8 @@ class CTCLoss(nn.Layer):
predicts = predicts[-1]
predicts = predicts.transpose((1, 0, 2))
N, B, _ = predicts.shape
preds_lengths = paddle.to_tensor([N] * B, dtype='int64')
preds_lengths = paddle.to_tensor(
[N] * B, dtype='int64', place=paddle.CPUPlace())
labels = batch[1].astype("int32")
label_lengths = batch[2].astype('int64')
loss = self.loss_func(predicts, labels, preds_lengths, label_lengths)
......
ppocr/losses/rec_pren_loss.py 0 → 100644
浏览文件 @ 4b214948
# copyright (c) 2022 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from paddle import nn
class PRENLoss(nn.Layer):
def __init__(self, **kwargs):
super(PRENLoss, self).__init__()
# note: 0 is padding idx
self.loss_func = nn.CrossEntropyLoss(reduction='mean', ignore_index=0)
def forward(self, predicts, batch):
loss = self.loss_func(predicts, batch[1].astype('int64'))
return {'loss': loss}
ppocr/modeling/backbones/__init__.py
浏览文件 @ 4b214948
......@@ -30,9 +30,10 @@ def build_backbone(config, model_type):
from .rec_resnet_31 import ResNet31
from .rec_resnet_aster import ResNet_ASTER
from .rec_micronet import MicroNet
from .rec_efficientb3_pren import EfficientNetb3_PREN
support_dict = [
'MobileNetV1Enhance', 'MobileNetV3', 'ResNet', 'ResNetFPN', 'MTB',
"ResNet31", "ResNet_ASTER", 'MicroNet'
"ResNet31", "ResNet_ASTER", 'MicroNet', 'EfficientNetb3_PREN'
]
elif model_type == "e2e":
from .e2e_resnet_vd_pg import ResNet
......@@ -45,8 +46,11 @@ def build_backbone(config, model_type):
from .table_mobilenet_v3 import MobileNetV3
support_dict = ["ResNet", "MobileNetV3"]
elif model_type == 'vqa':
from .vqa_layoutlm import LayoutLMForSer, LayoutXLMForSer, LayoutXLMForRe
support_dict = ["LayoutLMForSer", "LayoutXLMForSer", 'LayoutXLMForRe']
from .vqa_layoutlm import LayoutLMForSer, LayoutLMv2ForSer, LayoutLMv2ForRe, LayoutXLMForSer, LayoutXLMForRe
support_dict = [
"LayoutLMForSer", "LayoutLMv2ForSer", 'LayoutLMv2ForRe',
"LayoutXLMForSer", 'LayoutXLMForRe'
]
else:
raise NotImplementedError
......
ppocr/modeling/backbones/rec_efficientb3_pren.py 0 → 100644
浏览文件 @ 4b214948
# copyright (c) 2022 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Code is refer from:
https://github.com/RuijieJ/pren/blob/main/Nets/EfficientNet.py
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import math
from collections import namedtuple
import paddle
import paddle.nn as nn
import paddle.nn.functional as F
__all__ = ['EfficientNetb3']
class EffB3Params:
@staticmethod
def get_global_params():
"""
The fllowing are efficientnetb3's arch superparams, but to fit for scene
text recognition task, the resolution(image_size) here is changed
from 300 to 64.
"""
GlobalParams = namedtuple('GlobalParams', [
'drop_connect_rate', 'width_coefficient', 'depth_coefficient',
'depth_divisor', 'image_size'
])
global_params = GlobalParams(
drop_connect_rate=0.3,
width_coefficient=1.2,
depth_coefficient=1.4,
depth_divisor=8,
image_size=64)
return global_params
@staticmethod
def get_block_params():
BlockParams = namedtuple('BlockParams', [
'kernel_size', 'num_repeat', 'input_filters', 'output_filters',
'expand_ratio', 'id_skip', 'se_ratio', 'stride'
])
block_params = [
BlockParams(3, 1, 32, 16, 1, True, 0.25, 1),
BlockParams(3, 2, 16, 24, 6, True, 0.25, 2),
BlockParams(5, 2, 24, 40, 6, True, 0.25, 2),
BlockParams(3, 3, 40, 80, 6, True, 0.25, 2),
BlockParams(5, 3, 80, 112, 6, True, 0.25, 1),
BlockParams(5, 4, 112, 192, 6, True, 0.25, 2),
BlockParams(3, 1, 192, 320, 6, True, 0.25, 1)
]
return block_params
class EffUtils:
@staticmethod
def round_filters(filters, global_params):
"""Calculate and round number of filters based on depth multiplier."""
multiplier = global_params.width_coefficient
if not multiplier:
return filters
divisor = global_params.depth_divisor
filters *= multiplier
new_filters = int(filters + divisor / 2) // divisor * divisor
if new_filters < 0.9 * filters:
new_filters += divisor
return int(new_filters)
@staticmethod
def round_repeats(repeats, global_params):
"""Round number of filters based on depth multiplier."""
multiplier = global_params.depth_coefficient
if not multiplier:
return repeats
return int(math.ceil(multiplier * repeats))
class ConvBlock(nn.Layer):
def __init__(self, block_params):
super(ConvBlock, self).__init__()
self.block_args = block_params
self.has_se = (self.block_args.se_ratio is not None) and \
(0 < self.block_args.se_ratio <= 1)
self.id_skip = block_params.id_skip
# expansion phase
self.input_filters = self.block_args.input_filters
output_filters = \
self.block_args.input_filters * self.block_args.expand_ratio
if self.block_args.expand_ratio != 1:
self.expand_conv = nn.Conv2D(
self.input_filters, output_filters, 1, bias_attr=False)
self.bn0 = nn.BatchNorm(output_filters)
# depthwise conv phase
k = self.block_args.kernel_size
s = self.block_args.stride
self.depthwise_conv = nn.Conv2D(
output_filters,
output_filters,
groups=output_filters,
kernel_size=k,
stride=s,
padding='same',
bias_attr=False)
self.bn1 = nn.BatchNorm(output_filters)
# squeeze and excitation layer, if desired
if self.has_se:
num_squeezed_channels = max(1,
int(self.block_args.input_filters *
self.block_args.se_ratio))
self.se_reduce = nn.Conv2D(output_filters, num_squeezed_channels, 1)
self.se_expand = nn.Conv2D(num_squeezed_channels, output_filters, 1)
# output phase
self.final_oup = self.block_args.output_filters
self.project_conv = nn.Conv2D(
output_filters, self.final_oup, 1, bias_attr=False)
self.bn2 = nn.BatchNorm(self.final_oup)
self.swish = nn.Swish()
def drop_connect(self, inputs, p, training):
if not training:
return inputs
batch_size = inputs.shape[0]
keep_prob = 1 - p
random_tensor = keep_prob
random_tensor += paddle.rand([batch_size, 1, 1, 1], dtype=inputs.dtype)
random_tensor = paddle.to_tensor(random_tensor, place=inputs.place)
binary_tensor = paddle.floor(random_tensor)
output = inputs / keep_prob * binary_tensor
return output
def forward(self, inputs, drop_connect_rate=None):
# expansion and depthwise conv
x = inputs
if self.block_args.expand_ratio != 1:
x = self.swish(self.bn0(self.expand_conv(inputs)))
x = self.swish(self.bn1(self.depthwise_conv(x)))
# squeeze and excitation
if self.has_se:
x_squeezed = F.adaptive_avg_pool2d(x, 1)
x_squeezed = self.se_expand(self.swish(self.se_reduce(x_squeezed)))
x = F.sigmoid(x_squeezed) * x
x = self.bn2(self.project_conv(x))
# skip conntection and drop connect
if self.id_skip and self.block_args.stride == 1 and \
self.input_filters == self.final_oup:
if drop_connect_rate:
x = self.drop_connect(
x, p=drop_connect_rate, training=self.training)
x = x + inputs
return x
class EfficientNetb3_PREN(nn.Layer):
def __init__(self, in_channels):
super(EfficientNetb3_PREN, self).__init__()
self.blocks_params = EffB3Params.get_block_params()
self.global_params = EffB3Params.get_global_params()
self.out_channels = []
# stem
stem_channels = EffUtils.round_filters(32, self.global_params)
self.conv_stem = nn.Conv2D(
in_channels, stem_channels, 3, 2, padding='same', bias_attr=False)
self.bn0 = nn.BatchNorm(stem_channels)
self.blocks = []
# to extract three feature maps for fpn based on efficientnetb3 backbone
self.concerned_block_idxes = [7, 17, 25]
concerned_idx = 0
for i, block_params in enumerate(self.blocks_params):
block_params = block_params._replace(
input_filters=EffUtils.round_filters(block_params.input_filters,
self.global_params),
output_filters=EffUtils.round_filters(
block_params.output_filters, self.global_params),
num_repeat=EffUtils.round_repeats(block_params.num_repeat,
self.global_params))
self.blocks.append(
self.add_sublayer("{}-0".format(i), ConvBlock(block_params)))
concerned_idx += 1
if concerned_idx in self.concerned_block_idxes:
self.out_channels.append(block_params.output_filters)
if block_params.num_repeat > 1:
block_params = block_params._replace(
input_filters=block_params.output_filters, stride=1)
for j in range(block_params.num_repeat - 1):
self.blocks.append(
self.add_sublayer('{}-{}'.format(i, j + 1),
ConvBlock(block_params)))
concerned_idx += 1
if concerned_idx in self.concerned_block_idxes:
self.out_channels.append(block_params.output_filters)
self.swish = nn.Swish()
def forward(self, inputs):
outs = []
x = self.swish(self.bn0(self.conv_stem(inputs)))
for idx, block in enumerate(self.blocks):
drop_connect_rate = self.global_params.drop_connect_rate
if drop_connect_rate:
drop_connect_rate *= float(idx) / len(self.blocks)
x = block(x, drop_connect_rate=drop_connect_rate)
if idx in self.concerned_block_idxes:
outs.append(x)
return outs
ppocr/modeling/backbones/vqa_layoutlm.py
浏览文件 @ 4b214948
......@@ -21,12 +21,14 @@ from paddle import nn
from paddlenlp.transformers import LayoutXLMModel, LayoutXLMForTokenClassification, LayoutXLMForRelationExtraction
from paddlenlp.transformers import LayoutLMModel, LayoutLMForTokenClassification
from paddlenlp.transformers import LayoutLMv2Model, LayoutLMv2ForTokenClassification, LayoutLMv2ForRelationExtraction
__all__ = ["LayoutXLMForSer", 'LayoutLMForSer']
pretrained_model_dict = {
LayoutXLMModel: 'layoutxlm-base-uncased',
LayoutLMModel: 'layoutlm-base-uncased'
LayoutLMModel: 'layoutlm-base-uncased',
LayoutLMv2Model: 'layoutlmv2-base-uncased'
}
......@@ -58,12 +60,34 @@ class NLPBaseModel(nn.Layer):
self.out_channels = 1
class LayoutXLMForSer(NLPBaseModel):
class LayoutLMForSer(NLPBaseModel):
def __init__(self, num_classes, pretrained=True, checkpoints=None,
**kwargs):
super(LayoutXLMForSer, self).__init__(
LayoutXLMModel,
LayoutXLMForTokenClassification,
super(LayoutLMForSer, self).__init__(
LayoutLMModel,
LayoutLMForTokenClassification,
'ser',
pretrained,
checkpoints,
num_classes=num_classes)
def forward(self, x):
x = self.model(
input_ids=x[0],
bbox=x[2],
attention_mask=x[4],
token_type_ids=x[5],
position_ids=None,
output_hidden_states=False)
return x
class LayoutLMv2ForSer(NLPBaseModel):
def __init__(self, num_classes, pretrained=True, checkpoints=None,
**kwargs):
super(LayoutLMv2ForSer, self).__init__(
LayoutLMv2Model,
LayoutLMv2ForTokenClassification,
'ser',
pretrained,
checkpoints,
......@@ -82,12 +106,12 @@ class LayoutXLMForSer(NLPBaseModel):
return x[0]
class LayoutLMForSer(NLPBaseModel):
class LayoutXLMForSer(NLPBaseModel):
def __init__(self, num_classes, pretrained=True, checkpoints=None,
**kwargs):
super(LayoutLMForSer, self).__init__(
LayoutLMModel,
LayoutLMForTokenClassification,
super(LayoutXLMForSer, self).__init__(
LayoutXLMModel,
LayoutXLMForTokenClassification,
'ser',
pretrained,
checkpoints,
......@@ -97,10 +121,33 @@ class LayoutLMForSer(NLPBaseModel):
x = self.model(
input_ids=x[0],
bbox=x[2],
image=x[3],
attention_mask=x[4],
token_type_ids=x[5],
position_ids=None,
output_hidden_states=False)
head_mask=None,
labels=None)
return x[0]
class LayoutLMv2ForRe(NLPBaseModel):
def __init__(self, pretrained=True, checkpoints=None, **kwargs):
super(LayoutLMv2ForRe, self).__init__(LayoutLMv2Model,
LayoutLMv2ForRelationExtraction,
're', pretrained, checkpoints)
def forward(self, x):
x = self.model(
input_ids=x[0],
bbox=x[1],
labels=None,
image=x[2],
attention_mask=x[3],
token_type_ids=x[4],
position_ids=None,
head_mask=None,
entities=x[5],
relations=x[6])
return x
......
ppocr/modeling/heads/__init__.py
浏览文件 @ 4b214948
......@@ -31,6 +31,7 @@ def build_head(config):
from .rec_nrtr_head import Transformer
from .rec_sar_head import SARHead
from .rec_aster_head import AsterHead
from .rec_pren_head import PRENHead
# cls head
from .cls_head import ClsHead
......@@ -43,7 +44,7 @@ def build_head(config):
support_dict = [
'DBHead', 'PSEHead', 'FCEHead', 'EASTHead', 'SASTHead', 'CTCHead',
'ClsHead', 'AttentionHead', 'SRNHead', 'PGHead', 'Transformer',
'TableAttentionHead', 'SARHead', 'AsterHead', 'SDMGRHead'
'TableAttentionHead', 'SARHead', 'AsterHead', 'SDMGRHead', 'PRENHead'
]
#table head
......
ppocr/modeling/heads/rec_pren_head.py 0 → 100644
浏览文件 @ 4b214948
# copyright (c) 2022 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from paddle import nn
from paddle.nn import functional as F
class PRENHead(nn.Layer):
def __init__(self, in_channels, out_channels, **kwargs):
super(PRENHead, self).__init__()
self.linear = nn.Linear(in_channels, out_channels)
def forward(self, x, targets=None):
predicts = self.linear(x)
if not self.training:
predicts = F.softmax(predicts, axis=2)
return predicts
ppocr/modeling/necks/__init__.py
浏览文件 @ 4b214948
......@@ -24,9 +24,10 @@ def build_neck(config):
from .table_fpn import TableFPN
from .fpn import FPN
from .fce_fpn import FCEFPN
from .pren_fpn import PRENFPN
support_dict = [
'FPN', 'FCEFPN', 'DBFPN', 'EASTFPN', 'SASTFPN', 'SequenceEncoder',
'PGFPN', 'TableFPN'
'PGFPN', 'TableFPN', 'PRENFPN'
]
module_name = config.pop('name')
......
ppocr/modeling/necks/pren_fpn.py 0 → 100644
浏览文件 @ 4b214948
# copyright (c) 2022 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Code is refer from:
https://github.com/RuijieJ/pren/blob/main/Nets/Aggregation.py
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import paddle
from paddle import nn
import paddle.nn.functional as F
class PoolAggregate(nn.Layer):
def __init__(self, n_r, d_in, d_middle=None, d_out=None):
super(PoolAggregate, self).__init__()
if not d_middle:
d_middle = d_in
if not d_out:
d_out = d_in
self.d_in = d_in
self.d_middle = d_middle
self.d_out = d_out
self.act = nn.Swish()
self.n_r = n_r
self.aggs = self._build_aggs()
def _build_aggs(self):
aggs = []
for i in range(self.n_r):
aggs.append(
self.add_sublayer(
'{}'.format(i),
nn.Sequential(
('conv1', nn.Conv2D(
self.d_in, self.d_middle, 3, 2, 1, bias_attr=False)
), ('bn1', nn.BatchNorm(self.d_middle)),
('act', self.act), ('conv2', nn.Conv2D(
self.d_middle, self.d_out, 3, 2, 1, bias_attr=False
)), ('bn2', nn.BatchNorm(self.d_out)))))
return aggs
def forward(self, x):
b = x.shape[0]
outs = []
for agg in self.aggs:
y = agg(x)
p = F.adaptive_avg_pool2d(y, 1)
outs.append(p.reshape((b, 1, self.d_out)))
out = paddle.concat(outs, 1)
return out
class WeightAggregate(nn.Layer):
def __init__(self, n_r, d_in, d_middle=None, d_out=None):
super(WeightAggregate, self).__init__()
if not d_middle:
d_middle = d_in
if not d_out:
d_out = d_in
self.n_r = n_r
self.d_out = d_out
self.act = nn.Swish()
self.conv_n = nn.Sequential(
('conv1', nn.Conv2D(
d_in, d_in, 3, 1, 1,
bias_attr=False)), ('bn1', nn.BatchNorm(d_in)),
('act1', self.act), ('conv2', nn.Conv2D(
d_in, n_r, 1, bias_attr=False)), ('bn2', nn.BatchNorm(n_r)),
('act2', nn.Sigmoid()))
self.conv_d = nn.Sequential(
('conv1', nn.Conv2D(
d_in, d_middle, 3, 1, 1,
bias_attr=False)), ('bn1', nn.BatchNorm(d_middle)),
('act1', self.act), ('conv2', nn.Conv2D(
d_middle, d_out, 1,
bias_attr=False)), ('bn2', nn.BatchNorm(d_out)))
def forward(self, x):
b, _, h, w = x.shape
hmaps = self.conv_n(x)
fmaps = self.conv_d(x)
r = paddle.bmm(
hmaps.reshape((b, self.n_r, h * w)),
fmaps.reshape((b, self.d_out, h * w)).transpose((0, 2, 1)))
return r
class GCN(nn.Layer):
def __init__(self, d_in, n_in, d_out=None, n_out=None, dropout=0.1):
super(GCN, self).__init__()
if not d_out:
d_out = d_in
if not n_out:
n_out = d_in
self.conv_n = nn.Conv1D(n_in, n_out, 1)
self.linear = nn.Linear(d_in, d_out)
self.dropout = nn.Dropout(dropout)
self.act = nn.Swish()
def forward(self, x):
x = self.conv_n(x)
x = self.dropout(self.linear(x))
return self.act(x)
class PRENFPN(nn.Layer):
def __init__(self, in_channels, n_r, d_model, max_len, dropout):
super(PRENFPN, self).__init__()
assert len(in_channels) == 3, "in_channels' length must be 3."
c1, c2, c3 = in_channels # the depths are from big to small
# build fpn
assert d_model % 3 == 0, "{} can't be divided by 3.".format(d_model)
self.agg_p1 = PoolAggregate(n_r, c1, d_out=d_model // 3)
self.agg_p2 = PoolAggregate(n_r, c2, d_out=d_model // 3)
self.agg_p3 = PoolAggregate(n_r, c3, d_out=d_model // 3)
self.agg_w1 = WeightAggregate(n_r, c1, 4 * c1, d_model // 3)
self.agg_w2 = WeightAggregate(n_r, c2, 4 * c2, d_model // 3)
self.agg_w3 = WeightAggregate(n_r, c3, 4 * c3, d_model // 3)
self.gcn_pool = GCN(d_model, n_r, d_model, max_len, dropout)
self.gcn_weight = GCN(d_model, n_r, d_model, max_len, dropout)
self.out_channels = d_model
def forward(self, inputs):
f3, f5, f7 = inputs
rp1 = self.agg_p1(f3)
rp2 = self.agg_p2(f5)
rp3 = self.agg_p3(f7)
rp = paddle.concat([rp1, rp2, rp3], 2) # [b,nr,d]
rw1 = self.agg_w1(f3)
rw2 = self.agg_w2(f5)
rw3 = self.agg_w3(f7)
rw = paddle.concat([rw1, rw2, rw3], 2) # [b,nr,d]
y1 = self.gcn_pool(rp)
y2 = self.gcn_weight(rw)
y = 0.5 * (y1 + y2)
return y # [b,max_len,d]
ppocr/optimizer/__init__.py
浏览文件 @ 4b214948
......@@ -25,11 +25,8 @@ __all__ = ['build_optimizer']
def build_lr_scheduler(lr_config, epochs, step_each_epoch):
from . import learning_rate
lr_config.update({'epochs': epochs, 'step_each_epoch': step_each_epoch})
if 'name' in lr_config:
lr_name = lr_config.pop('name')
lr_name = lr_config.pop('name', 'Const')
lr = getattr(learning_rate, lr_name)(**lr_config)()
else:
lr = lr_config['learning_rate']
return lr
......
ppocr/optimizer/learning_rate.py
浏览文件 @ 4b214948
......@@ -276,3 +276,35 @@ class OneCycle(object):
end_lr=self.max_lr,
last_epoch=self.last_epoch)
return learning_rate
class Const(object):
"""
Const learning rate decay
Args:
learning_rate(float): initial learning rate
step_each_epoch(int): steps each epoch
last_epoch (int, optional): The index of last epoch. Can be set to restart training. Default: -1, means initial learning rate.
"""
def __init__(self,
learning_rate,
step_each_epoch,
warmup_epoch=0,
last_epoch=-1,
**kwargs):
super(Const, self).__init__()
self.learning_rate = learning_rate
self.last_epoch = last_epoch
self.warmup_epoch = round(warmup_epoch * step_each_epoch)
def __call__(self):
learning_rate = self.learning_rate
if self.warmup_epoch > 0:
learning_rate = lr.LinearWarmup(
learning_rate=learning_rate,
warmup_steps=self.warmup_epoch,
start_lr=0.0,
end_lr=self.learning_rate,
last_epoch=self.last_epoch)
return learning_rate
ppocr/postprocess/__init__.py
浏览文件 @ 4b214948
......@@ -25,8 +25,9 @@ from .db_postprocess import DBPostProcess, DistillationDBPostProcess
from .east_postprocess import EASTPostProcess
from .sast_postprocess import SASTPostProcess
from .fce_postprocess import FCEPostProcess
from .rec_postprocess import CTCLabelDecode, AttnLabelDecode, SRNLabelDecode, DistillationCTCLabelDecode, \
TableLabelDecode, NRTRLabelDecode, SARLabelDecode, SEEDLabelDecode
from .rec_postprocess import CTCLabelDecode, AttnLabelDecode, SRNLabelDecode, \
DistillationCTCLabelDecode, TableLabelDecode, NRTRLabelDecode, SARLabelDecode, \
SEEDLabelDecode, PRENLabelDecode
from .cls_postprocess import ClsPostProcess
from .pg_postprocess import PGPostProcess
from .vqa_token_ser_layoutlm_postprocess import VQASerTokenLayoutLMPostProcess
......@@ -40,7 +41,7 @@ def build_post_process(config, global_config=None):
'PGPostProcess', 'DistillationCTCLabelDecode', 'TableLabelDecode',
'DistillationDBPostProcess', 'NRTRLabelDecode', 'SARLabelDecode',
'SEEDLabelDecode', 'VQASerTokenLayoutLMPostProcess',
'VQAReTokenLayoutLMPostProcess'
'VQAReTokenLayoutLMPostProcess', 'PRENLabelDecode'
]
if config['name'] == 'PSEPostProcess':
......
ppocr/postprocess/rec_postprocess.py
浏览文件 @ 4b214948
......@@ -11,8 +11,8 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import string
import paddle
from paddle.nn import functional as F
import re
......@@ -652,3 +652,63 @@ class SARLabelDecode(BaseRecLabelDecode):
def get_ignored_tokens(self):
return [self.padding_idx]
class PRENLabelDecode(BaseRecLabelDecode):
""" Convert between text-label and text-index """
def __init__(self, character_dict_path=None, use_space_char=False,
**kwargs):
super(PRENLabelDecode, self).__init__(character_dict_path,
use_space_char)
def add_special_char(self, dict_character):
padding_str = '<PAD>' # 0
end_str = '<EOS>' # 1
unknown_str = '<UNK>' # 2
dict_character = [padding_str, end_str, unknown_str] + dict_character
self.padding_idx = 0
self.end_idx = 1
self.unknown_idx = 2
return dict_character
def decode(self, text_index, text_prob=None):
""" convert text-index into text-label. """
result_list = []
batch_size = len(text_index)
for batch_idx in range(batch_size):
char_list = []
conf_list = []
for idx in range(len(text_index[batch_idx])):
if text_index[batch_idx][idx] == self.end_idx:
break
if text_index[batch_idx][idx] in \
[self.padding_idx, self.unknown_idx]:
continue
char_list.append(self.character[int(text_index[batch_idx][
idx])])
if text_prob is not None:
conf_list.append(text_prob[batch_idx][idx])
else:
conf_list.append(1)
text = ''.join(char_list)
if len(text) > 0:
result_list.append((text, np.mean(conf_list)))
else:
# here confidence of empty recog result is 1
result_list.append(('', 1))
return result_list
def __call__(self, preds, label=None, *args, **kwargs):
preds = preds.numpy()
preds_idx = preds.argmax(axis=2)
preds_prob = preds.max(axis=2)
text = self.decode(preds_idx, preds_prob)
if label is None:
return text
label = self.decode(label)
return text, label
ppocr/utils/utility.py
浏览文件 @ 4b214948
......@@ -105,3 +105,22 @@ def set_seed(seed=1024):
random.seed(seed)
np.random.seed(seed)
paddle.seed(seed)
class AverageMeter:
def __init__(self):
self.reset()
def reset(self):
"""reset"""
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
"""update"""
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
ppstructure/README.md
浏览文件 @ 4b214948
English | [简体中文](README_ch.md)
- [1. Introduction](#1)
- [2. Update log](#2)
- [3. Features](#3)
- [4. Results](#4)
* [4.1 Layout analysis and table recognition](#41)
* [4.2 DOC-VQA](#42)
- [5. Quick start](#5)
- [6. PP-Structure System](#6)
* [6.1 Layout analysis and table recognition](#61)
* [6.2 DOC-VQA](#62)
- [7. Model List](#7)
<a name="1"></a>
- [1. Introduction](#1-introduction)
- [2. Update log](#2-update-log)
- [3. Features](#3-features)
- [4. Results](#4-results)
- [4.1 Layout analysis and table recognition](#41-layout-analysis-and-table-recognition)
- [4.2 DOC-VQA](#42-doc-vqa)
- [5. Quick start](#5-quick-start)
- [6. PP-Structure System](#6-pp-structure-system)
- [6.1 Layout analysis and table recognition](#61-layout-analysis-and-table-recognition)
- [6.1.1 Layout analysis](#611-layout-analysis)
- [6.1.2 Table recognition](#612-table-recognition)
- [6.2 DOC-VQA](#62-doc-vqa)
- [7. Model List](#7-model-list)
- [7.1 Layout analysis model](#71-layout-analysis-model)
- [7.2 OCR and table recognition model](#72-ocr-and-table-recognition-model)
- [7.3 DOC-VQA model](#73-doc-vqa-model)
## 1. Introduction
PP-Structure is an OCR toolkit that can be used for document analysis and processing with complex structures, designed to help developers better complete document understanding tasks
<a name="2"></a>
## 2. Update log
* 2022.02.12 DOC-VQA add LayoutLMv2 model。
* 2021.12.07 add [DOC-VQA SER and RE tasks](vqa/README.md)。
<a name="3"></a>
## 3. Features
The main features of PP-Structure are as follows:
......@@ -36,26 +36,19 @@ The main features of PP-Structure are as follows:
- Support custom training for layout analysis and table structure tasks
- Support Document Visual Question Answering (DOC-VQA) tasks: Semantic Entity Recognition (SER) and Relation Extraction (RE)
<a name="4"></a>
## 4. Results
<a name="41"></a>
### 4.1 Layout analysis and table recognition
<img src="../doc/table/ppstructure.GIF" width="100%"/>
The figure shows the pipeline of layout analysis + table recognition. The image is first divided into four areas of image, text, title and table by layout analysis, and then OCR detection and recognition is performed on the three areas of image, text and title, and the table is performed table recognition, where the image will also be stored for use.
<a name="42"></a>
### 4.2 DOC-VQA
* SER
![](./vqa/images/result_ser/zh_val_0_ser.jpg) | ![](./vqa/images/result_ser/zh_val_42_ser.jpg)
*
![](../doc/vqa/result_ser/zh_val_0_ser.jpg) | ![](../doc/vqa/result_ser/zh_val_42_ser.jpg)
---|---
Different colored boxes in the figure represent different categories. For xfun dataset, there are three categories: query, answer and header:
......@@ -69,25 +62,18 @@ The corresponding category and OCR recognition results are also marked at the to
* RE
![](./vqa/images/result_re/zh_val_21_re.jpg) | ![](./vqa/images/result_re/zh_val_40_re.jpg)
![](../doc/vqa/result_re/zh_val_21_re.jpg) | ![](../doc/vqa/result_re/zh_val_40_re.jpg)
---|---
In the figure, the red box represents the question, the blue box represents the answer, and the question and answer are connected by green lines. The corresponding category and OCR recognition results are also marked at the top left of the OCR detection box.
<a name="5"></a>
## 5. Quick start
Start from [Quick Installation](./docs/quickstart.md)
<a name="6"></a>
## 6. PP-Structure System
<a name="61"></a>
### 6.1 Layout analysis and table recognition
![pipeline](../doc/table/pipeline.jpg)
......@@ -96,45 +82,39 @@ In PP-Structure, the image will be divided into 5 types of areas **text, title,
#### 6.1.1 Layout analysis
Layout analysis classifies image by region, including the use of Python scripts of layout analysis tools, extraction of designated category detection boxes, performance indicators, and custom training layout analysis models. For details, please refer to [document](layout/README_en.md).
Layout analysis classifies image by region, including the use of Python scripts of layout analysis tools, extraction of designated category detection boxes, performance indicators, and custom training layout analysis models. For details, please refer to [document](layout/README.md).
#### 6.1.2 Table recognition
Table recognition converts table images into excel documents, which include the detection and recognition of table text and the prediction of table structure and cell coordinates. For detailed instructions, please refer to [document](table/README.md)
<a name="62"></a>
### 6.2 DOC-VQA
Document Visual Question Answering (DOC-VQA) if a type of Visual Question Answering (VQA), which includes Semantic Entity Recognition (SER) and Relation Extraction (RE) tasks. Based on SER task, text recognition and classification in images can be completed. Based on THE RE task, we can extract the relation of the text content in the image, such as judge the problem pair. For details, please refer to [document](vqa/README.md)
<a name="7"></a>
## 7. Model List
PP-Structure系列模型列表（更新中）
PP-Structure Series Model List (Updating)
* Layout analysis model
### 7.1 Layout analysis model
|model name|description|download|
| --- | --- | --- |
| ppyolov2_r50vd_dcn_365e_publaynet | The layout analysis model trained on the PubLayNet dataset can divide image into 5 types of areas **text, title, table, picture, and list** | [PubLayNet](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet.tar) |
* OCR and table recognition model
### 7.2 OCR and table recognition model
|model name|description|model size|download|
| --- | --- | --- | --- |
|ch_ppocr_mobile_slim_v2.0_det|Slim pruned lightweight model, supporting Chinese, English, multilingual text detection|2.6M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) |
|ch_ppocr_mobile_slim_v2.0_rec|Slim pruned and quantized lightweight model, supporting Chinese, English and number recognition|6M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_train.tar) |
|en_ppocr_mobile_v2.0_table_structure|Table structure prediction of English table scene trained on PubLayNet dataset|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
|ch_PP-OCRv2_det_slim|[New] Slim quantization with distillation lightweight model, supporting Chinese, English, multilingual text detection| 3M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar)|
|ch_PP-OCRv2_rec_slim|[New] Slim qunatization with distillation lightweight model, supporting Chinese, English, multilingual text recognition| 9M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_train.tar) |
|en_ppocr_mobile_v2.0_table_structure|Table structure prediction of English table scene trained on PubLayNet dataset| 18.6M |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
* DOC-VQA model
### 7.3 DOC-VQA model
|model name|description|model size|download|
| --- | --- | --- | --- |
|PP-Layout_v1.0_ser_pretrained|SER model trained on xfun Chinese dataset based on LayoutXLM|1.4G|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_ser_pretrained.tar) |
|PP-Layout_v1.0_re_pretrained|RE model trained on xfun Chinese dataset based on LayoutXLM|1.4G|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_re_pretrained.tar) |
|ser_LayoutXLM_xfun_zhd|SER model trained on xfun Chinese dataset based on LayoutXLM|1.4G|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
|re_LayoutXLM_xfun_zh|RE model trained on xfun Chinese dataset based on LayoutXLM|1.4G|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
If you need to use other models, you can download the model in [PPOCR model_list](../doc/doc_en/models_list_en.md) and [PPStructure model_list](./docs/model_list.md)
If you need to use other models, you can download the model in [PPOCR model_list](../doc/doc_en/models_list_en.md) and [PPStructure model_list](./docs/models_list.md)
ppstructure/README_ch.md
浏览文件 @ 4b214948
[English](README.md) | 简体中文
- [1. 简介](#1)
- [2. 近期更新](#2)
- [3. 特性](#3)
- [4. 效果展示](#4)
* [4.1 版面分析和表格识别](#41)
* [4.2 DOC-VQA](#42)
- [5. 快速体验](#5)
- [6. PP-Structure 介绍](#6)
* [6.1 版面分析+表格识别](#61)
* [6.2 DOC-VQA](#62)
- [7. 模型库](#7)
<a name="1"></a>
- [1. 简介](#1-简介)
- [2. 近期更新](#2-近期更新)
- [3. 特性](#3-特性)
- [4. 效果展示](#4-效果展示)
- [4.1 版面分析和表格识别](#41-版面分析和表格识别)
- [4.2 DOC-VQA](#42-doc-vqa)
- [5. 快速体验](#5-快速体验)
- [6. PP-Structure 介绍](#6-pp-structure-介绍)
- [6.1 版面分析+表格识别](#61-版面分析表格识别)
- [6.1.1 版面分析](#611-版面分析)
- [6.1.2 表格识别](#612-表格识别)
- [6.2 DOC-VQA](#62-doc-vqa)
- [7. 模型库](#7-模型库)
- [7.1 版面分析模型](#71-版面分析模型)
- [7.2 OCR和表格识别模型](#72-ocr和表格识别模型)
- [7.2 DOC-VQA 模型](#72-doc-vqa-模型)
## 1. 简介
PP-Structure是一个可用于复杂文档结构分析和处理的OCR工具包，旨在帮助开发者更好的完成文档理解相关任务。
<a name="2"></a>
## 2. 近期更新
* 2021.12.07 新增DOC-[VQA任务SER和RE](vqa/README.md)。
<a name="3"></a>
* 2022.02.12 DOC-VQA增加LayoutLMv2模型。
* 2021.12.07 新增[DOC-VQA任务SER和RE](vqa/README.md)。
## 3. 特性
......@@ -34,27 +35,19 @@ PP-Structure的主要特性如下：
- 支持版面分析和表格结构化两类任务自定义训练
- 支持文档视觉问答(Document Visual Question Answering，DOC-VQA)任务-语义实体识别(Semantic Entity Recognition，SER)和关系抽取(Relation Extraction，RE)
<a name="4"></a>
## 4. 效果展示
<a name="41"></a>
### 4.1 版面分析和表格识别
<img src="../doc/table/ppstructure.GIF" width="100%"/>
图中展示了版面分析+表格识别的整体流程，图片先有版面分析划分为图像、文本、标题和表格四种区域，然后对图像、文本和标题三种区域进行OCR的检测识别，对表格进行表格识别，其中图像还会被存储下来以便使用。
<a name="42"></a>
### 4.2 DOC-VQA
* SER
![](./vqa/images/result_ser/zh_val_0_ser.jpg) | ![](./vqa/images/result_ser/zh_val_42_ser.jpg)
![](../doc/vqa/result_ser/zh_val_0_ser.jpg) | ![](../doc/vqa/result_ser/zh_val_42_ser.jpg)
---|---
图中不同颜色的框表示不同的类别，对于XFUN数据集，有`QUESTION`, `ANSWER`, `HEADER` 3种类别
......@@ -67,24 +60,18 @@ PP-Structure的主要特性如下：
* RE
![](./vqa/images/result_re/zh_val_21_re.jpg) | ![](./vqa/images/result_re/zh_val_40_re.jpg)
![](../doc/vqa/result_re/zh_val_21_re.jpg) | ![](../doc/vqa/result_re/zh_val_40_re.jpg)
---|---
图中红色框表示问题，蓝色框表示答案，问题和答案之间使用绿色线连接。在OCR检测框的左上方也标出了对应的类别和OCR识别结果。
<a name="5"></a>
## 5. 快速体验
请参考[快速安装](./docs/quickstart.md)教程。
<a name="6"></a>
## 6. PP-Structure 介绍
<a name="61"></a>
### 6.1 版面分析+表格识别
![pipeline](../doc/table/pipeline.jpg)
......@@ -99,39 +86,34 @@ PP-Structure的主要特性如下：
表格识别将表格图片转换为excel文档，其中包含对于表格文本的检测和识别以及对于表格结构和单元格坐标的预测，详细说明参考[文档](table/README_ch.md)。
<a name="62"></a>
### 6.2 DOC-VQA
DOC-VQA指文档视觉问答，其中包括语义实体识别 (Semantic Entity Recognition, SER) 和关系抽取 (Relation Extraction, RE) 任务。基于 SER 任务，可以完成对图像中的文本识别与分类；基于 RE 任务，可以完成对图象中的文本内容的关系提取，如判断问题对(pair)，详细说明参考[文档](vqa/README.md)。
<a name="7"></a>
## 7. 模型库
PP-Structure系列模型列表（更新中）
* 版面分析模型
### 7.1 版面分析模型
|模型名称|模型简介|下载地址|
| --- | --- | --- |
| ppyolov2_r50vd_dcn_365e_publaynet | PubLayNet 数据集训练的版面分析模型，可以划分**文字、标题、表格、图片以及列表**5类区域 | [PubLayNet](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet.tar) |
* OCR和表格识别模型
### 7.2 OCR和表格识别模型
|模型名称|模型简介|模型大小|下载地址|
| --- | --- | --- | --- |
|ch_ppocr_mobile_slim_v2.0_det|slim裁剪版超轻量模型，支持中英文、多语种文本检测|2.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) |
|ch_ppocr_mobile_slim_v2.0_rec|slim裁剪量化版超轻量模型，支持中英文、数字识别|6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_train.tar) |
|ch_PP-OCRv2_det_slim|【最新】slim量化+蒸馏版超轻量模型，支持中英文、多语种文本检测| 3M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar)|
|ch_PP-OCRv2_rec_slim|【最新】slim量化版超轻量模型，支持中英文、数字识别| 9M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_train.tar) |
|en_ppocr_mobile_v2.0_table_structure|PubLayNet数据集训练的英文表格场景的表格结构预测|18.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
* DOC-VQA 模型
### 7.2 DOC-VQA 模型
|模型名称|模型简介|模型大小|下载地址|
| --- | --- | --- | --- |
|PP-Layout_v1.0_ser_pretrained|基于LayoutXLM在xfun中文数据集上训练的SER模型|1.4G|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_ser_pretrained.tar) |
|PP-Layout_v1.0_re_pretrained|基于LayoutXLM在xfun中文数据集上训练的RE模型|1.4G|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_re_pretrained.tar) |
|ser_LayoutXLM_xfun_zhd|基于LayoutXLM在xfun中文数据集上训练的SER模型|1.4G|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
|re_LayoutXLM_xfun_zh|基于LayoutXLM在xfun中文数据集上训练的RE模型|1.4G|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
更多模型下载，可以参考 [PPOCR model_list](../doc/doc_en/models_list.md) and [PPStructure model_list](./docs/model_list.md)
\ No newline at end of file
更多模型下载，可以参考 [PP-OCR model_list](../doc/doc_ch/models_list.md) and [PP-Structure model_list](./docs/models_list.md)
ppstructure/docs/installation.md
浏览文件 @ 4b214948
- [快速安装](#快速安装)
- [1. PaddlePaddle 和 PaddleOCR](#1-paddlepaddle-和-paddleocr)
- [2. 安装其他依赖](#2-安装其他依赖)
- [2.1 版面分析所需 Layout-Parser](#21-版面分析所需--layout-parser)
- [2.2 VQA所需依赖](#22--vqa所需依赖)
# 快速安装
## 1. PaddlePaddle 和 PaddleOCR
......
ppstructure/docs/kie.md
浏览文件 @ 4b214948
- [关键信息提取(Key Information Extraction)](#关键信息提取key-information-extraction)
- [1. 快速使用](#1-快速使用)
- [2. 执行训练](#2-执行训练)
- [3. 执行评估](#3-执行评估)
- [4. 参考文献](#4-参考文献)
# 关键信息提取(Key Information Extraction)
......@@ -7,11 +11,6 @@
SDMGR是一个关键信息提取算法，将每个检测到的文本区域分类为预定义的类别，如订单ID、发票号码，金额等。
* [1. 快速使用](#1-----)
* [2. 执行训练](#2-----)
* [3. 执行评估](#3-----)
<a name="1-----"></a>
## 1. 快速使用
训练和测试的数据采用wildreceipt数据集，通过如下指令下载数据集：
......@@ -36,7 +35,6 @@ python3.7 tools/infer_kie.py -c configs/kie/kie_unet_sdmgr.yml -o Global.checkpo
<img src="./imgs/0.png" width="800">
</div>
<a name="2-----"></a>
## 2. 执行训练
创建数据集软链到PaddleOCR/train_data目录下：
......@@ -50,7 +48,6 @@ ln -s ../../wildreceipt ./
```
python3.7 tools/train.py -c configs/kie/kie_unet_sdmgr.yml -o Global.save_model_dir=./output/kie/
```
<a name="3-----"></a>
## 3. 执行评估
```
......@@ -58,7 +55,7 @@ python3.7 tools/eval.py -c configs/kie/kie_unet_sdmgr.yml -o Global.checkpoints=
```
**参考文献：**
## 4. 参考文献
<!-- [ALGORITHM] -->
......
ppstructure/docs/kie_en.md
浏览文件 @ 4b214948
- [Key Information Extraction(KIE)](#key-information-extractionkie)
- [1. Quick Use](#1-quick-use)
- [2. Model Training](#2-model-training)
- [3. Model Evaluation](#3-model-evaluation)
- [4. Reference](#4-reference)
# Key Information Extraction(KIE)
......@@ -6,13 +10,6 @@ This section provides a tutorial example on how to quickly use, train, and evalu
[SDMGR(Spatial Dual-Modality Graph Reasoning)](https://arxiv.org/abs/2103.14470) is a KIE algorithm that classifies each detected text region into predefined categories, such as order ID, invoice number, amount, and etc.
* [1. Quick Use](#1-----)
* [2. Model Training](#2-----)
* [3. Model Evaluation](#3-----)
<a name="1-----"></a>
## 1. Quick Use
[Wildreceipt dataset](https://paperswithcode.com/dataset/wildreceipt) is used for this tutorial. It contains 1765 photos, with 25 classes, and 50000 text boxes, which can be downloaded by wget:
......@@ -37,7 +34,6 @@ The visualization results are shown in the figure below:
<img src="./imgs/0.png" width="800">
</div>
<a name="2-----"></a>
## 2. Model Training
Create a softlink to the folder, `PaddleOCR/train_data`:
......@@ -51,7 +47,6 @@ The configuration file used for training is `configs/kie/kie_unet_sdmgr.yml`. Th
```shell
python3.7 tools/train.py -c configs/kie/kie_unet_sdmgr.yml -o Global.save_model_dir=./output/kie/
```
<a name="3-----"></a>
## 3. Model Evaluation
......@@ -61,7 +56,7 @@ After training, you can execute the model evaluation with the following command:
python3.7 tools/eval.py -c configs/kie/kie_unet_sdmgr.yml -o Global.checkpoints=./output/kie/best_accuracy
```
**Reference:**
## 4. Reference
<!-- [ALGORITHM] -->
......
ppstructure/docs/model_list.md → ppstructure/docs/models_list.md
浏览文件 @ 4b214948
# Model List
- [PP-Structure 系列模型列表](#pp-structure-系列模型列表)
- [1. LayoutParser 模型](#1-layoutparser-模型)
- [2. OCR和表格识别模型](#2-ocr和表格识别模型)
- [2.1 OCR](#21-ocr)
- [2.2 表格识别模型](#22-表格识别模型)
- [3. VQA模型](#3-vqa模型)
- [4. KIE模型](#4-kie模型)
# PP-Structure 系列模型列表
## 1. LayoutParser 模型
......@@ -10,25 +19,33 @@
## 2. OCR和表格识别模型
### 2.1 OCR
|模型名称|模型简介|推理模型大小|下载地址|
| --- | --- | --- | --- |
|ch_ppocr_mobile_slim_v2.0_det|slim裁剪版超轻量模型，支持中英文、多语种文本检测|2.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) |
|ch_ppocr_mobile_slim_v2.0_rec|slim裁剪量化版超轻量模型，支持中英文、数字识别|6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_train.tar) |
|en_ppocr_mobile_v2.0_table_det|PubLayNet数据集训练的英文表格场景的文字检测|4.7M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_det_train.tar) |
|en_ppocr_mobile_v2.0_table_rec|PubLayNet数据集训练的英文表格场景的文字识别|6.9M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_rec_train.tar) |
|en_ppocr_mobile_v2.0_table_structure|PubLayNet数据集训练的英文表格场景的表格结构预测|18.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
如需要使用其他OCR模型，可以在 [model_list](../../doc/doc_ch/models_list.md) 下载模型或者使用自己训练好的模型配置到`det_model_dir`,`rec_model_dir`两个字段即可。
如需要使用其他OCR模型，可以在 [PP-OCR model_list](../../doc/doc_ch/models_list.md) 下载模型或者使用自己训练好的模型配置到 `det_model_dir`, `rec_model_dir`两个字段即可。
### 2.2 表格识别模型
|模型名称|模型简介|推理模型大小|下载地址|
| --- | --- | --- | --- |
|en_ppocr_mobile_v2.0_table_structure|PubLayNet数据集训练的英文表格场景的表格结构预测|18.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
## 3. VQA模型
|模型名称|模型简介|推理模型大小|下载地址|
| --- | --- | --- | --- |
|PP-Layout_v1.0_ser_pretrained|基于LayoutXLM在xfun中文数据集上训练的SER模型|1.4G|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
|PP-Layout_v1.0_re_pretrained|基于LayoutXLM在xfun中文数据集上训练的RE模型|1.4G|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
|ser_LayoutXLM_xfun_zh|基于LayoutXLM在xfun中文数据集上训练的SER模型|1.4G|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
|re_LayoutXLM_xfun_zh|基于LayoutXLM在xfun中文数据集上训练的RE模型|1.4G|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
|ser_LayoutLMv2_xfun_zh|基于LayoutLMv2在xfun中文数据集上训练的SER模型|778M|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh.tar) |
|re_LayoutLMv2_xfun_zh|基于LayoutLMv2在xfun中文数据集上训练的RE模型|765M|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutLMv2_xfun_zh.tar) |
|ser_LayoutLM_xfun_zh|基于LayoutLM在xfun中文数据集上训练的SER模型|430M|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar) |
## 3. KIE模型
## 4. KIE模型
|模型名称|模型简介|模型大小|下载地址|
| --- | --- | --- | --- |
|SDMGR|关键信息提取模型|-|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/kie_vgg16.tar)|
|SDMGR|关键信息提取模型|78M|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/kie_vgg16.tar)|
ppstructure/docs/quickstart.md
浏览文件 @ 4b214948
# PP-Structure 快速开始
* [1. 安装PaddleOCR whl包](#1)
* [2. 便捷使用](#2)
+ [2.1 命令行使用](#21)
+ [2.2 Python脚本使用](#22)
+ [2.3 返回结果说明](#23)
+ [2.4 参数说明](#24)
* [3. Python脚本使用](#3)
<a name="1"></a>
- [PP-Structure 快速开始](#pp-structure-快速开始)
- [1. 安装依赖包](#1-安装依赖包)
- [2. 便捷使用](#2-便捷使用)
- [2.1 命令行使用](#21-命令行使用)
- [2.2 Python脚本使用](#22-python脚本使用)
- [2.3 返回结果说明](#23-返回结果说明)
- [2.4 参数说明](#24-参数说明)
- [3. Python脚本使用](#3-python脚本使用)
## 1. 安装依赖包
......@@ -24,12 +22,8 @@ pip3 install -e .
```
<a name="2"></a>
## 2. 便捷使用
<a name="21"></a>
### 2.1 命令行使用
* 版面分析+表格识别
......@@ -41,8 +35,6 @@ paddleocr --image_dir=../doc/table/1.png --type=structure
请参考：[文档视觉问答](../vqa/README.md)。
<a name="22"></a>
### 2.2 Python脚本使用
* 版面分析+表格识别
......@@ -76,8 +68,6 @@ im_show.save('result.jpg')
请参考：[文档视觉问答](../vqa/README.md)。
<a name="23"></a>
### 2.3 返回结果说明
PP-Structure的返回结果为一个dict组成的list，示例如下
......@@ -103,8 +93,6 @@ dict 里各个字段说明如下
请参考：[文档视觉问答](../vqa/README.md)。
<a name="24"></a>
### 2.4 参数说明
| 字段 | 说明 | 默认值 |
......@@ -122,8 +110,6 @@ dict 里各个字段说明如下
运行完成后，每张图片会在`output`字段指定的目录下有一个同名目录，图片里的每个表格会存储为一个excel，图片区域会被裁剪之后保存下来，excel文件和图片名名为表格在图片里的坐标。
<a name="3"></a>
## 3. Python脚本使用
* 版面分析+表格识别
......
ppstructure/layout/README.md
浏览文件 @ 4b214948
English | [简体中文](README_ch.md)
- [Getting Started](#getting-started)
- [1. Install whl package](#1--install-whl-package)
- [2. Quick Start](#2-quick-start)
- [3. PostProcess](#3-postprocess)
- [4. Results](#4-results)
- [5. Training](#5-training)
# Getting Started
[1. Install whl package](#Install)
[2. Quick Start](#QuickStart)
[3. PostProcess](#PostProcess)
[4. Results](#Results)
[5. Training](#Training)
<a name="Install"></a>
## 1. Install whl package
```bash
wget https://paddleocr.bj.bcebos.com/whl/layoutparser-0.0.0-py3-none-any.whl
pip install -U layoutparser-0.0.0-py3-none-any.whl
```
<a name="QuickStart"></a>
## 2. Quick Start
Use LayoutParser to identify the layout of a document:
......@@ -77,8 +68,6 @@ The following model configurations and label maps are currently supported, which
* TableBank word and TableBank latex are trained on datasets of word documents and latex documents respectively;
* Download TableBank dataset contains both word and latex。
<a name="PostProcess"></a>
## 3. PostProcess
Layout parser contains multiple categories, if you only want to get the detection box for a specific category (such as the "Text" category), you can use the following code:
......@@ -119,7 +108,6 @@ Displays results with only the "Text" category：
<div align="center">
<img src="../../doc/table/result_text.jpg" width = "600" />
</div>
<a name="Results"></a>
## 4. Results
......@@ -134,8 +122,6 @@ Displays results with only the "Text" category：
​ **GPU：** a single NVIDIA Tesla P40
<a name="Training"></a>
## 5. Training
The above model is based on [PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection). If you want to train your own layout parser model，please refer to：[train_layoutparser_model](train_layoutparser_model.md)
ppstructure/layout/README_ch.md
浏览文件 @ 4b214948
[English](README.md) | 简体中文
- [版面分析使用说明](#版面分析使用说明)
- [1. 安装whl包](#1--安装whl包)
- [2. 使用](#2-使用)
- [3. 后处理](#3-后处理)
- [4. 指标](#4-指标)
- [5. 训练版面分析模型](#5-训练版面分析模型)
# 版面分析使用说明
[1. 安装whl包](#安装whl包)
[2. 使用](#使用)
[3. 后处理](#后处理)
[4. 指标](#指标)
[5. 训练版面分析模型](#训练版面分析模型)
<a name="安装whl包"></a>
## 1. 安装whl包
```bash
pip install -U https://paddleocr.bj.bcebos.com/whl/layoutparser-0.0.0-py3-none-any.whl
```
<a name="使用"></a>
## 2. 使用
使用layoutparser识别给定文档的布局：
......@@ -76,8 +68,6 @@ show_img.show()
* TableBank word和TableBank latex分别在word文档、latex文档数据集训练；
* 下载的TableBank数据集里同时包含word和latex。
<a name="后处理"></a>
## 3. 后处理
版面分析检测包含多个类别，如果只想获取指定类别(如"Text"类别)的检测框、可以使用下述代码：
......@@ -119,8 +109,6 @@ show_img.show()
<img src="../../doc/table/result_text.jpg" width = "600" />
</div>
<a name="指标"></a>
## 4. 指标
| Dataset | mAP | CPU time cost | GPU time cost |
......@@ -134,8 +122,6 @@ show_img.show()
​ **GPU：** a single NVIDIA Tesla P40
<a name="训练版面分析模型"></a>
## 5. 训练版面分析模型
上述模型基于[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection) 训练，如果您想训练自己的版面分析模型，请参考：[train_layoutparser_model](train_layoutparser_model_ch.md)
ppstructure/layout/train_layoutparser_model.md
浏览文件 @ 4b214948
# Training layout-parse
[1. Installation](#Installation)
​ [1.1 Requirements](#Requirements)
​ [1.2 Install PaddleDetection](#Install_PaddleDetection)
[2. Data preparation](#Data_reparation)
[3. Configuration](#Configuration)
English | [简体中文](train_layoutparser_model_ch.md)
- [Training layout-parse](#training-layout-parse)
- [1. Installation](#1--installation)
- [1.1 Requirements](#11-requirements)
- [1.2 Install PaddleDetection](#12-install-paddledetection)
- [2. Data preparation](#2-data-preparation)
- [3. Configuration](#3-configuration)
- [4. Training](#4-training)
- [5. Prediction](#5-prediction)
- [6. Deployment](#6-deployment)
- [6.1 Export model](#61-export-model)
- [6.2 Inference](#62-inference)
[4. Training](#Training)
[5. Prediction](#Prediction)
[6. Deployment](#Deployment)
​ [6.1 Export model](#Export_model)
​ [6.2 Inference](#Inference)
<a name="Installation"></a>
# Training layout-parse
## 1. Installation
<a name="Requirements"></a>
### 1.1 Requirements
- PaddlePaddle 2.1
......@@ -35,8 +24,6 @@
- CUDA >= 10.1
- cuDNN >= 7.6
<a name="Install_PaddleDetection"></a>
### 1.2 Install PaddleDetection
```bash
......@@ -51,8 +38,6 @@ pip install -r requirements.txt
For more installation tutorials, please refer to： [Install doc](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/INSTALL_cn.md)
<a name="Data_preparation"></a>
## 2. Data preparation
Download the [PubLayNet](https://github.com/ibm-aur-nlp/PubLayNet) dataset
......@@ -80,8 +65,6 @@ PubLayNet directory structure after decompressing ：
For other datasets，please refer to [the PrepareDataSet]((https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/PrepareDataSet.md) )
<a name="Configuration"></a>
## 3. Configuration
We use the `configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml` configuration for training，the configuration file is as follows
......@@ -113,8 +96,6 @@ The `ppyolov2_r50vd_dcn_365e_coco.yml` configuration depends on other configurat
Modify the preceding files, such as the dataset path and batch size etc.
<a name="Training"></a>
## 4. Training
PaddleDetection provides single-card/multi-card training mode to meet various training needs of users:
......@@ -146,8 +127,6 @@ python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/ppy
Note: If you encounter "`Out of memory error`" , try reducing `batch_size` in the `ppyolov2_reader.yml` file
prediction<a name="Prediction"></a>
## 5. Prediction
Set parameters and use PaddleDetection to predict：
......@@ -159,14 +138,10 @@ python tools/infer.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml --infer
`--draw_threshold` is an optional parameter. According to the calculation of [NMS](https://ieeexplore.ieee.org/document/1699659), different threshold will produce different results, ` keep_top_k ` represent the maximum amount of output target, the default value is 10. You can set different value according to your own actual situation。
<a name="Deployment"></a>
## 6. Deployment
Use your trained model in Layout Parser
<a name="Export_model"></a>
### 6.1 Export model
n the process of model training, the model file saved contains the process of forward prediction and back propagation. In the actual industrial deployment, there is no need for back propagation. Therefore, the model should be translated into the model format required by the deployment. The `tools/export_model.py` script is provided in PaddleDetection to export the model.
......@@ -183,8 +158,6 @@ The prediction model is exported to `inference/ppyolov2_r50vd_dcn_365e_coco` ,in
More model export tutorials, please refer to：[EXPORT_MODEL](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/deploy/EXPORT_MODEL.md)
<a name="Inference"></a>
### 6.2 Inference
`model_path` represent the trained model path, and layoutparser is used to predict:
......@@ -194,8 +167,6 @@ import layoutparser as lp
model = lp.PaddleDetectionLayoutModel(model_path="inference/ppyolov2_r50vd_dcn_365e_coco", threshold=0.5,label_map={0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"},enforce_cpu=True,enable_mkldnn=True)
```
***
More PaddleDetection training tutorials，please reference：[PaddleDetection Training](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/GETTING_STARTED_cn.md)
......
ppstructure/layout/train_layoutparser_model_ch.md
浏览文件 @ 4b214948
# 训练版面分析
[1. 安装](#安装)
​ [1.1 环境要求](#环境要求)
​ [1.2 安装PaddleDetection](#安装PaddleDetection)
[2. 准备数据](#准备数据)
[3. 配置文件改动和说明](#配置文件改动和说明)
[4. PaddleDetection训练](#训练)
[5. PaddleDetection预测](#预测)
[6. 预测部署](#预测部署)
​ [6.1 模型导出](#模型导出)
​ [6.2 layout parser预测](#layout_parser预测)
[English](train_layoutparser_model.md) | 简体中文
- [训练版面分析](#训练版面分析)
- [1. 安装](#1-安装)
- [1.1 环境要求](#11-环境要求)
- [1.2 安装PaddleDetection](#12-安装paddledetection)
- [2. 准备数据](#2-准备数据)
- [3. 配置文件改动和说明](#3-配置文件改动和说明)
- [4. PaddleDetection训练](#4-paddledetection训练)
- [5. PaddleDetection预测](#5-paddledetection预测)
- [6. 预测部署](#6-预测部署)
- [6.1 模型导出](#61-模型导出)
- [6.2 layout_parser预测](#62-layout_parser预测)
<a name="安装"></a>
# 训练版面分析
## 1. 安装
<a name="环境要求"></a>
### 1.1 环境要求
- PaddlePaddle 2.1
......@@ -35,8 +24,6 @@
- CUDA >= 10.1
- cuDNN >= 7.6
<a name="安装PaddleDetection"></a>
### 1.2 安装PaddleDetection
```bash
......@@ -51,8 +38,6 @@ pip install -r requirements.txt
更多安装教程，请参考: [Install doc](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/INSTALL_cn.md)
<a name="数据准备"></a>
## 2. 准备数据
下载 [PubLayNet](https://github.com/ibm-aur-nlp/PubLayNet) 数据集：
......@@ -80,8 +65,6 @@ tar -xvf publaynet.tar.gz
如果使用其它数据集，请参考[准备训练数据](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/PrepareDataSet.md)
<a name="配置文件改动和说明"></a>
## 3. 配置文件改动和说明
我们使用 `configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml`配置进行训练，配置文件摘要如下：
......@@ -113,8 +96,6 @@ weights: output/ppyolov2_r50vd_dcn_365e_coco/model_final
根据实际情况，修改上述文件，比如数据集路径、batch size等。
<a name="训练"></a>
## 4. PaddleDetection训练
PaddleDetection提供了单卡/多卡训练模式，满足用户多种训练需求
......@@ -146,8 +127,6 @@ python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/ppy
注意：如果遇到 "`Out of memory error`" 问题, 尝试在 `ppyolov2_reader.yml` 文件中调小`batch_size`
<a name="预测"></a>
## 5. PaddleDetection预测
设置参数，使用PaddleDetection预测：
......@@ -159,14 +138,10 @@ python tools/infer.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml --infer
`--draw_threshold` 是个可选参数. 根据 [NMS](https://ieeexplore.ieee.org/document/1699659) 的计算，不同阈值会产生不同的结果 `keep_top_k`表示设置输出目标的最大数量，默认值为100，用户可以根据自己的实际情况进行设定。
<a name="预测部署"></a>
## 6. 预测部署
在layout parser中使用自己训练好的模型。
<a name="模型导出"></a>
### 6.1 模型导出
在模型训练过程中保存的模型文件是包含前向预测和反向传播的过程，在实际的工业部署则不需要反向传播，因此需要将模型进行导成部署需要的模型格式。 在PaddleDetection中提供了 `tools/export_model.py`脚本来导出模型。
......@@ -183,8 +158,6 @@ python tools/export_model.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml
更多模型导出教程，请参考：[EXPORT_MODEL](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/deploy/EXPORT_MODEL.md)
<a name="layout parser预测"></a>
### 6.2 layout_parser预测
`model_path`指定训练好的模型路径，使用layout parser进行预测：
......
ppstructure/table/README.md
浏览文件 @ 4b214948
- [Table Recognition](#table-recognition)
- [1. pipeline](#1-pipeline)
- [2. Performance](#2-performance)
- [3. How to use](#3-how-to-use)
- [3.1 quick start](#31-quick-start)
- [3.2 Train](#32-train)
- [3.3 Eval](#33-eval)
- [3.4 Inference](#34-inference)
# Table Recognition
## 1. pipeline
......@@ -51,10 +61,10 @@ After running, the excel sheet of each picture will be saved in the directory sp
In this chapter, we only introduce the training of the table structure model, For model training of [text detection](../../doc/doc_en/detection_en.md) and [text recognition](../../doc/doc_en/recognition_en.md), please refer to the corresponding documents
#### data preparation
* data preparation
The training data uses public data set [PubTabNet](https://arxiv.org/abs/1911.10683 ), Can be downloaded from the official [website](https://github.com/ibm-aur-nlp/PubTabNet) 。The PubTabNet data set contains about 500,000 images, as well as annotations in html format。
#### Start training
* Start training
*If you are installing the cpu version of paddle, please modify the `use_gpu` field in the configuration file to false*
```shell
# single GPU training
......@@ -67,7 +77,7 @@ python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/
In the above instruction, use `-c` to select the training to use the `configs/table/table_mv3.yml` configuration file.
For a detailed explanation of the configuration file, please refer to [config](../../doc/doc_en/config_en.md).
#### load trained model and continue training
* load trained model and continue training
If you expect to load trained model and continue the training again, you can specify the parameter `Global.checkpoints` as the model path to be loaded.
......
ppstructure/table/README_ch.md
浏览文件 @ 4b214948
# 表格识别
- [表格识别](#表格识别)
- [1. 表格识别 pipeline](#1-表格识别-pipeline)
- [2. 性能](#2-性能)
- [3. 使用](#3-使用)
- [3.1 快速开始](#31-快速开始)
- [3.2 训练](#32-训练)
- [3.3 评估](#33-评估)
- [3.4 预测](#34-预测)
* [1. 表格识别 pipeline](#1)
* [2. 性能](#2)
* [3. 使用](#3)
+ [3.1 快速开始](#31)
+ [3.2 训练](#32)
+ [3.3 评估](#33)
+ [3.4 预测](#34)
# 表格识别
<a name="1"></a>
## 1. 表格识别 pipeline
表格识别主要包含三个模型
......@@ -28,7 +28,6 @@
4. 单元格的识别结果和表格结构一起构造表格的html字符串。
<a name="2"></a>
## 2. 性能
我们在 PubTabNet<sup>[1]</sup> 评估数据集上对算法进行了评估，性能如下
......@@ -38,9 +37,8 @@
| EDD<sup>[2]</sup> | 88.3 |
| Ours | 93.32 |
<a name="3"></a>
## 3. 使用
<a name="31"></a>
### 3.1 快速开始
```python
......@@ -61,14 +59,17 @@ python3 table/predict_table.py --det_model_dir=inference/en_ppocr_mobile_v2.0_ta
运行完成后，每张图片的excel表格会保存到output字段指定的目录下
note: 上述模型是在 PubLayNet 数据集上训练的表格识别模型，仅支持英文扫描场景，如需识别其他场景需要自己训练模型后替换 `det_model_dir`,`rec_model_dir`,`table_model_dir`三个字段即可。
<a name="32"></a>
### 3.2 训练
在这一章节中，我们仅介绍表格结构模型的训练，[文字检测](../../doc/doc_ch/detection.md)和[文字识别](../../doc/doc_ch/recognition.md)的模型训练请参考对应的文档。
#### 数据准备
* 数据准备
训练数据使用公开数据集PubTabNet ([论文](https://arxiv.org/abs/1911.10683)，[下载地址](https://github.com/ibm-aur-nlp/PubTabNet))。PubTabNet数据集包含约50万张表格数据的图像，以及图像对应的html格式的注释。
#### 启动训练
* 启动训练
*如果您安装的是cpu版本，请将配置文件中的 `use_gpu` 字段修改为false*
```shell
# 单机单卡训练
......@@ -79,7 +80,7 @@ python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/
上述指令中，通过-c 选择训练使用configs/table/table_mv3.yml配置文件。有关配置文件的详细解释，请参考[链接](../../doc/doc_ch/config.md)。
#### 断点训练
* 断点训练
如果训练程序中断，如果希望加载训练中断的模型从而恢复训练，可以通过指定Global.checkpoints指定要加载的模型路径：
```shell
......@@ -88,7 +89,6 @@ python3 tools/train.py -c configs/table/table_mv3.yml -o Global.checkpoints=./yo
**注意**：`Global.checkpoints`的优先级高于`Global.pretrain_weights`的优先级，即同时指定两个参数时，优先加载`Global.checkpoints`指定的模型，如果`Global.checkpoints`指定的模型路径有误，会加载`Global.pretrain_weights`指定的模型。
<a name="33"></a>
### 3.3 评估
表格使用 [TEDS(Tree-Edit-Distance-based Similarity)](https://github.com/ibm-aur-nlp/PubTabNet/tree/master/src) 作为模型的评估指标。在进行模型评估之前，需要将pipeline中的三个模型分别导出为inference模型(我们已经提供好)，还需要准备评估的gt， gt示例如下:
......@@ -113,7 +113,6 @@ python3 table/eval_table.py --det_model_dir=path/to/det_model_dir --rec_model_di
```bash
teds: 93.32
```
<a name="34"></a>
### 3.4 预测
```python
......
ppstructure/vqa/README.md
浏览文件 @ 4b214948
- [文档视觉问答（DOC-VQA）](#文档视觉问答doc-vqa)
- [1. 简介](#1-简介)
- [2. 性能](#2-性能)
- [3. 效果演示](#3-效果演示)
- [3.1 SER](#31-ser)
- [3.2 RE](#32-re)
- [4. 安装](#4-安装)
- [4.1 安装依赖](#41-安装依赖)
- [4.2 安装PaddleOCR（包含 PP-OCR 和 VQA）](#42-安装paddleocr包含-pp-ocr-和-vqa)
- [5. 使用](#5-使用)
- [5.1 数据和预训练模型准备](#51-数据和预训练模型准备)
- [5.2 SER](#52-ser)
- [5.3 RE](#53-re)
- [6. 参考链接](#6-参考链接)
# 文档视觉问答（DOC-VQA）
## 1. 简介
VQA指视觉问答，主要针对图像内容进行提问和回答,DOC-VQA是VQA任务中的一种，DOC-VQA主要针对文本图像的文字内容提出问题。
PP-Structure 里的 DOC-VQA算法基于PaddleNLP自然语言处理算法库进行开发。
......@@ -16,23 +34,23 @@ PP-Structure 里的 DOC-VQA算法基于PaddleNLP自然语言处理算法库进
本项目是 [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/pdf/2104.08836.pdf) 在 Paddle 2.2上的开源实现，
包含了在 [XFUND数据集](https://github.com/doc-analysis/XFUND) 上的微调代码。
## 1 性能
## 2. 性能
我们在 [XFUN](https://github.com/doc-analysis/XFUND) 的中文数据集上对算法进行了评估，性能如下
| 模型 | 任务 | hmean | 模型下载地址 |
|:---:|:---:|:---:| :---:|
| LayoutXLM | RE | 0.7483 | [链接](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
| LayoutXLM | SER | 0.9038 | [链接](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
| LayoutXLM | RE | 0.7483 | [链接](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
| LayoutLMv2 | SER | 0.8544 | [链接](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh.tar)
| LayoutLMv2 | RE | 0.6777 | [链接](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutLMv2_xfun_zh.tar) |
| LayoutLM | SER | 0.7731 | [链接](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar) |
## 2. 效果演示
## 3. 效果演示
**注意：** 测试图片来源于XFUN数据集。
### 2.1 SER
### 3.1 SER
![](../../doc/vqa/result_ser/zh_val_0_ser.jpg) | ![](../../doc/vqa/result_ser/zh_val_42_ser.jpg)
---|---
......@@ -45,8 +63,7 @@ PP-Structure 里的 DOC-VQA算法基于PaddleNLP自然语言处理算法库进
在OCR检测框的左上方也标出了对应的类别和OCR识别结果。
### 2.2 RE
### 3.2 RE
![](../../doc/vqa/result_re/zh_val_21_re.jpg) | ![](../../doc/vqa/result_re/zh_val_40_re.jpg)
---|---
......@@ -54,10 +71,9 @@ PP-Structure 里的 DOC-VQA算法基于PaddleNLP自然语言处理算法库进
图中红色框表示问题，蓝色框表示答案，问题和答案之间使用绿色线连接。在OCR检测框的左上方也标出了对应的类别和OCR识别结果。
## 4. 安装
## 3. 安装
### 3.1 安装依赖
### 4.1 安装依赖
- **（1) 安装PaddlePaddle**
......@@ -73,8 +89,7 @@ python3 -m pip install "paddlepaddle>=2.2" -i https://mirror.baidu.com/pypi/simp
```
更多需求，请参照[安装文档](https://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。
### 3.2 安装PaddleOCR（包含 PP-OCR 和 VQA ）
### 4.2 安装PaddleOCR（包含 PP-OCR 和 VQA）
- **（1）pip快速安装PaddleOCR whl包（仅预测）**
......@@ -99,10 +114,9 @@ git clone https://gitee.com/paddlepaddle/PaddleOCR
python3 -m pip install -r ppstructure/vqa/requirements.txt
```
## 4. 使用
## 5. 使用
### 4.1 数据和预训练模型准备
### 5.1 数据和预训练模型准备
如果希望直接体验预测过程，可以下载我们提供的预训练模型，跳过训练过程，直接预测即可。
......@@ -125,7 +139,7 @@ wget https://paddleocr.bj.bcebos.com/dataset/XFUND.tar
python3 ppstructure/vqa/helper/trans_xfun_data.py --ori_gt_path=path/to/json_path --output_path=path/to/save_path
```
### 4.2 SER任务
### 5.2 SER
启动训练之前，需要修改下面的四个字段
......@@ -164,7 +178,7 @@ CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/vqa/ser/layoutxlm.yml -o
使用如下命令即可完成`OCR引擎 + SER`的串联预测
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/infer_vqa_token_ser.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=PP-Layout_v1.0_ser_pretrained/ Global.infer_img=doc/vqa/input/zh_val_42.jpg
CUDA_VISIBLE_DEVICES=0 python3 tools/infer_vqa_token_ser.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=ser_LayoutXLM_xfun_zh/ Global.infer_img=doc/vqa/input/zh_val_42.jpg
```
最终会在`config.Global.save_res_path`字段所配置的目录下保存预测结果可视化图像以及预测结果文本文件，预测结果文本文件名为`infer_results.txt`。
......@@ -178,8 +192,7 @@ export CUDA_VISIBLE_DEVICES=0
python3 helper/eval_with_label_end2end.py --gt_json_path XFUND/zh_val/xfun_normalize_val.json --pred_json_path output_res/infer_results.txt
```
### 3.3 RE任务
### 5.3 RE
* 启动训练
......@@ -219,13 +232,12 @@ CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/vqa/re/layoutxlm.yml -o
使用如下命令即可完成`OCR引擎 + SER + RE`的串联预测
```shell
export CUDA_VISIBLE_DEVICES=0
python3 tools/infer_vqa_token_ser_re.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=PP-Layout_v1.0_re_pretrained/ Global.infer_img=doc/vqa/input/zh_val_21.jpg -c_ser configs/vqa/ser/layoutxlm.yml -o_ser Architecture.Backbone.checkpoints=PP-Layout_v1.0_ser_pretrained/
python3 tools/infer_vqa_token_ser_re.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=re_LayoutXLM_xfun_zh/ Global.infer_img=doc/vqa/input/zh_val_21.jpg -c_ser configs/vqa/ser/layoutxlm.yml -o_ser Architecture.Backbone.checkpoints=ser_LayoutXLM_xfun_zh/
```
最终会在`config.Global.save_res_path`字段所配置的目录下保存预测结果可视化图像以及预测结果文本文件，预测结果文本文件名为`infer_results.txt`。
## 参考链接
## 6. 参考链接
- LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding, https://arxiv.org/pdf/2104.08836.pdf
- microsoft/unilm/layoutxlm, https://github.com/microsoft/unilm/tree/master/layoutxlm
......
test_tipc/benchmark_train.sh 0 → 100644
浏览文件 @ 4b214948
#!/bin/bash
source test_tipc/common_func.sh
# set env
python=python
export str_tmp=$(echo `pip list|grep paddlepaddle-gpu|awk -F ' ' '{print $2}'`)
export frame_version=${str_tmp%%.post*}
export frame_commit=$(echo `${python} -c "import paddle;print(paddle.version.commit)"`)
# run benchmark sh
# Usage:
# bash run_benchmark_train.sh config.txt params
# or
# bash run_benchmark_train.sh config.txt
function func_parser_params(){
strs=$1
IFS="="
array=(${strs})
tmp=${array[1]}
echo ${tmp}
}
function func_sed_params(){
filename=$1
line=$2
param_value=$3
params=`sed -n "${line}p" $filename`
IFS=":"
array=(${params})
key=${array[0]}
value=${array[1]}
new_params="${key}:${param_value}"
IFS=";"
cmd="sed -i '${line}s/.*/${new_params}/' '${filename}'"
eval $cmd
}
function set_gpu_id(){
string=$1
_str=${string:1:6}
IFS="C"
arr=(${_str})
M=${arr[0]}
P=${arr[1]}
gn=`expr $P - 1`
gpu_num=`expr $gn / $M`
seq=`seq -s "," 0 $gpu_num`
echo $seq
}
function get_repo_name(){
IFS=";"
cur_dir=$(pwd)
IFS="/"
arr=(${cur_dir})
echo ${arr[-1]}
}
FILENAME=$1
# copy FILENAME as new
new_filename="./test_tipc/benchmark_train.txt"
cmd=`yes|cp $FILENAME $new_filename`
FILENAME=$new_filename
# MODE must be one of ['benchmark_train']
MODE=$2
PARAMS=$3
# bash test_tipc/benchmark_train.sh test_tipc/configs/det_mv3_db_v2_0/train_benchmark.txt benchmark_train dynamic_bs8_null_DP_N1C1
IFS=$'\n'
# parser params from train_benchmark.txt
dataline=`cat $FILENAME`
# parser params
IFS=$'\n'
lines=(${dataline})
model_name=$(func_parser_value "${lines[1]}")
# 获取benchmark_params所在的行数
line_num=`grep -n "train_benchmark_params" $FILENAME | cut -d ":" -f 1`
# for train log parser
batch_size=$(func_parser_value "${lines[line_num]}")
line_num=`expr $line_num + 1`
fp_items=$(func_parser_value "${lines[line_num]}")
line_num=`expr $line_num + 1`
epoch=$(func_parser_value "${lines[line_num]}")
line_num=`expr $line_num + 1`
profile_option_key=$(func_parser_key "${lines[line_num]}")
profile_option_params=$(func_parser_value "${lines[line_num]}")
profile_option="${profile_option_key}:${profile_option_params}"
line_num=`expr $line_num + 1`
flags_value=$(func_parser_value "${lines[line_num]}")
# set flags
IFS=";"
flags_list=(${flags_value})
for _flag in ${flags_list[*]}; do
cmd="export ${_flag}"
eval $cmd
done
# set log_name
repo_name=$(get_repo_name )
SAVE_LOG=${BENCHMARK_LOG_DIR:-$(pwd)} # */benchmark_log
mkdir -p "${SAVE_LOG}/benchmark_log/"
status_log="${SAVE_LOG}/benchmark_log/results.log"
# The number of lines in which train params can be replaced.
line_python=3
line_gpuid=4
line_precision=6
line_epoch=7
line_batchsize=9
line_profile=13
line_eval_py=24
line_export_py=30
func_sed_params "$FILENAME" "${line_eval_py}" "null"
func_sed_params "$FILENAME" "${line_export_py}" "null"
func_sed_params "$FILENAME" "${line_python}" "$python"
# if params
if [ ! -n "$PARAMS" ] ;then
# PARAMS input is not a word.
IFS="|"
batch_size_list=(${batch_size})
fp_items_list=(${fp_items})
device_num_list=(N1C4)
run_mode="DP"
else
# parser params from input: modeltype_bs${bs_item}_${fp_item}_${run_mode}_${device_num}
IFS="_"
params_list=(${PARAMS})
model_type=${params_list[0]}
batch_size=${params_list[1]}
batch_size=`echo ${batch_size} | tr -cd "[0-9]" `
precision=${params_list[2]}
# run_process_type=${params_list[3]}
run_mode=${params_list[3]}
device_num=${params_list[4]}
IFS=";"
if [ ${precision} = "null" ];then
precision="fp32"
fi
fp_items_list=($precision)
batch_size_list=($batch_size)
device_num_list=($device_num)
fi
IFS="|"
for batch_size in ${batch_size_list[*]}; do
for precision in ${fp_items_list[*]}; do
for device_num in ${device_num_list[*]}; do
# sed batchsize and precision
func_sed_params "$FILENAME" "${line_precision}" "$precision"
func_sed_params "$FILENAME" "${line_batchsize}" "$MODE=$batch_size"
func_sed_params "$FILENAME" "${line_epoch}" "$MODE=$epoch"
gpu_id=$(set_gpu_id $device_num)
if [ ${#gpu_id} -le 1 ];then
run_process_type="SingleP"
log_path="$SAVE_LOG/profiling_log"
mkdir -p $log_path
log_name="${repo_name}_${model_name}_bs${batch_size}_${precision}_${run_process_type}_${run_mode}_${device_num}_profiling"
func_sed_params "$FILENAME" "${line_gpuid}" "0" # sed used gpu_id
# set profile_option params
tmp=`sed -i "${line_profile}s/.*/${profile_option}/" "${FILENAME}"`
# run test_train_inference_python.sh
cmd="bash test_tipc/test_train_inference_python.sh ${FILENAME} benchmark_train > ${log_path}/${log_name} 2>&1 "
echo $cmd
eval $cmd
eval "cat ${log_path}/${log_name}"
# without profile
log_path="$SAVE_LOG/train_log"
speed_log_path="$SAVE_LOG/index"
mkdir -p $log_path
mkdir -p $speed_log_path
log_name="${repo_name}_${model_name}_bs${batch_size}_${precision}_${run_process_type}_${run_mode}_${device_num}_log"
speed_log_name="${repo_name}_${model_name}_bs${batch_size}_${precision}_${run_process_type}_${run_mode}_${device_num}_speed"
func_sed_params "$FILENAME" "${line_profile}" "null" # sed profile_id as null
cmd="bash test_tipc/test_train_inference_python.sh ${FILENAME} benchmark_train > ${log_path}/${log_name} 2>&1 "
echo $cmd
job_bt=`date '+%Y%m%d%H%M%S'`
eval $cmd
job_et=`date '+%Y%m%d%H%M%S'`
export model_run_time=$((${job_et}-${job_bt}))
eval "cat ${log_path}/${log_name}"
# parser log
_model_name="${model_name}_bs${batch_size}_${precision}_${run_process_type}_${run_mode}"
cmd="${python} ${BENCHMARK_ROOT}/scripts/analysis.py --filename ${log_path}/${log_name} \
--speed_log_file '${speed_log_path}/${speed_log_name}' \
--model_name ${_model_name} \
--base_batch_size ${batch_size} \
--run_mode ${run_mode} \
--run_process_type ${run_process_type} \
--fp_item ${precision} \
--keyword ips: \
--skip_steps 2 \
--device_num ${device_num} \
--speed_unit samples/s \
--convergence_key loss: "
echo $cmd
eval $cmd
last_status=${PIPESTATUS[0]}
status_check $last_status "${cmd}" "${status_log}"
else
IFS=";"
unset_env=`unset CUDA_VISIBLE_DEVICES`
run_process_type="MultiP"
log_path="$SAVE_LOG/train_log"
speed_log_path="$SAVE_LOG/index"
mkdir -p $log_path
mkdir -p $speed_log_path
log_name="${repo_name}_${model_name}_bs${batch_size}_${precision}_${run_process_type}_${run_mode}_${device_num}_log"
speed_log_name="${repo_name}_${model_name}_bs${batch_size}_${precision}_${run_process_type}_${run_mode}_${device_num}_speed"
func_sed_params "$FILENAME" "${line_gpuid}" "$gpu_id" # sed used gpu_id
func_sed_params "$FILENAME" "${line_profile}" "null" # sed --profile_option as null
cmd="bash test_tipc/test_train_inference_python.sh ${FILENAME} benchmark_train > ${log_path}/${log_name} 2>&1 "
echo $cmd
job_bt=`date '+%Y%m%d%H%M%S'`
eval $cmd
job_et=`date '+%Y%m%d%H%M%S'`
export model_run_time=$((${job_et}-${job_bt}))
eval "cat ${log_path}/${log_name}"
# parser log
_model_name="${model_name}_bs${batch_size}_${precision}_${run_process_type}_${run_mode}"
cmd="${python} ${BENCHMARK_ROOT}/scripts/analysis.py --filename ${log_path}/${log_name} \
--speed_log_file '${speed_log_path}/${speed_log_name}' \
--model_name ${_model_name} \
--base_batch_size ${batch_size} \
--run_mode ${run_mode} \
--run_process_type ${run_process_type} \
--fp_item ${precision} \
--keyword ips: \
--skip_steps 2 \
--device_num ${device_num} \
--speed_unit images/s \
--convergence_key loss: "
echo $cmd
eval $cmd
last_status=${PIPESTATUS[0]}
status_check $last_status "${cmd}" "${status_log}"
fi
done
done
done
test_tipc/configs/det_mv3_db_v2.0/train_infer_python.txt → test_tipc/configs/det_mv3_db_v2_0/train_infer_python.txt
浏览文件 @ 4b214948
===========================train_params===========================
model_name:det_mv3_db_v2.0
model_name:det_mv3_db_v2_0
python:python3.7
gpu_list:0|0,1
Global.use_gpu:True|True
......@@ -49,3 +49,9 @@ inference:tools/infer/predict_det.py
null:null
--benchmark:True
null:null
===========================train_benchmark_params==========================
batch_size:8|16
fp_items:fp32|fp16
epoch:2
--profiler_options:batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile
flags:FLAGS_eager_delete_tensor_gb=0.0;FLAGS_fraction_of_gpu_memory_to_use=0.98;FLAGS_conv_workspace_size_limit=4096
\ No newline at end of file
test_tipc/configs/det_r18_vd_v2_0/train_infer_python.txt 0 → 100644
浏览文件 @ 4b214948
===========================train_params===========================
model_name:det_r18_db_v2_0
python:python3.7
gpu_list:0|0,1
Global.use_gpu:True|True
Global.auto_cast:null
Global.epoch_num:lite_train_lite_infer=2|whole_train_whole_infer=300
Global.save_model_dir:./output/
Train.loader.batch_size_per_card:lite_train_lite_infer=2|whole_train_lite_infer=4
Global.pretrained_model:null
train_model_name:latest
train_infer_img_dir:./train_data/icdar2015/text_localization/ch4_test_images/
null:null
##
trainer:norm_train
norm_train:tools/train.py -c configs/det/det_res18_db_v2.0.yml -o
quant_export:null
fpgm_export:null
distill_train:null
null:null
null:null
##
===========================eval_params===========================
eval:null
null:null
##
===========================infer_params===========================
Global.save_inference_dir:./output/
Global.checkpoints:
norm_export:null
quant_export:null
fpgm_export:null
distill_export:null
export1:null
export2:null
##
train_model:null
infer_export:null
infer_quant:False
inference:tools/infer/predict_det.py
--use_gpu:True|False
--enable_mkldnn:True|False
--cpu_threads:1|6
--rec_batch_num:1
--use_tensorrt:False|True
--precision:fp32|fp16|int8
--det_model_dir:
--image_dir:./inference/ch_det_data_50/all-sum-510/
--save_log_path:null
--benchmark:True
null:null
===========================train_benchmark_params==========================
batch_size:8|16
fp_items:fp32|fp16
epoch:2
--profiler_options:batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile
test_tipc/configs/det_r50_vd_east_v2.0/train_infer_python.txt → test_tipc/configs/det_r50_vd_east_v2_0/train_infer_python.txt
浏览文件 @ 4b214948
===========================train_params===========================
model_name:det_r50_vd_east_v2.0
model_name:det_r50_vd_east_v2_0
python:python3.7
gpu_list:0
Global.use_gpu:True|True
......@@ -13,7 +13,7 @@ train_infer_img_dir:./train_data/icdar2015/text_localization/ch4_test_images/
null:null
##
trainer:norm_train
norm_train:tools/train.py -c test_tipc/configs/det_r50_vd_east_v2.0/det_r50_vd_east.yml -o
norm_train:tools/train.py -c test_tipc/configs/det_r50_vd_east_v2_0/det_r50_vd_east.yml -o
pact_train:null
fpgm_train:null
distill_train:null
......@@ -27,7 +27,7 @@ null:null
===========================infer_params===========================
Global.save_inference_dir:./output/
Global.checkpoints:
norm_export:tools/export_model.py -c test_tipc/configs/det_r50_vd_east_v2.0/det_r50_vd_east.yml -o
norm_export:tools/export_model.py -c test_tipc/configs/det_r50_vd_east_v2_0/det_r50_vd_east.yml -o
quant_export:null
fpgm_export:null
distill_export:null
......@@ -35,7 +35,7 @@ export1:null
export2:null
##
train_model:./inference/det_r50_vd_east_v2.0_train/best_accuracy
infer_export:tools/export_model.py -c test_tipc/configs/det_r50_vd_east_v2.0/det_r50_vd_east.yml -o
infer_export:tools/export_model.py -c test_tipc/configs/det_r50_vd_east_v2_0/det_r50_vd_east.yml -o
infer_quant:False
inference:tools/infer/predict_det.py
--use_gpu:True|False
......@@ -49,3 +49,8 @@ inference:tools/infer/predict_det.py
--save_log_path:null
--benchmark:True
--det_algorithm:EAST
===========================train_benchmark_params==========================
batch_size:8
fp_items:fp32|fp16
epoch:2
--profiler_options:batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile
\ No newline at end of file
test_tipc/configs/det_r50_vd_pse_v2.0/train_infer_python.txt → test_tipc/configs/det_r50_vd_pse_v2_0/train_infer_python.txt
浏览文件 @ 4b214948
===========================train_params===========================
model_name:det_r50_vd_pse_v2.0
model_name:det_r50_vd_pse_v2_0
python:python3.7
gpu_list:0
Global.use_gpu:True|True
......@@ -13,7 +13,7 @@ train_infer_img_dir:./train_data/icdar2015/text_localization/ch4_test_images/
null:null
##
trainer:norm_train
norm_train:tools/train.py -c test_tipc/configs/det_r50_vd_pse_v2.0/det_r50_vd_pse.yml -o
norm_train:tools/train.py -c test_tipc/configs/det_r50_vd_pse_v2_0/det_r50_vd_pse.yml -o
pact_train:null
fpgm_train:null
distill_train:null
......@@ -27,7 +27,7 @@ null:null
===========================infer_params===========================
Global.save_inference_dir:./output/
Global.checkpoints:
norm_export:tools/export_model.py -c test_tipc/configs/det_r50_vd_pse_v2.0/det_r50_vd_pse.yml -o
norm_export:tools/export_model.py -c test_tipc/configs/det_r50_vd_pse_v2_0/det_r50_vd_pse.yml -o
quant_export:null
fpgm_export:null
distill_export:null
......@@ -35,7 +35,7 @@ export1:null
export2:null
##
train_model:./inference/det_r50_vd_pse_v2.0_train/best_accuracy
infer_export:tools/export_model.py -c test_tipc/configs/det_r50_vd_pse_v2.0/det_r50_vd_pse.yml -o
infer_export:tools/export_model.py -c test_tipc/configs/det_r50_vd_pse_v2_0/det_r50_vd_pse.yml -o
infer_quant:False
inference:tools/infer/predict_det.py
--use_gpu:True|False
......@@ -49,3 +49,8 @@ inference:tools/infer/predict_det.py
--save_log_path:null
--benchmark:True
--det_algorithm:PSE
===========================train_benchmark_params==========================
batch_size:8
fp_items:fp32|fp16
epoch:2
--profiler_options:batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile
test_tipc/docs/benchmark_train.md 0 → 100644
浏览文件 @ 4b214948
# TIPC Linux端Benchmark测试文档
该文档为Benchmark测试说明，Benchmark预测功能测试的主程序为`benchmark_train.sh`，用于验证监控模型训练的性能。
# 1. 测试流程
## 1.1 准备数据和环境安装
运行`test_tipc/prepare.sh`，完成训练数据准备和安装环境流程。
```shell
# 运行格式：bash test_tipc/prepare.sh train_benchmark.txt mode
bash test_tipc/prepare.sh test_tipc/configs/det_mv3_db_v2_0/train_benchmark.txt benchmark_train
```
## 1.2 功能测试
执行`test_tipc/benchmark_train.sh`，完成模型训练和日志解析
```shell
# 运行格式：bash test_tipc/benchmark_train.sh train_benchmark.txt mode
bash test_tipc/benchmark_train.sh test_tipc/configs/det_mv3_db_v2_0/train_infer_python.txt benchmark_train
```
`test_tipc/benchmark_train.sh`支持根据传入的第三个参数实现只运行某一个训练配置，如下：
```shell
# 运行格式：bash test_tipc/benchmark_train.sh train_benchmark.txt mode
bash test_tipc/benchmark_train.sh test_tipc/configs/det_mv3_db_v2_0/train_infer_python.txt benchmark_train dynamic_bs8_fp32_DP_N1C1
```
dynamic_bs8_fp32_DP_N1C1为test_tipc/benchmark_train.sh传入的参数，格式如下：
`${modeltype}_${batch_size}_${fp_item}_${run_mode}_${device_num}`
包含的信息有：模型类型、batchsize大小、训练精度如fp32,fp16等、分布式运行模式以及分布式训练使用的机器信息如单机单卡（N1C1）。
## 2. 日志输出
运行后将保存模型的训练日志和解析日志，使用 `test_tipc/configs/det_mv3_db_v2_0/train_benchmark.txt` 参数文件的训练日志解析结果是：
```
{"model_branch": "dygaph", "model_commit": "7c39a1996b19087737c05d883fd346d2f39dbcc0", "model_name": "det_mv3_db_v2_0_bs8_fp32_SingleP_DP", "batch_size": 8, "fp_item": "fp32", "run_process_type": "SingleP", "run_mode": "DP", "convergence_value": "5.413110", "convergence_key": "loss:", "ips": 19.333, "speed_unit": "samples/s", "device_num": "N1C1", "model_run_time": "0", "frame_commit": "8cc09552473b842c651ead3b9848d41827a3dbab", "frame_version": "0.0.0"}
```
训练日志和日志解析结果保存在benchmark_log目录下，文件组织格式如下：
```
train_log/
├── index
│   ├── PaddleOCR_det_mv3_db_v2_0_bs8_fp32_SingleP_DP_N1C1_speed
│   └── PaddleOCR_det_mv3_db_v2_0_bs8_fp32_SingleP_DP_N1C4_speed
├── profiling_log
│   └── PaddleOCR_det_mv3_db_v2_0_bs8_fp32_SingleP_DP_N1C1_profiling
└── train_log
├── PaddleOCR_det_mv3_db_v2_0_bs8_fp32_SingleP_DP_N1C1_log
└── PaddleOCR_det_mv3_db_v2_0_bs8_fp32_SingleP_DP_N1C4_log
```
test_tipc/prepare.sh
浏览文件 @ 4b214948
......@@ -20,6 +20,25 @@ model_name=$(func_parser_value "${lines[1]}")
trainer_list=$(func_parser_value "${lines[14]}")
if [ ${MODE} = "benchmark_train" ];then
pip install -r requirements.txt
if [[ ${model_name} =~ "det_mv3_db_v2_0" || ${model_name} =~ "det_r50_vd_east_v2_0" || ${model_name} =~ "det_r50_vd_pse_v2_0" || ${model_name} =~ "det_r18_db_v2_0" ]];then
rm -rf ./train_data/icdar2015
wget -nc -P ./pretrain_models/ https://paddleocr.bj.bcebos.com/pretrained/MobileNetV3_large_x0_5_pretrained.pdparams --no-check-certificate
wget -nc -P ./train_data/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/test/icdar2015.tar --no-check-certificate
cd ./train_data/ && tar xf icdar2015.tar && cd ../
fi
if [[ ${model_name} =~ "det_r50_vd_east_v2_0" || ${model_name} =~ "det_r50_vd_pse_v2_0" ]];then
wget -nc -P ./pretrain_models/ https://paddleocr.bj.bcebos.com/pretrained/ResNet50_vd_ssld_pretrained.pdparams --no-check-certificate
wget -nc -P ./train_data/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/test/icdar2015.tar --no-check-certificate
cd ./train_data/ && tar xf icdar2015.tar && cd ../
fi
if [[ ${model_name} =~ "det_r18_db_v2_0" ]];then
wget -nc -P ./pretrain_models/ https://paddleocr.bj.bcebos.com/pretrained/ResNet18_vd_pretrained.pdparams --no-check-certificate
wget -nc -P ./train_data/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/test/icdar2015.tar --no-check-certificate
cd ./train_data/ && tar xf icdar2015.tar && cd ../
fi
fi
if [ ${MODE} = "lite_train_lite_infer" ];then
# pretrain lite train data
......@@ -52,7 +71,7 @@ if [ ${MODE} = "lite_train_lite_infer" ];then
wget -nc -P ./train_data/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/test/total_text_lite.tar --no-check-certificate
cd ./train_data && tar xf total_text_lite.tar && ln -s total_text_lite total_text && cd ../
fi
if [ ${model_name} == "det_mv3_db_v2.0" ]; then
if [ ${model_name} == "det_mv3_db_v2_0" ]; then
wget -nc -P ./inference/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_mv3_db_v2.0_train.tar --no-check-certificate
cd ./inference/ && tar xf det_mv3_db_v2.0_train.tar && cd ../
fi
......@@ -211,7 +230,7 @@ elif [ ${MODE} = "whole_infer" ];then
wget -nc -P ./inference/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_sast_totaltext_v2.0_train.tar --no-check-certificate
cd ./inference/ && tar xf det_r50_vd_sast_totaltext_v2.0_train.tar && cd ../
fi
if [ ${model_name} == "det_mv3_db_v2.0" ]; then
if [ ${model_name} == "det_mv3_db_v2_0" ]; then
wget -nc -P ./inference/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_mv3_db_v2.0_train.tar --no-check-certificate
cd ./inference/ && tar xf det_mv3_db_v2.0_train.tar && tar xf ch_det_data_50.tar && cd ../
fi
......@@ -223,7 +242,7 @@ elif [ ${MODE} = "whole_infer" ];then
wget -nc -P ./inference/ https://paddleocr.bj.bcebos.com/dygraph_v2.1/en_det/det_mv3_pse_v2.0_train.tar --no-check-certificate
cd ./inference/ && tar xf det_mv3_pse_v2.0_train.tar & cd ../
fi
if [ ${model_name} == "det_r50_vd_pse_v2.0" ]; then
if [ ${model_name} == "det_r50_vd_pse_v2_0" ]; then
wget -nc -P ./inference/ https://paddleocr.bj.bcebos.com/dygraph_v2.1/en_det/det_r50_vd_pse_v2.0_train.tar --no-check-certificate
cd ./inference/ && tar xf det_r50_vd_pse_v2.0_train.tar & cd ../
fi
......@@ -231,7 +250,7 @@ elif [ ${MODE} = "whole_infer" ];then
wget -nc -P ./inference/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_mv3_east_v2.0_train.tar --no-check-certificate
cd ./inference/ && tar xf det_mv3_east_v2.0_train.tar & cd ../
fi
if [ ${model_name} == "det_r50_vd_east_v2.0" ]; then
if [ ${model_name} == "det_r50_vd_east_v2_0" ]; then
wget -nc -P ./inference/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_east_v2.0_train.tar --no-check-certificate
cd ./inference/ && tar xf det_r50_vd_east_v2.0_train.tar & cd ../
fi
......
test_tipc/readme.md
浏览文件 @ 4b214948
......@@ -28,32 +28,32 @@
| DB |ch_ppocr_mobile_v2.0_det_PACT | 检测 | 支持 | 多机多卡 <br> 混合精度 | PACT量化 | Paddle Inference: C++ <br> Paddle Serving: Python, C++ <br> Paddle-Lite: <br> (1) ARM CPU(C++) |
| DB |ch_ppocr_mobile_v2.0_det_KL | 检测 | 支持 | 多机多卡 <br> 混合精度 | 离线量化| Paddle Inference: C++ <br> Paddle Serving: Python, C++ <br> Paddle-Lite: <br> (1) ARM CPU(C++) |
| DB |ch_ppocr_server_v2.0_det | 检测 | 支持 | 多机多卡 <br> 混合精度 | - | Paddle Inference: C++ <br> Paddle Serving: Python, C++ |
| DB |ch_PP-OCRv2_det | 检测 |
| DB |ch_PP-OCRv2_det | 检测 | 支持 | 多机多卡 <br> 混合精度 | - | Paddle Inference: C++ <br> Paddle Serving: Python, C++ |
| CRNN |ch_ppocr_mobile_v2.0_rec | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | Paddle Inference: C++ <br> Paddle Serving: Python, C++ <br> Paddle-Lite: <br> (1) ARM CPU(C++) |
| CRNN |ch_ppocr_server_v2.0_rec | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | Paddle Inference: C++ <br> Paddle Serving: Python, C++ |
| CRNN |ch_PP-OCRv2_rec | 识别 |
| CRNN |ch_PP-OCRv2_rec | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | Paddle Inference: C++ <br> Paddle Serving: Python, C++ |
| PP-OCR |ch_ppocr_mobile_v2.0 | 检测+识别 | 支持 | 多机多卡 <br> 混合精度 | - | Paddle Inference: C++ <br> Paddle Serving: Python, C++ <br> Paddle-Lite: <br> (1) ARM CPU(C++) |
| PP-OCR |ch_ppocr_server_v2.0 | 检测+识别 | 支持 | 多机多卡 <br> 混合精度 | - | Paddle Inference: C++ <br> Paddle Serving: Python, C++ |
|PP-OCRv2|ch_PP-OCRv2 | 检测+识别 |
| DB |det_mv3_db_v2.0 | 检测 |
| DB |det_r50_vd_db_v2.0 | 检测 |
| EAST |det_mv3_east_v2.0 | 检测 |
| EAST |det_r50_vd_east_v2.0 | 检测 |
| PSENet |det_mv3_pse_v2.0 | 检测 |
| PSENet |det_r50_vd_pse_v2.0 | 检测 |
| SAST |det_r50_vd_sast_totaltext_v2.0 | 检测 |
| Rosetta|rec_mv3_none_none_ctc_v2.0 | 识别 |
| Rosetta|rec_r34_vd_none_none_ctc_v2.0 | 识别 |
| CRNN |rec_mv3_none_bilstm_ctc_v2.0 | 识别 |
| CRNN |rec_r34_vd_none_bilstm_ctc_v2.0| 识别 |
| StarNet|rec_mv3_tps_bilstm_ctc_v2.0 | 识别 |
| StarNet|rec_r34_vd_tps_bilstm_ctc_v2.0 | 识别 |
| RARE |rec_mv3_tps_bilstm_att_v2.0 | 识别 |
| RARE |rec_r34_vd_tps_bilstm_att_v2.0 | 识别 |
| SRN |rec_r50fpn_vd_none_srn | 识别 |
| NRTR |rec_mtb_nrtr | 识别 |
| SAR |rec_r31_sar | 识别 |
| PGNet |rec_r34_vd_none_none_ctc_v2.0 | 端到端|
|PP-OCRv2|ch_PP-OCRv2 | 检测+识别 | 支持 | 多机多卡 <br> 混合精度 | - | Paddle Inference: C++ <br> Paddle Serving: Python, C++ |
| DB |det_mv3_db_v2.0 | 检测 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| DB |det_r50_vd_db_v2.0 | 检测 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| EAST |det_mv3_east_v2.0 | 检测 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| EAST |det_r50_vd_east_v2.0 | 检测 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| PSENet |det_mv3_pse_v2.0 | 检测 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| PSENet |det_r50_vd_pse_v2.0 | 检测 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| SAST |det_r50_vd_sast_totaltext_v2.0 | 检测 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| Rosetta|rec_mv3_none_none_ctc_v2.0 | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| Rosetta|rec_r34_vd_none_none_ctc_v2.0 | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| CRNN |rec_mv3_none_bilstm_ctc_v2.0 | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| CRNN |rec_r34_vd_none_bilstm_ctc_v2.0| 识别 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| StarNet|rec_mv3_tps_bilstm_ctc_v2.0 | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| StarNet|rec_r34_vd_tps_bilstm_ctc_v2.0 | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| RARE |rec_mv3_tps_bilstm_att_v2.0 | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| RARE |rec_r34_vd_tps_bilstm_att_v2.0 | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| SRN |rec_r50fpn_vd_none_srn | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| NRTR |rec_mtb_nrtr | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| SAR |rec_r31_sar | 识别 | 支持 | 多机多卡 <br> 混合精度 | - | - |
| PGNet |rec_r34_vd_none_none_ctc_v2.0 | 端到端| 支持 | 多机多卡 <br> 混合精度 | - | - |
......
test_tipc/supplementary/readme.md
浏览文件 @ 4b214948
......@@ -3,7 +3,7 @@
Linux端基础训练预测功能测试的主程序为test_train_python.sh，可以测试基于Python的模型训练、评估等基本功能，包括裁剪、量化、蒸馏训练。
![](./tipc_train.png)
![](./test_tipc/tipc_train.png)
测试链条如上图所示，主要测试内容有带共享权重，自定义OP的模型的正常训练和slim相关功能训练流程是否正常。
......@@ -28,23 +28,30 @@ pip3 install -r requirements.txt
- 模式1：lite_train_lite_infer，使用少量数据训练，用于快速验证训练到预测的走通流程，不验证精度和速度；
```
bash test_tipc/test_train_python.sh ./test_tipc/ch_ppocr_mobile_v2.0_det/train_infer_python.txt 'lite_train_lite_infer'
bash test_tipc/test_train_python.sh ./test_tipc/train_infer_python.txt 'lite_train_lite_infer'
```
- 模式2：whole_train_whole_infer，使用全量数据训练，用于快速验证训练到预测的走通流程，验证模型最终训练精度；
```
bash test_tipc/test_train_python.sh ./test_tipc/ch_ppocr_mobile_v2.0_det/train_infer_python.txt 'whole_train_whole_infer'
bash test_tipc/test_train_python.sh ./test_tipc/train_infer_python.txt 'whole_train_whole_infer'
```
如果是运行量化裁剪等训练方式，需要使用不同的配置文件。量化训练的测试指令如下：
```
bash test_tipc/test_train_python.sh ./test_tipc/ch_ppocr_mobile_v2.0_det/train_infer_python_PACT.txt 'lite_train_lite_infer'
bash test_tipc/test_train_python.sh ./test_tipc/train_infer_python_PACT.txt 'lite_train_lite_infer'
```
同理，FPGM裁剪的运行方式如下：
```
bash test_tipc/test_train_python.sh ./test_tipc/ch_ppocr_mobile_v2.0_det/train_infer_python_FPGM.txt 'lite_train_lite_infer'
bash test_tipc/test_train_python.sh ./test_tipc/train_infer_python_FPGM.txt 'lite_train_lite_infer'
```
多机多卡的运行配置文件分别为 `train_infer_python_fleet.txt`, `train_infer_python_FPGM_fleet.txt` 和 `train_infer_python_PACT_fleet.txt`。
运行时，需要修改配置文件中的 `gpu_list:xx.xx.xx.xx,yy.yy.yy.yy;0,1`。 将 `xx.xx.xx.xx` 替换为具体的 `ip` 地址，各个`ip`地址之间用`,`分隔。 另外，和单机训练
不同，启动多机多卡训练需要在多机的每个节点上分别运行命令。以多机多卡量化训练为例，指令如下：
```
bash test_tipc/test_train_python.sh ./test_tipc/train_infer_python_PACT_fleet.txt 'lite_train_lite_infer'
```
运行相应指令后，在`test_tipc/output`文件夹下自动会保存运行日志。如'lite_train_lite_infer'模式运行后，在test_tipc/extra_output文件夹有以下文件：
......
test_tipc/supplementary/test_tipc/test_train_python.sh
浏览文件 @ 4b214948
......@@ -35,7 +35,6 @@ use_share_conv_key=$(func_parser_key "${lines[13]}")
use_share_conv_list=$(func_parser_value "${lines[13]}")
run_train_py=$(func_parser_value "${lines[14]}")
LOG_PATH="./test_tipc/extra_output"
mkdir -p ${LOG_PATH}
status_log="${LOG_PATH}/results_python.log"
......@@ -98,6 +97,8 @@ if [ ${MODE} = "lite_train_lite_infer" ] || [ ${MODE} = "whole_train_whole_infer
cmd="${python} ${run_train_py} ${set_use_gpu} ${set_save_model} ${set_epoch} ${set_pretrain} ${set_checkpoints} ${set_autocast} ${set_batchsize} ${set_use_custom_op} ${set_model_type} ${set_use_share_conv} ${set_amp_config}"
elif [ ${#ips} -le 26 ];then # train with multi-gpu
cmd="${python} -m paddle.distributed.launch --gpus=${gpu} ${run_train_py} ${set_use_gpu} ${set_save_model} ${set_epoch} ${set_pretrain} ${set_checkpoints} ${set_autocast} ${set_batchsize} ${set_use_custom_op} ${set_model_type} ${set_use_share_conv} ${set_amp_config}"
else
cmd="${python} -m paddle.distributed.launch --ips=${ips} --gpus=${gpu} ${run_train_py} ${set_use_gpu} ${set_save_model} ${set_epoch} ${set_pretrain} ${set_checkpoints} ${set_autocast} ${set_batchsize} ${set_use_custom_op} ${set_model_type} ${set_use_share_conv} ${set_amp_config}"
fi
# run train
......
test_tipc/supplementary/test_tipc/train_infer_python_FPGM.txt
浏览文件 @ 4b214948
......@@ -4,9 +4,9 @@ python:python3.7
gpu_list:0|0,1
use_gpu:True|True
AMP.use_amp:True|False
epoch:lite_train_lite_infer=20|whole_train_whole_infer=1000
epoch:lite_train_lite_infer=2|whole_train_whole_infer=1000
save_model_dir:./output/
TRAIN.batch_size:lite_train_lite_infer=2|whole_train_whole_infer=4
TRAIN.batch_size:lite_train_lite_infer=1280|whole_train_whole_infer=1280
pretrained_model:null
checkpoints:null
use_custom_relu:False|True
......
test_tipc/supplementary/test_tipc/train_infer_python_FPGM_fleet.txt 0 → 100644
浏览文件 @ 4b214948
===========================train_params===========================
model_name:ch_PPOCRv2_det
python:python3.7
gpu_list:xx.xx.xx.xx,yy.yy.yy.yy;0,1
use_gpu:True
AMP.use_amp:True|False
epoch:lite_train_lite_infer=2|whole_train_whole_infer=1000
save_model_dir:./output/
TRAIN.batch_size:lite_train_lite_infer=1280|whole_train_whole_infer=1280
pretrained_model:null
checkpoints:null
use_custom_relu:False|True
model_type:cls|cls_distill|cls_distill_multiopt
MODEL.siamese:False|True
norm_train:train.py -c mv3_large_x0_5.yml -o prune_train=True
quant_train:False
prune_train:False
test_tipc/supplementary/test_tipc/train_infer_python_PACT.txt
浏览文件 @ 4b214948
......@@ -4,9 +4,9 @@ python:python3.7
gpu_list:0|0,1
use_gpu:True|True
AMP.use_amp:True|False
epoch:lite_train_lite_infer=20|whole_train_whole_infer=1000
epoch:lite_train_lite_infer=2|whole_train_whole_infer=1000
save_model_dir:./output/
TRAIN.batch_size:lite_train_lite_infer=2|whole_train_whole_infer=4
TRAIN.batch_size:lite_train_lite_infer=1280|whole_train_whole_infer=1280
pretrained_model:null
checkpoints:null
use_custom_relu:False|True
......
test_tipc/supplementary/test_tipc/train_infer_python_PACT_fleet.txt 0 → 100644
浏览文件 @ 4b214948
===========================train_params===========================
model_name:ch_PPOCRv2_det
python:python3.7
gpu_list:xx.xx.xx.xx,yy.yy.yy.yy;0,1
use_gpu:True
AMP.use_amp:True|False
epoch:lite_train_lite_infer=2|whole_train_whole_infer=1000
save_model_dir:./output/
TRAIN.batch_size:lite_train_lite_infer=1280|whole_train_whole_infer=1280
pretrained_model:null
checkpoints:null
use_custom_relu:False|True
model_type:cls|cls_distill|cls_distill_multiopt
MODEL.siamese:False|True
norm_train:train.py -c mv3_large_x0_5.yml -o quant_train=True
quant_train:False
prune_train:False
test_tipc/supplementary/test_tipc/train_infer_python_fleet.txt 0 → 100644
浏览文件 @ 4b214948
===========================train_params===========================
model_name:ch_PPOCRv2_det
python:python3.7
gpu_list:xx.xx.xx.xx,yy.yy.yy.yy;0,1
use_gpu:True
AMP.use_amp:True|False
epoch:lite_train_lite_infer=2|whole_train_whole_infer=1000
save_model_dir:./output/
TRAIN.batch_size:lite_train_lite_infer=1280|whole_train_whole_infer=1280
pretrained_model:null
checkpoints:null
use_custom_relu:False|True
model_type:cls|cls_distill|cls_distill_multiopt
MODEL.siamese:False|True
norm_train: train.py -c mv3_large_x0_5.yml -o
quant_train:False
prune_train:False
test_tipc/test_train_inference_python.sh
浏览文件 @ 4b214948
......@@ -284,7 +284,6 @@ else
set_amp_config=" "
fi
for trainer in ${trainer_list[*]}; do
eval ${env}
flag_quant=False
if [ ${trainer} = ${pact_key} ]; then
run_train=${pact_trainer}
......@@ -344,6 +343,7 @@ else
# run eval
if [ ${eval_py} != "null" ]; then
eval ${env}
set_eval_params1=$(func_set_params "${eval_key1}" "${eval_value1}")
eval_cmd="${python} ${eval_py} ${set_eval_pretrain} ${set_use_gpu} ${set_eval_params1}"
eval $eval_cmd
......
tools/eval.py
浏览文件 @ 4b214948
......@@ -28,7 +28,6 @@ from ppocr.modeling.architectures import build_model
from ppocr.postprocess import build_post_process
from ppocr.metrics import build_metric
from ppocr.utils.save_load import load_model
from ppocr.utils.utility import print_dict
import tools.program as program
......
tools/export_model.py
浏览文件 @ 4b214948
......@@ -55,6 +55,12 @@ def export_single_model(model, arch_config, save_path, logger):
shape=[None, 3, 48, 160], dtype="float32"),
]
model = to_static(model, input_spec=other_shape)
elif arch_config["algorithm"] == "PREN":
other_shape = [
paddle.static.InputSpec(
shape=[None, 3, 64, 512], dtype="float32"),
]
model = to_static(model, input_spec=other_shape)
else:
infer_shape = [3, -1, -1]
if arch_config["model_type"] == "rec":
......
tools/infer/predict_system.py
浏览文件 @ 4b214948
......@@ -24,6 +24,7 @@ os.environ["FLAGS_allocator_strategy"] = 'auto_growth'
import cv2
import copy
import numpy as np
import json
import time
import logging
from PIL import Image
......@@ -92,11 +93,11 @@ class TextSystem(object):
self.draw_crop_rec_res(self.args.crop_res_save_dir, img_crop_list,
rec_res)
filter_boxes, filter_rec_res = [], []
for box, rec_reuslt in zip(dt_boxes, rec_res):
text, score = rec_reuslt
for box, rec_result in zip(dt_boxes, rec_res):
text, score = rec_result
if score >= self.drop_score:
filter_boxes.append(box)
filter_rec_res.append(rec_reuslt)
filter_rec_res.append(rec_result)
return filter_boxes, filter_rec_res
......@@ -128,6 +129,9 @@ def main(args):
is_visualize = True
font_path = args.vis_font_path
drop_score = args.drop_score
draw_img_save_dir = args.draw_img_save_dir
os.makedirs(draw_img_save_dir, exist_ok=True)
save_results = []
# warm up 10 times
if args.warmup:
......@@ -157,6 +161,14 @@ def main(args):
for text, score in rec_res:
logger.debug("{}, {:.3f}".format(text, score))
res = [{
"transcription": rec_res[idx][0],
"points": np.array(dt_boxes[idx]).astype(np.int32).tolist(),
} for idx in range(len(dt_boxes))]
save_pred = os.path.basename(image_file) + "\t" + json.dumps(
res, ensure_ascii=False) + "\n"
save_results.append(save_pred)
if is_visualize:
image = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
boxes = dt_boxes
......@@ -170,8 +182,6 @@ def main(args):
scores,
drop_score=drop_score,
font_path=font_path)
draw_img_save_dir = args.draw_img_save_dir
os.makedirs(draw_img_save_dir, exist_ok=True)
if flag:
image_file = image_file[:-3] + "png"
cv2.imwrite(
......@@ -185,6 +195,9 @@ def main(args):
text_sys.text_detector.autolog.report()
text_sys.text_recognizer.autolog.report()
with open(os.path.join(draw_img_save_dir, "system_results.txt"), 'w') as f:
f.writelines(save_results)
if __name__ == "__main__":
args = utility.parse_args()
......
tools/infer_cls.py
浏览文件 @ 4b214948
......@@ -73,8 +73,8 @@ def main():
images = paddle.to_tensor(images)
preds = model(images)
post_result = post_process_class(preds)
for rec_reuslt in post_result:
logger.info('\t result: {}'.format(rec_reuslt))
for rec_result in post_result:
logger.info('\t result: {}'.format(rec_result))
logger.info("success!")
......
tools/program.py
浏览文件 @ 4b214948
......@@ -21,7 +21,7 @@ import sys
import platform
import yaml
import time
import shutil
import datetime
import paddle
import paddle.distributed as dist
from tqdm import tqdm
......@@ -29,11 +29,10 @@ from argparse import ArgumentParser, RawDescriptionHelpFormatter
from ppocr.utils.stats import TrainingStats
from ppocr.utils.save_load import save_model
from ppocr.utils.utility import print_dict
from ppocr.utils.utility import print_dict, AverageMeter
from ppocr.utils.logging import get_logger
from ppocr.utils import profiler
from ppocr.data import build_dataloader
import numpy as np
class ArgsParser(ArgumentParser):
......@@ -48,7 +47,8 @@ class ArgsParser(ArgumentParser):
'--profiler_options',
type=str,
default=None,
help='The option of profiler, which should be in format \"key1=value1;key2=value2;key3=value3\".'
help='The option of profiler, which should be in format ' \
'\"key1=value1;key2=value2;key3=value3\".'
)
def parse_args(self, argv=None):
......@@ -99,7 +99,8 @@ def merge_config(config, opts):
sub_keys = key.split('.')
assert (
sub_keys[0] in config
), "the sub_keys can only be one of global_config: {}, but get: {}, please check your running command".format(
), "the sub_keys can only be one of global_config: {}, but get: " \
"{}, please check your running command".format(
config.keys(), sub_keys[0])
cur = config[sub_keys[0]]
for idx, sub_key in enumerate(sub_keys[1:]):
......@@ -129,6 +130,25 @@ def check_gpu(use_gpu):
pass
def check_xpu(use_xpu):
"""
Log error and exit when set use_xpu=true in paddlepaddle
cpu/gpu version.
"""
err = "Config use_xpu cannot be set as true while you are " \
"using paddlepaddle cpu/gpu version ! \nPlease try: \n" \
"\t1. Install paddlepaddle-xpu to run model on XPU \n" \
"\t2. Set use_xpu as false in config file to run " \
"model on CPU/GPU"
try:
if use_xpu and not paddle.is_compiled_with_xpu():
print(err)
sys.exit(1)
except Exception as e:
pass
def train(config,
train_dataloader,
valid_dataloader,
......@@ -145,6 +165,7 @@ def train(config,
scaler=None):
cal_metric_during_train = config['Global'].get('cal_metric_during_train',
False)
calc_epoch_interval = config['Global'].get('calc_epoch_interval', 1)
log_smooth_window = config['Global']['log_smooth_window']
epoch_num = config['Global']['epoch_num']
print_batch_step = config['Global']['print_batch_step']
......@@ -160,11 +181,13 @@ def train(config,
eval_batch_step = eval_batch_step[1]
if len(valid_dataloader) == 0:
logger.info(
'No Images in eval dataset, evaluation during training will be disabled'
'No Images in eval dataset, evaluation during training ' \
'will be disabled'
)
start_eval_step = 1e111
logger.info(
"During the training process, after the {}th iteration, an evaluation is run every {} iterations".
"During the training process, after the {}th iteration, " \
"an evaluation is run every {} iterations".
format(start_eval_step, eval_batch_step))
save_epoch_step = config['Global']['save_epoch_step']
save_model_dir = config['Global']['save_model_dir']
......@@ -189,10 +212,11 @@ def train(config,
start_epoch = best_model_dict[
'start_epoch'] if 'start_epoch' in best_model_dict else 1
train_reader_cost = 0.0
train_run_cost = 0.0
total_samples = 0
train_reader_cost = 0.0
train_batch_cost = 0.0
reader_start = time.time()
eta_meter = AverageMeter()
max_iter = len(train_dataloader) - 1 if platform.system(
) == "Windows" else len(train_dataloader)
......@@ -203,7 +227,6 @@ def train(config,
config, 'Train', device, logger, seed=epoch)
max_iter = len(train_dataloader) - 1 if platform.system(
) == "Windows" else len(train_dataloader)
for idx, batch in enumerate(train_dataloader):
profiler.add_profiler_step(profiler_options)
train_reader_cost += time.time() - reader_start
......@@ -214,7 +237,6 @@ def train(config,
if use_srn:
model_average = True
train_start = time.time()
# use amp
if scaler:
with paddle.amp.auto_cast():
......@@ -242,7 +264,19 @@ def train(config,
optimizer.step()
optimizer.clear_grad()
train_run_cost += time.time() - train_start
if cal_metric_during_train and epoch % calc_epoch_interval == 0: # only rec and cls need
batch = [item.numpy() for item in batch]
if model_type in ['table', 'kie']:
eval_class(preds, batch)
else:
post_result = post_process_class(preds, batch[1])
eval_class(post_result, batch)
metric = eval_class.get_metric()
train_stats.update(metric)
train_batch_time = time.time() - reader_start
train_batch_cost += train_batch_time
eta_meter.update(train_batch_time)
global_step += 1
total_samples += len(images)
......@@ -254,16 +288,6 @@ def train(config,
stats['lr'] = lr
train_stats.update(stats)
if cal_metric_during_train: # only rec and cls need
batch = [item.numpy() for item in batch]
if model_type in ['table', 'kie']:
eval_class(preds, batch)
else:
post_result = post_process_class(preds, batch[1])
eval_class(post_result, batch)
metric = eval_class.get_metric()
train_stats.update(metric)
if vdl_writer is not None and dist.get_rank() == 0:
for k, v in train_stats.get().items():
vdl_writer.add_scalar('TRAIN/{}'.format(k), v, global_step)
......@@ -273,19 +297,27 @@ def train(config,
(global_step > 0 and global_step % print_batch_step == 0) or
(idx >= len(train_dataloader) - 1)):
logs = train_stats.log()
strs = 'epoch: [{}/{}], global_step: {}, {}, avg_reader_cost: {:.5f} s, avg_batch_cost: {:.5f} s, avg_samples: {}, ips: {:.5f}'.format(
epoch, epoch_num, global_step, logs, train_reader_cost /
print_batch_step, (train_reader_cost + train_run_cost) /
print_batch_step, total_samples / print_batch_step,
total_samples / (train_reader_cost + train_run_cost))
eta_sec = ((epoch_num + 1 - epoch) * \
len(train_dataloader) - idx - 1) * eta_meter.avg
eta_sec_format = str(datetime.timedelta(seconds=int(eta_sec)))
strs = 'epoch: [{}/{}], global_step: {}, {}, avg_reader_cost: ' \
'{:.5f} s, avg_batch_cost: {:.5f} s, avg_samples: {}, ' \
'ips: {:.5f} samples/s, eta: {}'.format(
epoch, epoch_num, global_step, logs,
train_reader_cost / print_batch_step,
train_batch_cost / print_batch_step,
total_samples / print_batch_step,
total_samples / train_batch_cost, eta_sec_format)
logger.info(strs)
train_reader_cost = 0.0
train_run_cost = 0.0
total_samples = 0
train_reader_cost = 0.0
train_batch_cost = 0.0
# eval
if global_step > start_eval_step and \
(global_step - start_eval_step) % eval_batch_step == 0 and dist.get_rank() == 0:
(global_step - start_eval_step) % eval_batch_step == 0 \
and dist.get_rank() == 0:
if model_average:
Model_Average = paddle.incubate.optimizer.ModelAverage(
0.15,
......@@ -499,14 +531,24 @@ def preprocess(is_train=False):
use_gpu = config['Global']['use_gpu']
check_gpu(use_gpu)
# check if set use_xpu=True in paddlepaddle cpu/gpu version
use_xpu = False
if 'use_xpu' in config['Global']:
use_xpu = config['Global']['use_xpu']
check_xpu(use_xpu)
alg = config['Architecture']['algorithm']
assert alg in [
'EAST', 'DB', 'SAST', 'Rosetta', 'CRNN', 'STARNet', 'RARE', 'SRN',
'CLS', 'PGNet', 'Distillation', 'NRTR', 'TableAttn', 'SAR', 'PSE',
'SEED', 'SDMGR', 'LayoutXLM', 'LayoutLM', 'FCE'
'SEED', 'SDMGR', 'LayoutXLM', 'LayoutLM', 'PREN', 'FCE'
]
device = 'gpu:{}'.format(dist.ParallelEnv().dev_id) if use_gpu else 'cpu'
device = 'cpu'
if use_gpu:
device = 'gpu:{}'.format(dist.ParallelEnv().dev_id)
if use_xpu:
device = 'xpu'
device = paddle.set_device(device)
config['Global']['distributed'] = dist.get_world_size() != 1
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或