merge upstream

PPOCRLabel/libs/shape.py

@@ -48,6 +48,7 @@ class Shape(object):
 
     def __init__(self, label=None, line_color=None, difficult=False, key_cls="None", paintLabel=False):
         self.label = label
+        self.idx = 0
         self.points = []
         self.fill = False
         self.selected = False

README.md

@@ -72,6 +72,7 @@ PaddleOCR support a variety of cutting-edge algorithms related to OCR, and devel
 <img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG"  width = "200" height = "200" />
 </div>
 <a name="Supported-Chinese-model-list"></a>
+
 ## PP-OCR Series Model List（Update on September 8th）
 
 | Model introduction                                           | Model name                   | Recommended scene | Detection model                                              | Direction classifier                                         | Recognition model                                            |

README_ch.md

@@ -71,6 +71,8 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
 ## 《动手学OCR》电子书
 - [《动手学OCR》电子书📚](./doc/doc_ch/ocr_book.md)
 
+## 场景应用
+- PaddleOCR场景应用覆盖通用，制造、金融、交通行业的主要OCR垂类应用，在PP-OCR、PP-Structure的通用能力基础之上，以notebook的形式展示利用场景数据微调、模型优化方法、数据增广等内容，为开发者快速落地OCR应用提供示范与启发。详情可查看[README](./applications)。
 
 <a name="开源社区"></a>
 ## 开源社区

applications/PCB字符识别/gen_data/corpus/text.txt

+5ZQ
+I4UL
+PWL
+SNOG
+ZL02
+1C30
+O3H
+YHRS
+N03S
+1U5Y
+JTK
+EN4F
+YKJ
+DWNH
+R42W
+X0V
+4OF5
+08AM
+Y93S
+GWE2
+0KR
+9U2A
+DBQ
+Y6J
+ROZ
+K06
+KIEY
+NZQJ
+UN1B
+6X4
\ No newline at end of file

applications/PCB字符识别/gen_data/gen.py

+# copyright (c) 2020 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.
+"""
+This code is refer from:
+https://github.com/zcswdt/Color_OCR_image_generator
+"""
+import os
+import random
+from PIL import Image, ImageDraw, ImageFont
+import json
+import argparse
+
+
+def get_char_lines(txt_root_path):
+    """
+    desc:get corpus line
+    """
+    txt_files = os.listdir(txt_root_path)
+    char_lines = []
+    for txt in txt_files:
+        f = open(os.path.join(txt_root_path, txt), mode='r', encoding='utf-8')
+        lines = f.readlines()
+        f.close()
+        for line in lines:
+            char_lines.append(line.strip())
+        return char_lines
+
+
+def get_horizontal_text_picture(image_file, chars, fonts_list, cf):
+    """
+    desc:gen horizontal text picture
+    """
+    img = Image.open(image_file)
+    if img.mode != 'RGB':
+        img = img.convert('RGB')
+    img_w, img_h = img.size
+
+    # random choice font
+    font_path = random.choice(fonts_list)
+    # random choice font size
+    font_size = random.randint(cf.font_min_size, cf.font_max_size)
+    font = ImageFont.truetype(font_path, font_size)
+
+    ch_w = []
+    ch_h = []
+    for ch in chars:
+        wt, ht = font.getsize(ch)
+        ch_w.append(wt)
+        ch_h.append(ht)
+    f_w = sum(ch_w)
+    f_h = max(ch_h)
+
+    # add space
+    char_space_width = max(ch_w)
+    f_w += (char_space_width * (len(chars) - 1))
+
+    x1 = random.randint(0, img_w - f_w)
+    y1 = random.randint(0, img_h - f_h)
+    x2 = x1 + f_w
+    y2 = y1 + f_h
+
+    crop_y1 = y1
+    crop_x1 = x1
+    crop_y2 = y2
+    crop_x2 = x2
+
+    best_color = (0, 0, 0)
+    draw = ImageDraw.Draw(img)
+    for i, ch in enumerate(chars):
+        draw.text((x1, y1), ch, best_color, font=font)
+        x1 += (ch_w[i] + char_space_width)
+    crop_img = img.crop((crop_x1, crop_y1, crop_x2, crop_y2))
+    return crop_img, chars
+
+
+def get_vertical_text_picture(image_file, chars, fonts_list, cf):
+    """
+    desc:gen vertical text picture
+    """
+    img = Image.open(image_file)
+    if img.mode != 'RGB':
+        img = img.convert('RGB')
+    img_w, img_h = img.size
+    # random choice font
+    font_path = random.choice(fonts_list)
+    # random choice font size
+    font_size = random.randint(cf.font_min_size, cf.font_max_size)
+    font = ImageFont.truetype(font_path, font_size)
+
+    ch_w = []
+    ch_h = []
+    for ch in chars:
+        wt, ht = font.getsize(ch)
+        ch_w.append(wt)
+        ch_h.append(ht)
+    f_w = max(ch_w)
+    f_h = sum(ch_h)
+
+    x1 = random.randint(0, img_w - f_w)
+    y1 = random.randint(0, img_h - f_h)
+    x2 = x1 + f_w
+    y2 = y1 + f_h
+
+    crop_y1 = y1
+    crop_x1 = x1
+    crop_y2 = y2
+    crop_x2 = x2
+
+    best_color = (0, 0, 0)
+    draw = ImageDraw.Draw(img)
+    i = 0
+    for ch in chars:
+        draw.text((x1, y1), ch, best_color, font=font)
+        y1 = y1 + ch_h[i]
+        i = i + 1
+    crop_img = img.crop((crop_x1, crop_y1, crop_x2, crop_y2))
+    crop_img = crop_img.transpose(Image.ROTATE_90)
+    return crop_img, chars
+
+
+def get_fonts(fonts_path):
+    """
+    desc: get all fonts
+    """
+    font_files = os.listdir(fonts_path)
+    fonts_list=[]
+    for font_file in font_files:
+        font_path=os.path.join(fonts_path, font_file)
+        fonts_list.append(font_path)
+    return fonts_list
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--num_img', type=int, default=30, help="Number of images to generate")
+    parser.add_argument('--font_min_size', type=int, default=11)
+    parser.add_argument('--font_max_size', type=int, default=12,
+                        help="Help adjust the size of the generated text and the size of the picture")
+    parser.add_argument('--bg_path', type=str, default='./background',
+                        help='The generated text pictures will be pasted onto the pictures of this folder')
+    parser.add_argument('--det_bg_path', type=str, default='./det_background',
+                        help='The generated text pictures will use the pictures of this folder as the background')
+    parser.add_argument('--fonts_path', type=str, default='../../StyleText/fonts',
+                        help='The font used to generate the picture')
+    parser.add_argument('--corpus_path', type=str, default='./corpus',
+                        help='The corpus used to generate the text picture')
+    parser.add_argument('--output_dir', type=str, default='./output/', help='Images save dir')
+
+
+    cf = parser.parse_args()
+    # save path
+    if not os.path.exists(cf.output_dir):
+        os.mkdir(cf.output_dir)
+
+    # get corpus
+    txt_root_path = cf.corpus_path
+    char_lines = get_char_lines(txt_root_path=txt_root_path)
+
+    # get all fonts
+    fonts_path = cf.fonts_path
+    fonts_list = get_fonts(fonts_path)
+
+    # rec bg
+    img_root_path = cf.bg_path
+    imnames=os.listdir(img_root_path)
+    
+    # det bg
+    det_bg_path = cf.det_bg_path
+    bg_pics = os.listdir(det_bg_path)
+
+    # OCR det files
+    det_val_file = open(cf.output_dir + 'det_gt_val.txt', 'w', encoding='utf-8')
+    det_train_file = open(cf.output_dir + 'det_gt_train.txt', 'w', encoding='utf-8')
+    # det imgs
+    det_save_dir = 'imgs/'
+    if not os.path.exists(cf.output_dir + det_save_dir):
+        os.mkdir(cf.output_dir + det_save_dir)
+    det_val_save_dir = 'imgs_val/'
+    if not os.path.exists(cf.output_dir + det_val_save_dir):
+        os.mkdir(cf.output_dir + det_val_save_dir)
+
+    # OCR rec files
+    rec_val_file = open(cf.output_dir + 'rec_gt_val.txt', 'w', encoding='utf-8')
+    rec_train_file = open(cf.output_dir + 'rec_gt_train.txt', 'w', encoding='utf-8')
+    # rec imgs
+    rec_save_dir = 'rec_imgs/'
+    if not os.path.exists(cf.output_dir + rec_save_dir):
+        os.mkdir(cf.output_dir + rec_save_dir)
+    rec_val_save_dir = 'rec_imgs_val/'
+    if not os.path.exists(cf.output_dir + rec_val_save_dir):
+        os.mkdir(cf.output_dir + rec_val_save_dir)
+
+
+    val_ratio = cf.num_img * 0.2  # val dataset ratio
+
+    print('start generating...')
+    for i in range(0, cf.num_img):
+        imname = random.choice(imnames)
+        img_path = os.path.join(img_root_path, imname)
+
+        rnd = random.random()
+        # gen horizontal text picture
+        if rnd < 0.5:
+            gen_img, chars = get_horizontal_text_picture(img_path, char_lines[i], fonts_list, cf)
+            ori_w, ori_h = gen_img.size
+            gen_img = gen_img.crop((0, 3, ori_w, ori_h))
+        # gen vertical text picture
+        else:
+            gen_img, chars = get_vertical_text_picture(img_path, char_lines[i], fonts_list, cf)
+            ori_w, ori_h = gen_img.size
+            gen_img = gen_img.crop((3, 0, ori_w, ori_h))
+
+        ori_w, ori_h = gen_img.size
+
+        # rec imgs
+        save_img_name = str(i).zfill(4) + '.jpg'
+        if i < val_ratio:
+            save_dir = os.path.join(rec_val_save_dir, save_img_name)
+            line = save_dir + '\t' + char_lines[i] + '\n'
+            rec_val_file.write(line)
+        else:
+            save_dir = os.path.join(rec_save_dir, save_img_name)
+            line = save_dir + '\t' + char_lines[i] + '\n'
+            rec_train_file.write(line)
+        gen_img.save(cf.output_dir + save_dir, quality = 95, subsampling=0)
+
+        # det img
+        # random choice bg
+        bg_pic = random.sample(bg_pics, 1)[0]
+        det_img = Image.open(os.path.join(det_bg_path, bg_pic))
+        # the PCB position is fixed, modify it according to your own scenario
+        if bg_pic == '1.png':
+            x1 = 38
+            y1 = 3
+        else:
+            x1 = 34
+            y1 = 1
+
+        det_img.paste(gen_img, (x1, y1))
+        # text pos
+        chars_pos = [[x1, y1], [x1 + ori_w, y1], [x1 + ori_w, y1 + ori_h], [x1, y1 + ori_h]]
+        label = [{"transcription":char_lines[i], "points":chars_pos}]
+        if i < val_ratio:
+            save_dir = os.path.join(det_val_save_dir, save_img_name)
+            det_val_file.write(save_dir + '\t' + json.dumps(
+                    label, ensure_ascii=False) + '\n')
+        else:
+            save_dir = os.path.join(det_save_dir, save_img_name)
+            det_train_file.write(save_dir + '\t' + json.dumps(
+                    label, ensure_ascii=False) + '\n')
+        det_img.save(cf.output_dir + save_dir, quality = 95, subsampling=0)

applications/README.md

+# 场景应用
+
+PaddleOCR场景应用覆盖通用，制造、金融、交通行业的主要OCR垂类应用，在PP-OCR、PP-Structure的通用能力基础之上，以notebook的形式展示利用场景数据微调、模型优化方法、数据增广等内容，为开发者快速落地OCR应用提供示范与启发。
+
+> 如需下载全部垂类模型，可以扫描下方二维码，关注公众号填写问卷后，加入PaddleOCR官方交流群获取20G OCR学习大礼包（内含《动手学OCR》电子书、课程回放视频、前沿论文等重磅资料）
+
+<div align="center">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+
+
+> 如果您是企业开发者且未在下述场景中找到合适的方案，可以填写[OCR应用合作调研问卷](https://paddle.wjx.cn/vj/QwF7GKw.aspx)，免费与官方团队展开不同层次的合作，包括但不限于问题抽象、确定技术方案、项目答疑、共同研发等。如果您已经使用PaddleOCR落地项目，也可以填写此问卷，与飞桨平台共同宣传推广，提升企业技术品宣。期待您的提交！
+
+## 通用
+
+| 类别                                              | 亮点     | 类别       | 亮点         |
+| ------------------------------------------------- | -------- | ---------- | ------------ |
+| [高精度中文识别模型SVTR](./高精度中文识别模型.md) | 新增模型 | 手写体识别 | 新增字形支持 |
+
+## 制造
+
+| 类别                                                         | 亮点                           | 类别                                        | 亮点                 |
+| ------------------------------------------------------------ | ------------------------------ | ------------------------------------------- | -------------------- |
+| [数码管识别](./光功率计数码管字符识别/光功率计数码管字符识别.md) | 数码管数据合成、漏识别调优     | 电表识别                                    | 大分辨率图像检测调优 |
+| [液晶屏读数识别](./液晶屏读数识别.md)                        | 检测模型蒸馏、Serving部署      | [PCB文字识别](./PCB字符识别/PCB字符识别.md) | 小尺寸文本检测与识别 |
+| [包装生产日期](./包装生产日期识别.md)                        | 点阵字符合成、过曝过暗文字识别 | 液晶屏缺陷检测                              | 非文字字符识别       |
+
+## 金融
+
+| 类别                           | 亮点                     | 类别         | 亮点                  |
+| ------------------------------ | ------------------------ | ------------ | --------------------- |
+| [表单VQA](./多模态表单识别.md) | 多模态通用表单结构化提取 | 通用卡证识别 | 通用结构化提取        |
+| 增值税发票                     | 尽请期待                 | 身份证识别   | 结构化提取、图像阴影  |
+| 印章检测与识别                 | 端到端弯曲文本识别       | 合同比对     | 密集文本检测、NLP串联 |
+
+## 交通
+
+| 类别                            | 亮点                           | 类别       | 亮点     |
+| ------------------------------- | ------------------------------ | ---------- | -------- |
+| [车牌识别](./轻量级车牌识别.md) | 多角度图像、轻量模型、端侧部署 | 快递单识别 | 尽请期待 |
+| 驾驶证/行驶证识别               | 尽请期待                       |            |          |
\ No newline at end of file

applications/光功率计数码管字符识别/corpus/digital.txt

+46.39
+40.08
+89.52
+-71.93
+23.19
+-81.02
+-34.09
+05.87
+-67.80
+-51.56
+-34.58
+37.91
+56.98
+29.01
+-90.13
+35.55
+66.07
+-90.35
+-50.93
+42.42
+21.40
+-30.99
+-71.78
+25.60
+-48.69
+-72.28
+-17.55
+-99.93
+-47.35
+-64.89
+-31.28
+-90.01
+05.17
+30.91
+30.56
+-06.90
+79.05
+67.74
+-32.31
+94.22
+28.75
+51.03
+-58.96

applications/光功率计数码管字符识别/光功率计数码管字符识别.md

+# 光功率计数码管字符识别
+
+本案例将使用OCR技术自动识别光功率计显示屏文字，通过本章您可以掌握：
+
+- PaddleOCR快速使用
+- 数据合成方法
+- 数据挖掘方法
+- 基于现有数据微调
+
+## 1. 背景介绍
+
+光功率计（optical power meter ）是指用于测量绝对光功率或通过一段光纤的光功率相对损耗的仪器。在光纤系统中，测量光功率是最基本的，非常像电子学中的万用表；在光纤测量中，光功率计是重负荷常用表。
+
+<img src="https://bkimg.cdn.bcebos.com/pic/a08b87d6277f9e2f999f5e3e1c30e924b899f35a?x-bce-process=image/watermark,image_d2F0ZXIvYmFpa2U5Mg==,g_7,xp_5,yp_5/format,f_auto" width="400">
+
+目前光功率计缺少将数据直接输出的功能，需要人工读数。这一项工作单调重复，如果可以使用机器替代人工，将节约大量成本。针对上述问题，希望通过摄像头拍照->智能读数的方式高效地完成此任务。
+
+为实现智能读数，通常会采取文本检测+文本识别的方案：
+
+第一步，使用文本检测模型定位出光功率计中的数字部分；
+
+第二步，使用文本识别模型获得准确的数字和单位信息。
+
+本项目主要介绍如何完成第二步文本识别部分，包括：真实评估集的建立、训练数据的合成、基于 PP-OCRv3 和 SVTR_Tiny 两个模型进行训练，以及评估和推理。
+
+本项目难点如下：
+
+- 光功率计数码管字符数据较少，难以获取。
+- 数码管中小数点占像素较少，容易漏识别。
+
+针对以上问题， 本例选用 PP-OCRv3 和 SVTR_Tiny 两个高精度模型训练，同时提供了真实数据挖掘案例和数据合成案例。基于 PP-OCRv3 模型，在构建的真实评估集上精度从 52% 提升至 72%，SVTR_Tiny 模型精度可达到 78.9%。
+
+aistudio项目链接: [光功率计数码管字符识别](https://aistudio.baidu.com/aistudio/projectdetail/4049044?contributionType=1)
+
+## 2. PaddleOCR 快速使用
+
+PaddleOCR 旨在打造一套丰富、领先、且实用的OCR工具库，助力开发者训练出更好的模型，并应用落地。
+
+![](https://github.com/PaddlePaddle/PaddleOCR/raw/release/2.5/doc/imgs_results/ch_ppocr_mobile_v2.0/test_add_91.jpg)
+
+
+官方提供了适用于通用场景的高精轻量模型，首先使用官方提供的 PP-OCRv3 模型预测图片，验证下当前模型在光功率计场景上的效果。
+
+- 准备环境
+
+```
+python3 -m pip install -U pip
+python3 -m pip install paddleocr
+```
+
+
+- 测试效果
+
+测试图：
+
+![](https://ai-studio-static-online.cdn.bcebos.com/8dca91f016884e16ad9216d416da72ea08190f97d87b4be883f15079b7ebab9a)
+
+
+```
+paddleocr --lang=ch --det=Fase --image_dir=data
+```
+
+得到如下测试结果：
+
+```
+('.7000', 0.6885431408882141)
+```
+
+发现数字识别较准，然而对负号和小数点识别不准确。 由于PP-OCRv3的训练数据大多为通用场景数据，在特定的场景上效果可能不够好。因此需要基于场景数据进行微调。
+
+下面就主要介绍如何在光功率计（数码管）场景上微调训练。
+
+
+## 3. 开始训练
+
+### 3.1 数据准备
+
+特定的工业场景往往很难获取开源的真实数据集，光功率计也是如此。在实际工业场景中，可以通过摄像头采集的方法收集大量真实数据，本例中重点介绍数据合成方法和真实数据挖掘方法，如何利用有限的数据优化模型精度。
+
+数据集分为两个部分：合成数据，真实数据， 其中合成数据由 text_renderer 工具批量生成得到， 真实数据通过爬虫等方式在百度图片中搜索并使用 PPOCRLabel 标注得到。
+
+
+- 合成数据
+
+本例中数据合成工具使用的是 [text_renderer](https://github.com/Sanster/text_renderer)， 该工具可以合成用于文本识别训练的文本行数据:
+
+![](https://github.com/oh-my-ocr/text_renderer/raw/master/example_data/effect_layout_image/char_spacing_compact.jpg)
+
+![](https://github.com/oh-my-ocr/text_renderer/raw/master/example_data/effect_layout_image/color_image.jpg)
+
+
+```
+export https_proxy=http://172.19.57.45:3128
+git clone https://github.com/oh-my-ocr/text_renderer
+```
+
+```
+import os
+python3 setup.py develop
+python3 -m pip install -r docker/requirements.txt
+python3 main.py \
+    --config example_data/example.py \
+    --dataset img \
+    --num_processes 2 \
+    --log_period 10
+```
+
+给定字体和语料，就可以合成较为丰富样式的文本行数据。 光功率计识别场景，目标是正确识别数码管文本，因此需要收集部分数码管字体，训练语料，用于合成文本识别数据。
+
+将收集好的语料存放在 example_data 路径下：
+
+```
+ln -s ./fonts/DS* text_renderer/example_data/font/
+ln -s ./corpus/digital.txt text_renderer/example_data/text/
+```
+
+修改 text_renderer/example_data/font_list/font_list.txt ,选择需要的字体开始合成：
+
+```
+python3 main.py \
+    --config example_data/digital_example.py \
+    --dataset img \
+    --num_processes 2 \
+    --log_period 10
+```
+
+合成图片会被存在目录 text_renderer/example_data/digital/chn_data 下
+
+查看合成的数据样例：
+
+![](https://ai-studio-static-online.cdn.bcebos.com/7d5774a273f84efba5b9ce7fd3f86e9ef24b6473e046444db69fa3ca20ac0986)
+
+
+- 真实数据挖掘
+
+模型训练需要使用真实数据作为评价指标，否则很容易过拟合到简单的合成数据中。没有开源数据的情况下，可以利用部分无标注数据+标注工具获得真实数据。
+
+
+1. 数据搜集
+
+使用[爬虫工具](https://github.com/Joeclinton1/google-images-download.git)获得无标注数据
+
+2. [PPOCRLabel](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.5/PPOCRLabel) 完成半自动标注
+
+PPOCRLabel是一款适用于OCR领域的半自动化图形标注工具，内置PP-OCR模型对数据自动标注和重新识别。使用Python3和PyQT5编写，支持矩形框标注、表格标注、不规则文本标注、关键信息标注模式，导出格式可直接用于PaddleOCR检测和识别模型的训练。
+
+![](https://github.com/PaddlePaddle/PaddleOCR/raw/release/2.5/PPOCRLabel/data/gif/steps_en.gif)
+
+
+收集完数据后就可以进行分配了，验证集中一般都是真实数据，训练集中包含合成数据+真实数据。本例中标注了155张图片，其中训练集和验证集的数目为100和55。
+
+
+最终 `data` 文件夹应包含以下几部分：
+
+```
+|-data
+  |- synth_train.txt
+  |- real_train.txt
+  |- real_eval.txt
+  |- synthetic_data
+      |- word_001.png
+      |- word_002.jpg
+      |- word_003.jpg
+      | ...
+  |- real_data
+      |- word_001.png
+      |- word_002.jpg
+      |- word_003.jpg
+      | ...
+  ...
+```
+
+### 3.2 模型选择
+
+本案例提供了2种文本识别模型：PP-OCRv3 识别模型 和 SVTR_Tiny：
+
+[PP-OCRv3 识别模型](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_ch/PP-OCRv3_introduction.md)：PP-OCRv3的识别模块是基于文本识别算法SVTR优化。SVTR不再采用RNN结构，通过引入Transformers结构更加有效地挖掘文本行图像的上下文信息，从而提升文本识别能力。并进行了一系列结构改进加速模型预测。
+
+[SVTR_Tiny](https://arxiv.org/abs/2205.00159):SVTR提出了一种用于场景文本识别的单视觉模型，该模型在patch-wise image tokenization框架内，完全摒弃了序列建模，在精度具有竞争力的前提下，模型参数量更少，速度更快。
+
+以上两个策略在自建中文数据集上的精度和速度对比如下：
+
+| ID | 策略 |  模型大小 | 精度 | 预测耗时（CPU + MKLDNN)|
+|-----|-----|--------|----| --- |
+| 01 | PP-OCRv2 | 8M | 74.8% | 8.54ms |
+| 02 | SVTR_Tiny | 21M | 80.1% | 97ms |
+| 03 | SVTR_LCNet(h32) | 12M | 71.9% | 6.6ms |
+| 04 | SVTR_LCNet(h48) | 12M | 73.98% | 7.6ms |
+| 05 | + GTC | 12M | 75.8% | 7.6ms |
+| 06 | + TextConAug | 12M | 76.3% | 7.6ms |
+| 07 | + TextRotNet | 12M | 76.9% | 7.6ms |
+| 08 | + UDML | 12M | 78.4% | 7.6ms |
+| 09 | + UIM | 12M | 79.4% | 7.6ms |
+
+
+### 3.3 开始训练
+
+首先下载 PaddleOCR 代码库
+
+```
+git clone -b release/2.5 https://github.com/PaddlePaddle/PaddleOCR.git
+```
+
+PaddleOCR提供了训练脚本、评估脚本和预测脚本，本节将以 PP-OCRv3 中文识别模型为例：
+
+**Step1：下载预训练模型**
+
+首先下载 pretrain model，您可以下载训练好的模型在自定义数据上进行finetune
+
+```
+cd PaddleOCR/
+# 下载PP-OCRv3 中文预训练模型
+wget -P ./pretrain_models/ https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_train.tar
+# 解压模型参数
+cd pretrain_models
+tar -xf ch_PP-OCRv3_rec_train.tar && rm -rf ch_PP-OCRv3_rec_train.tar
+```
+
+**Step2：自定义字典文件**
+
+接下来需要提供一个字典（{word_dict_name}.txt），使模型在训练时，可以将所有出现的字符映射为字典的索引。
+
+因此字典需要包含所有希望被正确识别的字符，{word_dict_name}.txt需要写成如下格式，并以 `utf-8` 编码格式保存：
+
+```
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+-
+.
+```
+
+word_dict.txt 每行有一个单字，将字符与数字索引映射在一起，“3.14” 将被映射成 [3, 11, 1, 4]
+
+* 内置字典
+
+PaddleOCR内置了一部分字典，可以按需使用。
+
+`ppocr/utils/ppocr_keys_v1.txt` 是一个包含6623个字符的中文字典
+
+`ppocr/utils/ic15_dict.txt` 是一个包含36个字符的英文字典
+
+* 自定义字典
+
+内置字典面向通用场景，具体的工业场景中，可能需要识别特殊字符，或者只需识别某几个字符，此时自定义字典会更提升模型精度。例如在光功率计场景中，需要识别数字和单位。
+
+遍历真实数据标签中的字符，制作字典`digital_dict.txt`如下所示：
+
+```
+-
+.
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+B
+E
+F
+H
+L
+N
+T
+W
+d
+k
+m
+n
+o
+z
+```
+
+
+
+
+**Step3：修改配置文件**
+
+为了更好的使用预训练模型，训练推荐使用[ch_PP-OCRv3_rec_distillation.yml](../../configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml)配置文件，并参考下列说明修改配置文件：
+
+以 `ch_PP-OCRv3_rec_distillation.yml` 为例：
+```
+Global:
+  ...
+  # 添加自定义字典，如修改字典请将路径指向新字典
+  character_dict_path: ppocr/utils/dict/digital_dict.txt
+  ...
+  # 识别空格
+  use_space_char: True
+
+
+Optimizer:
+  ...
+  # 添加学习率衰减策略
+  lr:
+    name: Cosine
+    learning_rate: 0.001
+  ...
+
+...
+
+Train:
+  dataset:
+    # 数据集格式，支持LMDBDataSet以及SimpleDataSet
+    name: SimpleDataSet
+    # 数据集路径
+    data_dir: ./data/
+    # 训练集标签文件
+    label_file_list:
+    - ./train_data/digital_img/digital_train.txt  #11w
+    - ./train_data/digital_img/real_train.txt     #100
+    - ./train_data/digital_img/dbm_img/dbm.txt    #3w
+    ratio_list:
+    - 0.3
+    - 1.0
+    - 1.0
+    transforms:
+      ...
+      - RecResizeImg:
+          # 修改 image_shape 以适应长文本
+          image_shape: [3, 48, 320]
+      ...
+  loader:
+    ...
+    # 单卡训练的batch_size
+    batch_size_per_card: 256
+    ...
+
+Eval:
+  dataset:
+    # 数据集格式，支持LMDBDataSet以及SimpleDataSet
+    name: SimpleDataSet
+    # 数据集路径
+    data_dir: ./data
+    # 验证集标签文件
+    label_file_list:
+    - ./train_data/digital_img/real_val.txt
+    transforms:
+      ...
+      - RecResizeImg:
+          # 修改 image_shape 以适应长文本
+          image_shape: [3, 48, 320]
+      ...
+  loader:
+    # 单卡验证的batch_size
+    batch_size_per_card: 256
+    ...
+```
+**注意，训练/预测/评估时的配置文件请务必与训练一致。**
+
+**Step4：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*
+
+```
+# GPU训练 支持单卡，多卡训练
+# 训练数码管数据 训练日志会自动保存为 "{save_model_dir}" 下的train.log
+
+#单卡训练（训练周期长，不建议）
+python3 tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model=./pretrain_models/ch_PP-OCRv3_rec_train/best_accuracy
+
+#多卡训练，通过--gpus参数指定卡号
+python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model=./pretrain_models/en_PP-OCRv3_rec_train/best_accuracy
+```
+
+
+PaddleOCR支持训练和评估交替进行, 可以在 `configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml` 中修改 `eval_batch_step` 设置评估频率，默认每500个iter评估一次。评估过程中默认将最佳acc模型，保存为 `output/ch_PP-OCRv3_rec_distill/best_accuracy` 。
+
+如果验证集很大，测试将会比较耗时，建议减少评估次数，或训练完再进行评估。
+
+### SVTR_Tiny 训练
+
+SVTR_Tiny 训练步骤与上面一致，SVTR支持的配置和模型训练权重可以参考[算法介绍文档](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_ch/algorithm_rec_svtr.md)
+
+**Step1：下载预训练模型**
+
+```
+# 下载 SVTR_Tiny 中文识别预训练模型和配置文件
+wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/rec_svtr_tiny_none_ctc_ch_train.tar
+# 解压模型参数
+tar -xf rec_svtr_tiny_none_ctc_ch_train.tar && rm -rf rec_svtr_tiny_none_ctc_ch_train.tar
+```
+**Step2：自定义字典文件**
+
+字典依然使用自定义的 digital_dict.txt
+
+**Step3：修改配置文件**
+
+配置文件中对应修改字典路径和数据路径
+
+**Step4：启动训练**
+
+```
+## 单卡训练
+python tools/train.py -c rec_svtr_tiny_none_ctc_ch_train/rec_svtr_tiny_6local_6global_stn_ch.yml \
+           -o Global.pretrained_model=./rec_svtr_tiny_none_ctc_ch_train/best_accuracy
+```
+
+### 3.4 验证效果
+
+如需获取已训练模型，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
+<div align="left">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+将下载或训练完成的模型放置在对应目录下即可完成模型推理
+
+* 指标评估
+
+训练中模型参数默认保存在`Global.save_model_dir`目录下。在评估指标时，需要设置`Global.checkpoints`指向保存的参数文件。评估数据集可以通过 `configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml`  修改Eval中的 `label_file_path` 设置。
+
+```
+# GPU 评估， Global.checkpoints 为待测权重
+python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.checkpoints={path/to/weights}/best_accuracy
+```
+
+* 测试识别效果
+
+使用 PaddleOCR 训练好的模型，可以通过以下脚本进行快速预测。
+
+默认预测图片存储在 `infer_img` 里，通过 `-o Global.checkpoints` 加载训练好的参数文件：
+
+根据配置文件中设置的 `save_model_dir` 和 `save_epoch_step` 字段，会有以下几种参数被保存下来：
+
+```
+output/rec/
+├── best_accuracy.pdopt  
+├── best_accuracy.pdparams  
+├── best_accuracy.states  
+├── config.yml  
+├── iter_epoch_3.pdopt  
+├── iter_epoch_3.pdparams  
+├── iter_epoch_3.states  
+├── latest.pdopt  
+├── latest.pdparams  
+├── latest.states  
+└── train.log
+```
+
+其中 best_accuracy.* 是评估集上的最优模型；iter_epoch_x.* 是以 `save_epoch_step` 为间隔保存下来的模型；latest.* 是最后一个epoch的模型。
+
+```
+# 预测英文结果
+python3 tools/infer_rec.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model={path/to/weights}/best_accuracy  Global.infer_img=test_digital.png
+```
+
+预测图片：
+
+![](https://ai-studio-static-online.cdn.bcebos.com/8dca91f016884e16ad9216d416da72ea08190f97d87b4be883f15079b7ebab9a)
+
+
+得到输入图像的预测结果：
+
+```
+infer_img: test_digital.png
+        result: ('-70.00', 0.9998967)
+```

applications/轻量级车牌识别.md

@@ -249,7 +249,7 @@ tar -xf ch_PP-OCRv3_det_distill_train.tar
 cd /home/aistudio/PaddleOCR
 ```
 
-预训练模型下载完成后，我们使用[ch_PP-OCRv3_det_student.yml](../configs/chepai/ch_PP-OCRv3_det_student.yml) 配置文件进行后续实验，在开始评估之前需要对配置文件中部分字段进行设置，具体如下：
+预训练模型下载完成后，我们使用[ch_PP-OCRv3_det_student.yml](../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_student.yml) 配置文件进行后续实验，在开始评估之前需要对配置文件中部分字段进行设置，具体如下：
 
 1. 模型存储和训练相关:
    1. Global.pretrained_model: 指向PP-OCRv3文本检测预训练模型地址
@@ -311,7 +311,6 @@ python tools/train.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_student.yml -o
 
 在上述命令中，通过`-o`的方式修改了配置文件中的参数。
 
-训练好的模型地址为: [det_ppocr_v3_finetune.tar](https://paddleocr.bj.bcebos.com/fanliku/license_plate_recognition/det_ppocr_v3_finetune.tar)
 
 **评估**
 
@@ -354,8 +353,6 @@ python3.7 deploy/slim/quantization/quant.py -c configs/det/ch_PP-OCRv3/ch_PP-OCR
     Eval.dataset.label_file_list=[/home/aistudio/data/CCPD2020/PPOCR/test/det.txt]
 ```
 
-训练好的模型地址为: [det_ppocr_v3_quant.tar](https://paddleocr.bj.bcebos.com/fanliku/license_plate_recognition/det_ppocr_v3_quant.tar)
-
 量化后指标对比如下
 
 |方案|hmeans| 模型大小 | 预测速度(lite) |
@@ -436,6 +433,12 @@ python tools/eval.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec.yml -o \
     Eval.dataset.label_file_list=[/home/aistudio/data/CCPD2020/PPOCR/test/rec.txt]
 ```
 
+如需获取已训练模型，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
+<div align="left">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+
+
 评估部分日志如下：
 ```bash
 [2022/05/12 19:52:02] ppocr INFO: load pretrain successful from models/ch_PP-OCRv3_rec_train/best_accuracy
@@ -528,7 +531,6 @@ python tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec.yml -o \
     Eval.dataset.data_dir=/home/aistudio/data/CCPD2020/PPOCR \
     Eval.dataset.label_file_list=[/home/aistudio/data/CCPD2020/PPOCR/test/rec.txt]
 ```
-训练好的模型地址为: [rec_ppocr_v3_finetune.tar](https://paddleocr.bj.bcebos.com/fanliku/license_plate_recognition/rec_ppocr_v3_finetune.tar)
 
 **评估**
 
@@ -570,7 +572,6 @@ python3.7 deploy/slim/quantization/quant.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_
     Eval.dataset.data_dir=/home/aistudio/data/CCPD2020/PPOCR \
     Eval.dataset.label_file_list=[/home/aistudio/data/CCPD2020/PPOCR/test/rec.txt]
 ```
-训练好的模型地址为: [rec_ppocr_v3_quant.tar](https://paddleocr.bj.bcebos.com/fanliku/license_plate_recognition/rec_ppocr_v3_quant.tar)
 
 量化后指标对比如下
 
@@ -787,12 +788,12 @@ python tools/infer/predict_system.py \
 
 - 端侧部署
 
-端侧部署我们采用基于 PaddleLite 的 cpp 推理。Paddle Lite是飞桨轻量化推理引擎，为手机、IOT端提供高效推理能力，并广泛整合跨平台硬件，为端侧部署及应用落地问题提供轻量化的部署方案。具体可参考 [PaddleOCR lite教程](../dygraph/deploy/lite/readme_ch.md)
+端侧部署我们采用基于 PaddleLite 的 cpp 推理。Paddle Lite是飞桨轻量化推理引擎，为手机、IOT端提供高效推理能力，并广泛整合跨平台硬件，为端侧部署及应用落地问题提供轻量化的部署方案。具体可参考 [PaddleOCR lite教程](../deploy/lite/readme_ch.md)
 
 
 ### 4.5 实验总结
 
-我们分别使用PP-OCRv3中英文超轻量预训练模型在车牌数据集上进行了直接评估和 fine-tune 和 fine-tune +量化3种方案的实验，并基于[PaddleOCR lite教程](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/deploy/lite/readme_ch.md)进行了速度测试，指标对比如下：
+我们分别使用PP-OCRv3中英文超轻量预训练模型在车牌数据集上进行了直接评估和 fine-tune 和 fine-tune +量化3种方案的实验，并基于[PaddleOCR lite教程](../deploy/lite/readme_ch.md)进行了速度测试，指标对比如下：
 
 - 检测
 

applications/高精度中文识别模型.md

+# 高精度中文场景文本识别模型SVTR
+
+## 1. 简介
+
+PP-OCRv3是百度开源的超轻量级场景文本检测识别模型库，其中超轻量的场景中文识别模型SVTR_LCNet使用了SVTR算法结构。为了保证速度，SVTR_LCNet将SVTR模型的Local Blocks替换为LCNet，使用两层Global Blocks。在中文场景中，PP-OCRv3识别主要使用如下优化策略：
+- GTC：Attention指导CTC训练策略；
+- TextConAug：挖掘文字上下文信息的数据增广策略；
+- TextRotNet：自监督的预训练模型；
+- UDML：联合互学习策略；
+- UIM：无标注数据挖掘方案。
+
+其中 *UIM：无标注数据挖掘方案* 使用了高精度的SVTR中文模型进行无标注文件的刷库，该模型在PP-OCRv3识别的数据集上训练，精度对比如下表。
+
+|中文识别算法|模型|UIM|精度|
+| --- | --- | --- |--- |
+|PP-OCRv3|SVTR_LCNet| w/o |78.4%|
+|PP-OCRv3|SVTR_LCNet| w |79.4%|
+|SVTR|SVTR-Tiny|-|82.5%|
+
+aistudio项目链接: [高精度中文场景文本识别模型SVTR](https://aistudio.baidu.com/aistudio/projectdetail/4263032)
+
+## 2. SVTR中文模型使用
+
+### 环境准备
+
+
+本任务基于Aistudio完成, 具体环境如下：
+
+- 操作系统: Linux
+- PaddlePaddle: 2.3
+- PaddleOCR: dygraph
+
+下载 PaddleOCR代码
+
+```bash
+git clone -b dygraph https://github.com/PaddlePaddle/PaddleOCR
+```
+
+安装依赖库
+
+```bash
+pip install -r PaddleOCR/requirements.txt -i https://mirror.baidu.com/pypi/simple
+```
+
+### 快速使用
+
+获取SVTR中文模型文件，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
+<div align="center">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+
+```bash
+# 解压模型文件
+tar xf svtr_ch_high_accuracy.tar
+```
+
+预测中文文本，以下图为例：
+![](../doc/imgs_words/ch/word_1.jpg)
+
+预测命令：
+
+```bash
+# CPU预测
+python tools/infer_rec.py -c configs/rec/rec_svtrnet_ch.yml -o Global.pretrained_model=./svtr_ch_high_accuracy/best_accuracy Global.infer_img=./doc/imgs_words/ch/word_1.jpg Global.use_gpu=False
+
+# GPU预测
+#python tools/infer_rec.py -c configs/rec/rec_svtrnet_ch.yml -o Global.pretrained_model=./svtr_ch_high_accuracy/best_accuracy Global.infer_img=./doc/imgs_words/ch/word_1.jpg Global.use_gpu=True
+```
+
+可以看到最后打印结果为
+- result: 韩国小馆    0.9853458404541016
+
+0.9853458404541016为预测置信度。
+
+### 推理模型导出与预测
+
+inference 模型（paddle.jit.save保存的模型） 一般是模型训练，把模型结构和模型参数保存在文件中的固化模型，多用于预测部署场景。 训练过程中保存的模型是checkpoints模型，保存的只有模型的参数，多用于恢复训练等。 与checkpoints模型相比，inference 模型会额外保存模型的结构信息，在预测部署、加速推理上性能优越，灵活方便，适合于实际系统集成。
+
+运行识别模型转inference模型命令，如下：
+
+```bash
+python tools/export_model.py -c configs/rec/rec_svtrnet_ch.yml -o Global.pretrained_model=./svtr_ch_high_accuracy/best_accuracy Global.save_inference_dir=./inference/svtr_ch
+```
+
+转换成功后，在目录下有三个文件：
+```shell
+inference/svtr_ch/
+    ├── inference.pdiparams         # 识别inference模型的参数文件
+    ├── inference.pdiparams.info    # 识别inference模型的参数信息，可忽略
+    └── inference.pdmodel           # 识别inference模型的program文件
+```
+
+inference模型预测，命令如下：
+
+```bash
+# CPU预测
+python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words/ch/word_1.jpg" --rec_algorithm='SVTR' --rec_model_dir=./inference/svtr_ch/ --rec_image_shape='3, 32, 320'  --rec_char_dict_path=ppocr/utils/ppocr_keys_v1.txt --use_gpu=False
+
+# GPU预测
+#python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words/ch/word_1.jpg" --rec_algorithm='SVTR' --rec_model_dir=./inference/svtr_ch/ --rec_image_shape='3, 32, 320'  --rec_char_dict_path=ppocr/utils/ppocr_keys_v1.txt --use_gpu=True
+```
+
+**注意**
+
+- 使用SVTR算法时，需要指定--rec_algorithm='SVTR'
+- 如果使用自定义字典训练的模型，需要将--rec_char_dict_path=ppocr/utils/ppocr_keys_v1.txt修改为自定义的字典
+- --rec_image_shape='3, 32, 320' 该参数不能去掉

configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml

@@ -28,7 +28,7 @@ Architecture:
       algorithm: DB
       Transform:
       Backbone:
-        name: ResNet
+        name: ResNet_vd
         layers: 18
       Neck:
         name: DBFPN

configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_distill.yml

@@ -45,7 +45,7 @@ Architecture:
       algorithm: DB
       Transform:
       Backbone:
-        name: ResNet
+        name: ResNet_vd
         layers: 18
       Neck:
         name: DBFPN

configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_dml.yml

@@ -65,7 +65,7 @@ Loss:
       - ["Student", "Teacher"]
       maps_name: "thrink_maps"
       weight: 1.0
-      act: "softmax"
+      # act: None
       model_name_pairs: ["Student", "Teacher"]
       key: maps
   - DistillationDBLoss:

configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml

@@ -61,7 +61,7 @@ Architecture:
       model_type: det
       algorithm: DB
       Backbone:
-        name: ResNet
+        name: ResNet_vd
         in_channels: 3
         layers: 50
       Neck:

configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_dml.yml

@@ -25,7 +25,7 @@ Architecture:
       model_type: det
       algorithm: DB
       Backbone:
-        name: ResNet
+        name: ResNet_vd
         in_channels: 3
         layers: 50
       Neck:
@@ -40,7 +40,7 @@ Architecture:
       model_type: det
       algorithm: DB
       Backbone:
-        name: ResNet
+        name: ResNet_vd
         in_channels: 3
         layers: 50
       Neck:
@@ -60,7 +60,7 @@ Loss:
       - ["Student", "Student2"]
       maps_name: "thrink_maps"
       weight: 1.0
-      act: "softmax"
+      # act: None
       model_name_pairs: ["Student", "Student2"]
       key: maps
   - DistillationDBLoss:

configs/det/ch_ppocr_v2.0/ch_det_res18_db_v2.0.yml

@@ -20,7 +20,7 @@ Architecture:
   algorithm: DB
   Transform:
   Backbone:
-    name: ResNet
+    name: ResNet_vd
     layers: 18
     disable_se: True
   Neck:

configs/det/det_r50_db++_ic15.yml

+Global:
+  debug: false
+  use_gpu: true
+  epoch_num: 1000
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/det_r50_icdar15/
+  save_epoch_step: 200
+  eval_batch_step:
+  - 0
+  - 2000
+  cal_metric_during_train: false
+  pretrained_model: ./pretrain_models/ResNet50_dcn_asf_synthtext_pretrained
+  checkpoints: null
+  save_inference_dir: null
+  use_visualdl: false
+  infer_img: doc/imgs_en/img_10.jpg
+  save_res_path: ./checkpoints/det_db/predicts_db.txt
+Architecture:
+  model_type: det
+  algorithm: DB++
+  Transform: null
+  Backbone:
+    name: ResNet
+    layers: 50
+    dcn_stage: [False, True, True, True]
+  Neck:
+    name: DBFPN
+    out_channels: 256
+    use_asf: True
+  Head:
+    name: DBHead
+    k: 50
+Loss:
+  name: DBLoss
+  balance_loss: true
+  main_loss_type: BCELoss
+  alpha: 5
+  beta: 10
+  ohem_ratio: 3
+Optimizer:
+  name: Momentum
+  momentum: 0.9
+  lr:
+    name: DecayLearningRate
+    learning_rate: 0.007
+    epochs: 1000
+    factor: 0.9
+    end_lr: 0
+  weight_decay: 0.0001
+PostProcess:
+  name: DBPostProcess
+  thresh: 0.3
+  box_thresh: 0.6
+  max_candidates: 1000
+  unclip_ratio: 1.5
+Metric:
+  name: DetMetric
+  main_indicator: hmean
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/icdar2015/text_localization/                                                                        
+    label_file_list:                                                                                            
+    - ./train_data/icdar2015/text_localization/train_icdar2015_label.txt                                                     
+    ratio_list:                                                                                                 
+    - 1.0 
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - DetLabelEncode: null
+    - IaaAugment:
+        augmenter_args:
+        - type: Fliplr
+          args:
+            p: 0.5
+        - type: Affine
+          args:
+            rotate:
+            - -10
+            - 10
+        - type: Resize
+          args:
+            size:
+            - 0.5
+            - 3
+    - EastRandomCropData:
+        size:
+        - 640
+        - 640
+        max_tries: 10
+        keep_ratio: true
+    - MakeShrinkMap:
+        shrink_ratio: 0.4
+        min_text_size: 8
+    - MakeBorderMap:
+        shrink_ratio: 0.4
+        thresh_min: 0.3
+        thresh_max: 0.7
+    - NormalizeImage:
+        scale: 1./255.
+        mean:
+        - 0.48109378172549
+        - 0.45752457890196
+        - 0.40787054090196
+        std:
+        - 1.0
+        - 1.0
+        - 1.0
+        order: hwc
+    - ToCHWImage: null
+    - KeepKeys:
+        keep_keys:
+        - image
+        - threshold_map
+        - threshold_mask
+        - shrink_map
+        - shrink_mask
+  loader:
+    shuffle: true
+    drop_last: false
+    batch_size_per_card: 4
+    num_workers: 8
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/icdar2015/text_localization                                                                          
+    label_file_list:                                                                                            
+    - ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - DetLabelEncode: null
+    - DetResizeForTest:
+        image_shape:
+        - 1152
+        - 2048
+    - NormalizeImage:
+        scale: 1./255.
+        mean:
+        - 0.48109378172549
+        - 0.45752457890196
+        - 0.40787054090196
+        std:
+        - 1.0
+        - 1.0
+        - 1.0
+        order: hwc
+    - ToCHWImage: null
+    - KeepKeys:
+        keep_keys:
+        - image
+        - shape
+        - polys
+        - ignore_tags
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 1
+    num_workers: 2
+profiler_options: null

configs/det/det_r50_db++_td_tr.yml

+Global:
+  debug: false
+  use_gpu: true
+  epoch_num: 1000
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/det_r50_td_tr/
+  save_epoch_step: 200
+  eval_batch_step:
+  - 0
+  - 2000
+  cal_metric_during_train: false
+  pretrained_model: ./pretrain_models/ResNet50_dcn_asf_synthtext_pretrained
+  checkpoints: null
+  save_inference_dir: null
+  use_visualdl: false
+  infer_img: doc/imgs_en/img_10.jpg
+  save_res_path: ./checkpoints/det_db/predicts_db.txt
+Architecture:
+  model_type: det
+  algorithm: DB++
+  Transform: null
+  Backbone:
+    name: ResNet
+    layers: 50
+    dcn_stage: [False, True, True, True]
+  Neck:
+    name: DBFPN
+    out_channels: 256
+    use_asf: True
+  Head:
+    name: DBHead
+    k: 50
+Loss:
+  name: DBLoss
+  balance_loss: true
+  main_loss_type: BCELoss
+  alpha: 5
+  beta: 10
+  ohem_ratio: 3
+Optimizer:
+  name: Momentum
+  momentum: 0.9
+  lr:
+    name: DecayLearningRate
+    learning_rate: 0.007
+    epochs: 1000
+    factor: 0.9
+    end_lr: 0
+  weight_decay: 0.0001
+PostProcess:
+  name: DBPostProcess
+  thresh: 0.3
+  box_thresh: 0.5
+  max_candidates: 1000
+  unclip_ratio: 1.5
+Metric:
+  name: DetMetric
+  main_indicator: hmean
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/                                                                                 
+    label_file_list:                                                                                            
+    - ./train_data/TD_TR/TD500/train_gt_labels.txt                                                           
+    - ./train_data/TD_TR/TR400/gt_labels.txt                                                                 
+    ratio_list:                                                                                                 
+    - 1.0                                                                                                       
+    - 1.0 
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - DetLabelEncode: null
+    - IaaAugment:
+        augmenter_args:
+        - type: Fliplr
+          args:
+            p: 0.5
+        - type: Affine
+          args:
+            rotate:
+            - -10
+            - 10
+        - type: Resize
+          args:
+            size:
+            - 0.5
+            - 3
+    - EastRandomCropData:
+        size:
+        - 640
+        - 640
+        max_tries: 10
+        keep_ratio: true
+    - MakeShrinkMap:
+        shrink_ratio: 0.4
+        min_text_size: 8
+    - MakeBorderMap:
+        shrink_ratio: 0.4
+        thresh_min: 0.3
+        thresh_max: 0.7
+    - NormalizeImage:
+        scale: 1./255.
+        mean:
+        - 0.48109378172549
+        - 0.45752457890196
+        - 0.40787054090196
+        std:
+        - 1.0
+        - 1.0
+        - 1.0
+        order: hwc
+    - ToCHWImage: null
+    - KeepKeys:
+        keep_keys:
+        - image
+        - threshold_map
+        - threshold_mask
+        - shrink_map
+        - shrink_mask
+  loader:
+    shuffle: true
+    drop_last: false
+    batch_size_per_card: 4
+    num_workers: 8
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/                                                                          
+    label_file_list:                                                                                            
+    - ./train_data/TD_TR/TD500/test_gt_labels.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - DetLabelEncode: null
+    - DetResizeForTest:
+        image_shape:
+        - 736
+        - 736
+        keep_ratio: True
+    - NormalizeImage:
+        scale: 1./255.
+        mean:
+        - 0.48109378172549
+        - 0.45752457890196
+        - 0.40787054090196
+        std:
+        - 1.0
+        - 1.0
+        - 1.0
+        order: hwc
+    - ToCHWImage: null
+    - KeepKeys:
+        keep_keys:
+        - image
+        - shape
+        - polys
+        - ignore_tags
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 1
+    num_workers: 2
+profiler_options: null

configs/det/det_r50_vd_db.yml

@@ -20,7 +20,7 @@ Architecture:
   algorithm: DB
   Transform:
   Backbone:
-    name: ResNet
+    name: ResNet_vd
     layers: 50
   Neck:
     name: DBFPN

configs/det/det_r50_vd_dcn_fce_ctw.yml

@@ -21,7 +21,7 @@ Architecture:
   algorithm: FCE
   Transform:
   Backbone:
-    name: ResNet
+    name: ResNet_vd
     layers: 50
     dcn_stage: [False, True, True, True]
     out_indices: [1,2,3]

configs/det/det_r50_vd_east.yml

@@ -20,7 +20,7 @@ Architecture:
   algorithm: EAST
   Transform:
   Backbone:
-    name: ResNet
+    name: ResNet_vd
     layers: 50
   Neck:
     name: EASTFPN

configs/det/det_r50_vd_pse.yml

@@ -20,7 +20,7 @@ Architecture:
   algorithm: PSE
   Transform:
   Backbone:
-    name: ResNet
+    name: ResNet_vd
     layers: 50
   Neck:
     name: FPN

configs/det/det_res18_db_v2.0.yml

@@ -20,7 +20,7 @@ Architecture:
   algorithm: DB
   Transform:
   Backbone:
-    name: ResNet
+    name: ResNet_vd
     layers: 18
     disable_se: True
   Neck:

configs/kie/kie_unet_sdmgr.yml

@@ -17,7 +17,7 @@ Global:
   checkpoints:
   save_inference_dir:
   use_visualdl: False
-  class_path: ./train_data/wildreceipt/class_list.txt
+  class_path: &class_path ./train_data/wildreceipt/class_list.txt
   infer_img: ./train_data/wildreceipt/1.txt
   save_res_path: ./output/sdmgr_kie/predicts_kie.txt
   img_scale: [ 1024, 512 ]
@@ -72,6 +72,7 @@ Train:
           order: 'hwc'
       - KieLabelEncode: # Class handling label
           character_dict_path: ./train_data/wildreceipt/dict.txt
+          class_path: *class_path
       - KieResize:
       - ToCHWImage:
       - KeepKeys:
@@ -88,7 +89,6 @@ Eval:
     data_dir: ./train_data/wildreceipt
     label_file_list:
       - ./train_data/wildreceipt/wildreceipt_test.txt
-      # - /paddle/data/PaddleOCR/train_data/wildreceipt/1.txt
     transforms:
       - DecodeImage: # load image
           img_mode: RGB

configs/rec/rec_mtb_nrtr.yml

@@ -9,7 +9,7 @@ Global:
   eval_batch_step: [0, 2000]
   cal_metric_during_train: True
   pretrained_model:
-  checkpoints: 
+  checkpoints:
   save_inference_dir:
   use_visualdl: False
   infer_img: doc/imgs_words_en/word_10.png
@@ -49,7 +49,7 @@ Architecture:
     
 
 Loss:
-  name: NRTRLoss
+  name: CELoss
   smoothing: True
 
 PostProcess:
@@ -68,8 +68,8 @@ Train:
           img_mode: BGR
           channel_first: False
       - NRTRLabelEncode: # Class handling label
-      - NRTRRecResizeImg:
-          image_shape: [100, 32]
+      - GrayRecResizeImg:
+          image_shape: [100, 32] # W H
           resize_type: PIL # PIL or OpenCV
       - KeepKeys:
           keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
@@ -82,14 +82,14 @@ Train:
 Eval:
   dataset:
     name: LMDBDataSet
-    data_dir: ./train_data/data_lmdb_release/evaluation/
+    data_dir: ./train_data/data_lmdb_release/validation/
     transforms:
       - DecodeImage: # load image
           img_mode: BGR
           channel_first: False
       - NRTRLabelEncode: # Class handling label
-      - NRTRRecResizeImg:
-          image_shape: [100, 32]
+      - GrayRecResizeImg:
+          image_shape: [100, 32] # W H
           resize_type: PIL # PIL or OpenCV
       - KeepKeys:
           keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
@@ -97,5 +97,5 @@ Eval:
     shuffle: False
     drop_last: False
     batch_size_per_card: 256
-    num_workers: 1
+    num_workers: 4
     use_shared_memory: False

configs/rec/rec_r45_abinet.yml

+Global:
+  use_gpu: True
+  epoch_num: 10
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/r45_abinet/
+  save_epoch_step: 1
+  # evaluation is run every 2000 iterations
+  eval_batch_step: [0, 2000]
+  cal_metric_during_train: True
+  pretrained_model:
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: False
+  infer_img: doc/imgs_words_en/word_10.png
+  # for data or label process
+  character_dict_path:
+  character_type: en
+  max_text_length: 25
+  infer_mode: False
+  use_space_char: False
+  save_res_path: ./output/rec/predicts_abinet.txt
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.99
+  clip_norm: 20.0
+  lr:
+    name: Piecewise
+    decay_epochs: [6]
+    values: [0.0001, 0.00001] 
+  regularizer:
+    name: 'L2'
+    factor: 0.
+
+Architecture:
+  model_type: rec
+  algorithm: ABINet
+  in_channels: 3
+  Transform:
+  Backbone:
+    name: ResNet45
+  Head:
+    name: ABINetHead
+    use_lang: True
+    iter_size: 3
+    
+
+Loss:
+  name: CELoss
+  ignore_index: &ignore_index 100 # Must be greater than the number of character classes
+
+PostProcess:
+  name: ABINetLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+Train:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/training/
+    transforms:
+      - DecodeImage: # load image
+          img_mode: RGB
+          channel_first: False
+      - ABINetRecAug:
+      - ABINetLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - ABINetRecResizeImg:
+          image_shape: [3, 32, 128]
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    batch_size_per_card: 96
+    drop_last: True
+    num_workers: 4
+
+Eval:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/validation/
+    transforms:
+      - DecodeImage: # load image
+          img_mode: RGB
+          channel_first: False
+      - ABINetLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - ABINetRecResizeImg:
+          image_shape: [3, 32, 128]
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 256
+    num_workers: 4
+    use_shared_memory: False

configs/rec/rec_svtrnet.yml

@@ -26,7 +26,7 @@ Optimizer:
   name: AdamW
   beta1: 0.9
   beta2: 0.99
-  epsilon: 0.00000008
+  epsilon: 8.e-8
   weight_decay: 0.05
   no_weight_decay_name: norm pos_embed
   one_dim_param_no_weight_decay: true
@@ -77,14 +77,13 @@ Metric:
 Train:
   dataset:
     name: LMDBDataSet
-    data_dir: ./train_data/data_lmdb_release/training/
+    data_dir: ./train_data/data_lmdb_release/training
     transforms:
       - DecodeImage: # load image
           img_mode: BGR
           channel_first: False
       - CTCLabelEncode: # Class handling label
-      - RecResizeImg:
-          character_dict_path:
+      - SVTRRecResizeImg:
           image_shape: [3, 64, 256]
           padding: False
       - KeepKeys:
@@ -98,14 +97,13 @@ Train:
 Eval:
   dataset:
     name: LMDBDataSet
-    data_dir: ./train_data/data_lmdb_release/validation/
+    data_dir: ./train_data/data_lmdb_release/validation
     transforms:
       - DecodeImage: # load image
           img_mode: BGR
           channel_first: False
       - CTCLabelEncode: # Class handling label
-      - RecResizeImg:
-          character_dict_path:
+      - SVTRRecResizeImg:
           image_shape: [3, 64, 256]
           padding: False
       - KeepKeys:

configs/rec/rec_svtrnet_ch.yml

+Global:
+  use_gpu: true
+  epoch_num: 100
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/svtr_ch_all/
+  save_epoch_step: 10
+  eval_batch_step:
+  - 0
+  - 2000
+  cal_metric_during_train: true
+  pretrained_model: null
+  checkpoints: null
+  save_inference_dir: null
+  use_visualdl: false
+  infer_img: doc/imgs_words/ch/word_1.jpg
+  character_dict_path: ppocr/utils/ppocr_keys_v1.txt
+  max_text_length: 25
+  infer_mode: false
+  use_space_char: true
+  save_res_path: ./output/rec/predicts_svtr_tiny_ch_all.txt
+Optimizer:
+  name: AdamW
+  beta1: 0.9
+  beta2: 0.99
+  epsilon: 8.0e-08
+  weight_decay: 0.05
+  no_weight_decay_name: norm pos_embed
+  one_dim_param_no_weight_decay: true
+  lr:
+    name: Cosine
+    learning_rate: 0.0005
+    warmup_epoch: 2
+Architecture:
+  model_type: rec
+  algorithm: SVTR
+  Transform: null
+  Backbone:
+    name: SVTRNet
+    img_size:
+    - 32
+    - 320
+    out_char_num: 40
+    out_channels: 96
+    patch_merging: Conv
+    embed_dim:
+    - 64
+    - 128
+    - 256
+    depth:
+    - 3
+    - 6
+    - 3
+    num_heads:
+    - 2
+    - 4
+    - 8
+    mixer:
+    - Local
+    - Local
+    - Local
+    - Local
+    - Local
+    - Local
+    - Global
+    - Global
+    - Global
+    - Global
+    - Global
+    - Global
+    local_mixer:
+    - - 7
+      - 11
+    - - 7
+      - 11
+    - - 7
+      - 11
+    last_stage: true
+    prenorm: false
+  Neck:
+    name: SequenceEncoder
+    encoder_type: reshape
+  Head:
+    name: CTCHead
+Loss:
+  name: CTCLoss
+PostProcess:
+  name: CTCLabelDecode
+Metric:
+  name: RecMetric
+  main_indicator: acc
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data
+    label_file_list:
+    - ./train_data/train_list.txt
+    ext_op_transform_idx: 1
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - RecConAug:
+        prob: 0.5
+        ext_data_num: 2
+        image_shape:
+        - 32
+        - 320
+        - 3
+    - RecAug: null
+    - CTCLabelEncode: null
+    - SVTRRecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+        padding: true
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: true
+    batch_size_per_card: 256
+    drop_last: true
+    num_workers: 8
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data
+    label_file_list:
+    - ./train_data/val_list.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - CTCLabelEncode: null
+    - SVTRRecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+        padding: true
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 256
+    num_workers: 2
+profiler_options: null

configs/rec/rec_vitstr_none_ce.yml

+Global:
+  use_gpu: True
+  epoch_num: 20
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/vitstr_none_ce/
+  save_epoch_step: 1
+  # evaluation is run every 2000 iterations after the 0th iteration#
+  eval_batch_step: [0, 2000]
+  cal_metric_during_train: True
+  pretrained_model:
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: False
+  infer_img: doc/imgs_words_en/word_10.png
+  # for data or label process
+  character_dict_path: ppocr/utils/EN_symbol_dict.txt
+  max_text_length: 25
+  infer_mode: False
+  use_space_char: False
+  save_res_path: ./output/rec/predicts_vitstr.txt
+
+
+Optimizer:
+  name: Adadelta
+  epsilon: 1.e-8
+  rho: 0.95
+  clip_norm: 5.0
+  lr:
+    learning_rate: 1.0
+
+Architecture:
+  model_type: rec
+  algorithm: ViTSTR
+  in_channels: 1
+  Transform:
+  Backbone:
+    name: ViTSTR
+    scale: tiny
+  Neck:
+    name: SequenceEncoder
+    encoder_type: reshape
+  Head:
+    name: CTCHead
+
+Loss:
+  name: CELoss
+  with_all: True
+  ignore_index: &ignore_index 0 # Must be zero or greater than the number of character classes
+
+PostProcess:
+  name: ViTSTRLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+Train:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/training/
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - ViTSTRLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - GrayRecResizeImg:
+          image_shape: [224, 224] # W H
+          resize_type: PIL # PIL or OpenCV
+          inter_type: 'Image.BICUBIC'
+          scale: false
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    batch_size_per_card: 48
+    drop_last: True
+    num_workers: 8
+
+Eval:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/validation/
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - ViTSTRLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - GrayRecResizeImg:
+          image_shape: [224, 224] # W H
+          resize_type: PIL # PIL or OpenCV
+          inter_type: 'Image.BICUBIC'
+          scale: false
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 256
+    num_workers: 2

configs/table/table_master.yml

+Global:
+  use_gpu: true
+  epoch_num: 17
+  log_smooth_window: 20
+  print_batch_step: 100
+  save_model_dir: ./output/table_master/
+  save_epoch_step: 17
+  eval_batch_step: [0,  6259]
+  cal_metric_during_train: true
+  pretrained_model: null
+  checkpoints: 
+  save_inference_dir: output/table_master/infer
+  use_visualdl: false
+  infer_img: ppstructure/docs/table/table.jpg
+  save_res_path: ./output/table_master
+  character_dict_path: ppocr/utils/dict/table_master_structure_dict.txt
+  infer_mode: false
+  max_text_length: 500
+  process_total_num: 0
+  process_cut_num: 0
+
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: MultiStepDecay
+    learning_rate: 0.001
+    milestones: [12, 15]
+    gamma: 0.1
+    warmup_epoch: 0.02
+  regularizer:
+    name: L2
+    factor: 0.0
+
+Architecture:
+  model_type: table
+  algorithm: TableMaster
+  Backbone:
+    name: TableResNetExtra
+    gcb_config:
+      ratio: 0.0625
+      headers: 1
+      att_scale: False
+      fusion_type: channel_add
+      layers: [False, True, True, True]
+    layers: [1,2,5,3]
+  Head:
+    name: TableMasterHead
+    hidden_size: 512
+    headers: 8
+    dropout: 0
+    d_ff: 2024
+    max_text_length: 500
+
+Loss:
+  name: TableMasterLoss
+  ignore_index: 42 # set to len of dict + 3
+
+PostProcess:
+  name: TableMasterLabelDecode
+  box_shape: pad
+
+Metric:
+  name: TableMetric
+  main_indicator: acc
+  compute_bbox_metric: False
+
+Train:
+  dataset:
+    name: PubTabDataSet
+    data_dir: train_data/table/pubtabnet/train/
+    label_file_list: [train_data/table/pubtabnet/PubTabNet_2.0.0_train.jsonl]
+    transforms:
+      - DecodeImage:
+          img_mode: BGR
+          channel_first: False
+      - TableMasterLabelEncode:
+          learn_empty_box: False
+          merge_no_span_structure: True
+          replace_empty_cell_token: True
+      - ResizeTableImage:
+          max_len: 480
+          resize_bboxes: True
+      - PaddingTableImage:
+          size: [480, 480]
+      - TableBoxEncode:
+          use_xywh: True
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [0.5, 0.5, 0.5]
+          std: [0.5, 0.5, 0.5]
+          order: hwc
+      - ToCHWImage: null
+      - KeepKeys:
+          keep_keys: [image, structure, bboxes, bbox_masks, shape]
+  loader:
+    shuffle: True
+    batch_size_per_card: 10
+    drop_last: True
+    num_workers: 8
+
+Eval:
+  dataset:
+    name: PubTabDataSet
+    data_dir: train_data/table/pubtabnet/train/
+    label_file_list: [train_data/table/pubtabnet/PubTabNet_2.0.0_val.jsonl]
+    transforms:
+      - DecodeImage:
+          img_mode: BGR
+          channel_first: False
+      - TableMasterLabelEncode:
+          learn_empty_box: False
+          merge_no_span_structure: True
+          replace_empty_cell_token: True
+      - ResizeTableImage:
+          max_len: 480
+          resize_bboxes: True
+      - PaddingTableImage:
+          size: [480, 480]
+      - TableBoxEncode:
+          use_xywh: True
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [0.5, 0.5, 0.5]
+          std: [0.5, 0.5, 0.5]
+          order: hwc
+      - ToCHWImage: null
+      - KeepKeys:
+          keep_keys: [image, structure, bboxes, bbox_masks, shape]
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 10
+    num_workers: 8
\ No newline at end of file

configs/table/table_mv3.yml

@@ -4,21 +4,20 @@ Global:
   log_smooth_window: 20
   print_batch_step: 5
   save_model_dir: ./output/table_mv3/
-  save_epoch_step: 3
+  save_epoch_step: 400
   # evaluation is run every 400 iterations after the 0th iteration
   eval_batch_step: [0, 400]
   cal_metric_during_train: True
   pretrained_model:
-  checkpoints: 
+  checkpoints:
   save_inference_dir:
   use_visualdl: False
-  infer_img: doc/table/table.jpg
+  infer_img: ppstructure/docs/table/table.jpg
+  save_res_path: output/table_mv3
   # for data or label process
   character_dict_path: ppocr/utils/dict/table_structure_dict.txt
   character_type: en
-  max_text_length: 100
-  max_elem_length: 800
-  max_cell_num: 500
+  max_text_length: 800
   infer_mode: False
   process_total_num: 0
   process_cut_num: 0
@@ -44,11 +43,8 @@ Architecture:
   Head:
     name: TableAttentionHead
     hidden_size: 256
-    l2_decay: 0.00001
     loc_type: 2
-    max_text_length: 100
-    max_elem_length: 800
-    max_cell_num: 500
+    max_text_length: 800
 
 Loss:
   name: TableAttentionLoss
@@ -61,28 +57,34 @@ PostProcess:
 Metric:
   name: TableMetric
   main_indicator: acc
+  compute_bbox_metric: false # cost many time, set False for training
 
 Train:
   dataset:
     name: PubTabDataSet
     data_dir: train_data/table/pubtabnet/train/
-    label_file_path: train_data/table/pubtabnet/PubTabNet_2.0.0_train.jsonl
+    label_file_list: [train_data/table/pubtabnet/PubTabNet_2.0.0_train.jsonl]
     transforms:
       - DecodeImage: # load image
           img_mode: BGR
           channel_first: False
+      - TableLabelEncode:
+          learn_empty_box: False
+          merge_no_span_structure: False
+          replace_empty_cell_token: False
+      - TableBoxEncode:
       - ResizeTableImage:
           max_len: 488
-      - TableLabelEncode:
       - NormalizeImage:
           scale: 1./255.
           mean: [0.485, 0.456, 0.406]
           std: [0.229, 0.224, 0.225]
           order: 'hwc'
       - PaddingTableImage:
+          size: [488, 488]
       - ToCHWImage:
       - KeepKeys:
-          keep_keys: ['image', 'structure', 'bbox_list', 'sp_tokens', 'bbox_list_mask']
+          keep_keys: [ 'image', 'structure', 'bboxes', 'bbox_masks', 'shape' ]
   loader:
     shuffle: True
     batch_size_per_card: 32
@@ -92,24 +94,29 @@ Train:
 Eval:
   dataset:
     name: PubTabDataSet
-    data_dir: train_data/table/pubtabnet/val/
-    label_file_path: train_data/table/pubtabnet/PubTabNet_2.0.0_val.jsonl
+    data_dir: /home/zhoujun20/table/PubTabNe/pubtabnet/val/
+    label_file_list: [/home/zhoujun20/table/PubTabNe/pubtabnet/val_500.jsonl]
     transforms:
       - DecodeImage: # load image
           img_mode: BGR
           channel_first: False
+      - TableLabelEncode:
+          learn_empty_box: False
+          merge_no_span_structure: False
+          replace_empty_cell_token: False
+      - TableBoxEncode:
       - ResizeTableImage:
           max_len: 488
-      - TableLabelEncode:
       - NormalizeImage:
           scale: 1./255.
           mean: [0.485, 0.456, 0.406]
           std: [0.229, 0.224, 0.225]
           order: 'hwc'
       - PaddingTableImage:
+          size: [488, 488]
       - ToCHWImage:
       - KeepKeys:
-          keep_keys: ['image', 'structure', 'bbox_list', 'sp_tokens', 'bbox_list_mask']
+          keep_keys: [ 'image', 'structure', 'bboxes', 'bbox_masks', 'shape' ]
   loader:
     shuffle: False
     drop_last: False

configs/vqa/re/layoutlmv2_funsd.yml

+Global:
+  use_gpu: True
+  epoch_num: &epoch_num 200
+  log_smooth_window: 10
+  print_batch_step: 10
+  save_model_dir: ./output/re_layoutlmv2_funsd
+  save_epoch_step: 2000
+  # evaluation is run every 10 iterations after the 0th iteration
+  eval_batch_step: [ 0, 57 ]
+  cal_metric_during_train: False
+  save_inference_dir:
+  use_visualdl: False
+  seed: 2022
+  infer_img: train_data/FUNSD/testing_data/images/83624198.png
+  save_res_path: ./output/re_layoutlmv2_funsd/res/
+
+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/FUNSD/training_data/images/
+    label_file_list: 
+      - ./train_data/FUNSD/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 train_data/FUNSD/class_list.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:
+          # dataloader will return list in this order
+          keep_keys: [ 'input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'image', 'entities', 'relations']
+  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/FUNSD/testing_data/images/
+    label_file_list: 
+      - ./train_data/FUNSD/test.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:
+          # dataloader will return list in this order
+          keep_keys: [ 'input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'image', 'entities', 'relations']
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 8
+    num_workers: 8
+    collate_fn: ListCollator

configs/vqa/re/layoutlmv2.yml → configs/vqa/re/layoutlmv2_xund_zh.yml

@@ -3,16 +3,16 @@ Global:
   epoch_num: &epoch_num 200
   log_smooth_window: 10
   print_batch_step: 10
-  save_model_dir: ./output/re_layoutlmv2/
+  save_model_dir: ./output/re_layoutlmv2_xfund_zh
   save_epoch_step: 2000
   # evaluation is run every 10 iterations after the 0th iteration
-  eval_batch_step: [ 0, 19 ]
+  eval_batch_step: [ 0, 57 ]
   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/
+  infer_img: ppstructure/docs/vqa/input/zh_val_21.jpg
+  save_res_path: ./output/re_layoutlmv2_xfund_zh/res/
 
 Architecture:
   model_type: vqa
@@ -21,7 +21,7 @@ Architecture:
   Backbone:
     name: LayoutLMv2ForRe
     pretrained: True
-    checkpoints: 
+    checkpoints:
 
 Loss:
   name: LossFromOutput
@@ -52,7 +52,7 @@ Train:
     name: SimpleDataSet
     data_dir: train_data/XFUND/zh_train/image
     label_file_list: 
-      - train_data/XFUND/zh_train/xfun_normalize_train.json
+      - train_data/XFUND/zh_train/train.json
     ratio_list: [ 1.0 ]
     transforms:
       - DecodeImage: # load image
@@ -61,7 +61,7 @@ Train:
       - VQATokenLabelEncode: # Class handling label
           contains_re: True
           algorithm: *algorithm
-          class_path: &class_path ppstructure/vqa/labels/labels_ser.txt
+          class_path: &class_path train_data/XFUND/class_list_xfun.txt
       - VQATokenPad:
           max_seq_len: &max_seq_len 512
           return_attention_mask: True
@@ -77,7 +77,7 @@ Train:
           order: 'hwc'
       - ToCHWImage:
       - KeepKeys:
-          keep_keys: [ 'input_ids', 'bbox', 'image', 'attention_mask', 'token_type_ids','entities', 'relations'] # dataloader will return list in this order
+          keep_keys: [ 'input_ids', 'bbox', 'attention_mask', 'token_type_ids','image', 'entities', 'relations'] # dataloader will return list in this order
   loader:
     shuffle: True
     drop_last: False
@@ -90,7 +90,7 @@ Eval:
     name: SimpleDataSet
     data_dir: train_data/XFUND/zh_val/image
     label_file_list:
-      - train_data/XFUND/zh_val/xfun_normalize_val.json
+      - train_data/XFUND/zh_val/val.json
     transforms:
       - DecodeImage: # load image
           img_mode: RGB
@@ -114,7 +114,7 @@ Eval:
           order: 'hwc'
       - ToCHWImage:
       - KeepKeys:
-          keep_keys: [ 'input_ids', 'bbox', 'image', 'attention_mask', 'token_type_ids','entities', 'relations'] # dataloader will return list in this order
+          keep_keys: [ 'input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'image','entities', 'relations'] # dataloader will return list in this order
   loader:
     shuffle: False
     drop_last: False

configs/vqa/re/layoutxlm_funsd.yml

+Global:
+  use_gpu: True
+  epoch_num: &epoch_num 200
+  log_smooth_window: 10
+  print_batch_step: 10
+  save_model_dir: ./output/re_layoutxlm_funsd
+  save_epoch_step: 2000
+  # evaluation is run every 10 iterations after the 0th iteration
+  eval_batch_step: [ 0, 57 ]
+  cal_metric_during_train: False
+  save_inference_dir:
+  use_visualdl: False
+  seed: 2022
+  infer_img: train_data/FUNSD/testing_data/images/83624198.png
+  save_res_path: ./output/re_layoutxlm_funsd/res/
+
+Architecture:
+  model_type: vqa
+  algorithm: &algorithm "LayoutXLM"
+  Transform:
+  Backbone:
+    name: LayoutXLMForRe
+    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/FUNSD/training_data/images/
+    label_file_list: 
+      - ./train_data/FUNSD/train_v4.json
+      # - ./train_data/FUNSD/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 ./train_data/FUNSD/class_list.txt
+          use_textline_bbox_info: &use_textline_bbox_info True
+      - 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
+          mean: [ 123.675, 116.28, 103.53 ]
+          std: [ 58.395, 57.12, 57.375 ]
+          order: 'hwc'
+      - ToCHWImage:
+      - KeepKeys:
+          # dataloader will return list in this order
+          keep_keys: [ 'input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'image', 'entities', 'relations']
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 8
+    num_workers: 16
+    collate_fn: ListCollator
+
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/FUNSD/testing_data/images/
+    label_file_list: 
+      - ./train_data/FUNSD/test_v4.json
+      # - ./train_data/FUNSD/test.json
+    transforms:
+      - DecodeImage: # load image
+          img_mode: RGB
+          channel_first: False
+      - VQATokenLabelEncode: # Class handling label
+          contains_re: True
+          algorithm: *algorithm
+          class_path: *class_path
+          use_textline_bbox_info: *use_textline_bbox_info
+      - 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
+          mean: [ 123.675, 116.28, 103.53 ]
+          std: [ 58.395, 57.12, 57.375 ]
+          order: 'hwc'
+      - ToCHWImage:
+      - KeepKeys:
+          # dataloader will return list in this order
+          keep_keys: [ 'input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'image', 'entities', 'relations']
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 8
+    num_workers: 8
+    collate_fn: ListCollator

configs/vqa/re/layoutxlm.yml → configs/vqa/re/layoutxlm_xfund_zh.yml

@@ -11,7 +11,7 @@ Global:
   save_inference_dir:
   use_visualdl: False
   seed: 2022
-  infer_img: doc/vqa/input/zh_val_21.jpg
+  infer_img: ppstructure/docs/vqa/input/zh_val_21.jpg
   save_res_path: ./output/re/
 
 Architecture:
@@ -52,7 +52,7 @@ Train:
     name: SimpleDataSet
     data_dir: train_data/XFUND/zh_train/image
     label_file_list: 
-      - train_data/XFUND/zh_train/xfun_normalize_train.json
+      - train_data/XFUND/zh_train/train.json
     ratio_list: [ 1.0 ]
     transforms:
       - DecodeImage: # load image
@@ -61,7 +61,7 @@ Train:
       - VQATokenLabelEncode: # Class handling label
           contains_re: True
           algorithm: *algorithm
-          class_path: &class_path ppstructure/vqa/labels/labels_ser.txt
+          class_path: &class_path train_data/XFUND/class_list_xfun.txt
       - VQATokenPad:
           max_seq_len: &max_seq_len 512
           return_attention_mask: True
@@ -77,7 +77,7 @@ Train:
           order: 'hwc'
       - ToCHWImage:
       - KeepKeys:
-          keep_keys: [ 'input_ids', 'bbox', 'image', 'attention_mask', 'token_type_ids','entities', 'relations'] # dataloader will return list in this order
+          keep_keys: [ 'input_ids', 'bbox','attention_mask', 'token_type_ids', 'image', 'entities', 'relations'] # dataloader will return list in this order
   loader:
     shuffle: True
     drop_last: False
@@ -90,7 +90,7 @@ Eval:
     name: SimpleDataSet
     data_dir: train_data/XFUND/zh_val/image
     label_file_list:
-      - train_data/XFUND/zh_val/xfun_normalize_val.json
+      - train_data/XFUND/zh_val/val.json
     transforms:
       - DecodeImage: # load image
           img_mode: RGB
@@ -114,7 +114,7 @@ Eval:
           order: 'hwc'
       - ToCHWImage:
       - KeepKeys:
-          keep_keys: [ 'input_ids', 'bbox', 'image', 'attention_mask', 'token_type_ids','entities', 'relations'] # dataloader will return list in this order
+          keep_keys: [ 'input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'image', 'entities', 'relations'] # dataloader will return list in this order
   loader:
     shuffle: False
     drop_last: False

deploy/avh/.gitignore

+include/inputs.h
+include/outputs.h
+
+__pycache__/
+build/
\ No newline at end of file

deploy/avh/requirements.txt

+paddlepaddle
+numpy
+opencv-python
\ No newline at end of file

deploy/lite/config.txt

@@ -4,4 +4,5 @@ det_db_box_thresh  0.5
 det_db_unclip_ratio  1.6
 det_db_use_dilate 0
 det_use_polygon_score 1
-use_direction_classify  1
\ No newline at end of file
+use_direction_classify  1
+rec_image_height  32
\ No newline at end of file