未验证 提交 349b7d38 编写于 作者: 天涯古巷's avatar 天涯古巷 提交者: GitHub

Merge branch 'PaddlePaddle:release/2.3' into release/2.3

......@@ -141,6 +141,7 @@ Train:
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- CopyPaste:
- IaaAugment:
augmenter_args:
- { 'type': Fliplr, 'args': { 'p': 0.5 } }
......
......@@ -91,7 +91,7 @@ Optimizer:
PostProcess:
name: DistillationDBPostProcess
model_name: ["Student", "Student2"]
model_name: ["Student"]
key: head_out
thresh: 0.3
box_thresh: 0.6
......
......@@ -8,7 +8,7 @@ Global:
# evaluation is run every 5000 iterations after the 4000th iteration
eval_batch_step: [4000, 5000]
cal_metric_during_train: False
pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained/
pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained
checkpoints:
save_inference_dir:
use_visualdl: False
......
......@@ -8,7 +8,7 @@ Global:
# evaluation is run every 5000 iterations after the 4000th iteration
eval_batch_step: [4000, 5000]
cal_metric_during_train: False
pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained/
pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained
checkpoints:
save_inference_dir:
use_visualdl: False
......
Global:
use_gpu: true
epoch_num: 50
epoch_num: 400
log_smooth_window: 20
print_batch_step: 5
save_model_dir: ./output/table_mv3/
save_epoch_step: 5
save_epoch_step: 3
# evaluation is run every 400 iterations after the 0th iteration
eval_batch_step: [0, 400]
cal_metric_during_train: True
......@@ -12,18 +12,17 @@ Global:
checkpoints:
save_inference_dir:
use_visualdl: False
infer_img: doc/imgs_words/ch/word_1.jpg
infer_img: doc/table/table.jpg
# 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: 500
max_elem_length: 800
max_cell_num: 500
infer_mode: False
process_total_num: 0
process_cut_num: 0
Optimizer:
name: Adam
beta1: 0.9
......@@ -41,13 +40,15 @@ Architecture:
Backbone:
name: MobileNetV3
scale: 1.0
model_name: small
disable_se: True
model_name: large
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
Loss:
name: TableAttentionLoss
......
......@@ -18,7 +18,7 @@ import paddlehub as hub
from tools.infer.utility import base64_to_cv2
from tools.infer.predict_det import TextDetector
from tools.infer.utility import parse_args
from deploy.hubserving.ocr_system.params import read_params
from deploy.hubserving.ocr_det.params import read_params
@moduleinfo(
......
......@@ -50,7 +50,7 @@ PaddleOCR基于动态图开源的文本识别算法列表:
- [x] STAR-Net([paper](http://www.bmva.org/bmvc/2016/papers/paper043/index.html))[11]
- [x] RARE([paper](https://arxiv.org/abs/1603.03915v1))[12]
- [x] SRN([paper](https://arxiv.org/abs/2003.12294))[5]
- [x] NRTR([paper](https://arxiv.org/abs/1806.00926v2))
- [x] NRTR([paper](https://arxiv.org/abs/1806.00926v2))[13]
参考[DTRB][3](https://arxiv.org/abs/1904.01906)文字识别训练和评估流程,使用MJSynth和SynthText两个文字识别数据集训练,在IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE数据集上进行评估,算法效果如下:
......@@ -78,4 +78,3 @@ PaddleOCR文本检测算法的训练和使用请参考文档教程中[模型训
## 3. 模型推理
上述模型中除PP-OCR系列模型以外,其余模型仅支持基于Python引擎的推理,具体内容可参考[基于Python预测引擎推理](./inference.md)
......@@ -112,4 +112,14 @@
year={2016}
}
13.NRTR
@misc{sheng2019nrtr,
title={NRTR: A No-Recurrence Sequence-to-Sequence Model For Scene Text Recognition},
author={Fenfen Sheng and Zhineng Chen and Bo Xu},
year={2019},
eprint={1806.00926},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
......@@ -11,7 +11,10 @@
#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 refered from:
https://github.com/songdejia/EAST/blob/master/data_utils.py
"""
import math
import cv2
import numpy as np
......@@ -24,10 +27,10 @@ __all__ = ['EASTProcessTrain']
class EASTProcessTrain(object):
def __init__(self,
image_shape = [512, 512],
background_ratio = 0.125,
min_crop_side_ratio = 0.1,
min_text_size = 10,
image_shape=[512, 512],
background_ratio=0.125,
min_crop_side_ratio=0.1,
min_text_size=10,
**kwargs):
self.input_size = image_shape[1]
self.random_scale = np.array([0.5, 1, 2.0, 3.0])
......@@ -282,12 +285,7 @@ class EASTProcessTrain(object):
1.0 / max(min(poly_h, poly_w), 1.0)
return score_map, geo_map, training_mask
def crop_area(self,
im,
polys,
tags,
crop_background=False,
max_tries=50):
def crop_area(self, im, polys, tags, crop_background=False, max_tries=50):
"""
make random crop from the input image
:param im:
......
......@@ -11,6 +11,11 @@
# 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/WenmuZhou/DBNet.pytorch/blob/master/data_loader/modules/iaa_augment.py
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......
# -*- coding:utf-8 -*-
# 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/WenmuZhou/DBNet.pytorch/blob/master/data_loader/modules/make_border_map.py
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......
# -*- coding:utf-8 -*-
# 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/WenmuZhou/DBNet.pytorch/blob/master/data_loader/modules/make_shrink_map.py
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......
# -*- coding:utf-8 -*-
# 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/WenmuZhou/DBNet.pytorch/blob/master/data_loader/modules/random_crop_data.py
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......
......@@ -11,7 +11,10 @@
#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 part code is refered from:
https://github.com/songdejia/EAST/blob/master/data_utils.py
"""
import math
import cv2
import numpy as np
......
......@@ -11,7 +11,10 @@
# 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/RubanSeven/Text-Image-Augmentation-python/blob/master/augment.py
"""
import numpy as np
from .warp_mls import WarpMLS
......
......@@ -11,7 +11,10 @@
# 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/RubanSeven/Text-Image-Augmentation-python/blob/master/warp_mls.py
"""
import numpy as np
......
......@@ -11,7 +11,10 @@
# 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/WenmuZhou/DBNet.pytorch/blob/master/models/losses/basic_loss.py
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......
......@@ -11,6 +11,10 @@
# 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/WenmuZhou/DBNet.pytorch/blob/master/models/losses/DB_loss.py
"""
from __future__ import absolute_import
from __future__ import division
......
......@@ -12,6 +12,8 @@
# See the License for the specific language governing permissions and
# limitations under the License.
# This code is refer from: https://github.com/PaddlePaddle/PaddleClas/blob/develop/ppcls/arch/backbone/legendary_models/pp_lcnet.py
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......
......@@ -75,7 +75,7 @@ class AttentionHead(nn.Layer):
probs_step, axis=1)], axis=1)
next_input = probs_step.argmax(axis=1)
targets = next_input
probs = paddle.nn.functional.softmax(probs, axis=2)
return probs
......
......@@ -23,14 +23,22 @@ import numpy as np
class TableAttentionHead(nn.Layer):
def __init__(self, in_channels, hidden_size, loc_type, in_max_len=488, **kwargs):
def __init__(self,
in_channels,
hidden_size,
loc_type,
in_max_len=488,
max_text_length=100,
max_elem_length=800,
max_cell_num=500,
**kwargs):
super(TableAttentionHead, self).__init__()
self.input_size = in_channels[-1]
self.hidden_size = hidden_size
self.elem_num = 30
self.max_text_length = 100
self.max_elem_length = 500
self.max_cell_num = 500
self.max_text_length = max_text_length
self.max_elem_length = max_elem_length
self.max_cell_num = max_cell_num
self.structure_attention_cell = AttentionGRUCell(
self.input_size, hidden_size, self.elem_num, use_gru=False)
......@@ -42,11 +50,11 @@ class TableAttentionHead(nn.Layer):
self.loc_generator = nn.Linear(hidden_size, 4)
else:
if self.in_max_len == 640:
self.loc_fea_trans = nn.Linear(400, self.max_elem_length+1)
self.loc_fea_trans = nn.Linear(400, self.max_elem_length + 1)
elif self.in_max_len == 800:
self.loc_fea_trans = nn.Linear(625, self.max_elem_length+1)
self.loc_fea_trans = nn.Linear(625, self.max_elem_length + 1)
else:
self.loc_fea_trans = nn.Linear(256, self.max_elem_length+1)
self.loc_fea_trans = nn.Linear(256, self.max_elem_length + 1)
self.loc_generator = nn.Linear(self.input_size + hidden_size, 4)
def _char_to_onehot(self, input_char, onehot_dim):
......@@ -69,7 +77,7 @@ class TableAttentionHead(nn.Layer):
output_hiddens = []
if self.training and targets is not None:
structure = targets[0]
for i in range(self.max_elem_length+1):
for i in range(self.max_elem_length + 1):
elem_onehots = self._char_to_onehot(
structure[:, i], onehot_dim=self.elem_num)
(outputs, hidden), alpha = self.structure_attention_cell(
......@@ -96,7 +104,7 @@ class TableAttentionHead(nn.Layer):
alpha = None
max_elem_length = paddle.to_tensor(self.max_elem_length)
i = 0
while i < max_elem_length+1:
while i < max_elem_length + 1:
elem_onehots = self._char_to_onehot(
temp_elem, onehot_dim=self.elem_num)
(outputs, hidden), alpha = self.structure_attention_cell(
......@@ -119,7 +127,7 @@ class TableAttentionHead(nn.Layer):
loc_concat = paddle.concat([output, loc_fea], axis=2)
loc_preds = self.loc_generator(loc_concat)
loc_preds = F.sigmoid(loc_preds)
return {'structure_probs':structure_probs, 'loc_preds':loc_preds}
return {'structure_probs': structure_probs, 'loc_preds': loc_preds}
class AttentionGRUCell(nn.Layer):
......
......@@ -11,6 +11,10 @@
# 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/clovaai/deep-text-recognition-benchmark/blob/master/modules/transformation.py
"""
from __future__ import absolute_import
from __future__ import division
......@@ -231,7 +235,8 @@ class GridGenerator(nn.Layer):
""" Return inv_delta_C which is needed to calculate T """
F = self.F
hat_eye = paddle.eye(F, dtype='float64') # F x F
hat_C = paddle.norm(C.reshape([1, F, 2]) - C.reshape([F, 1, 2]), axis=2) + hat_eye
hat_C = paddle.norm(
C.reshape([1, F, 2]) - C.reshape([F, 1, 2]), axis=2) + hat_eye
hat_C = (hat_C**2) * paddle.log(hat_C)
delta_C = paddle.concat( # F+3 x F+3
[
......
......@@ -11,7 +11,10 @@
# 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 refered from:
https://github.com/WenmuZhou/DBNet.pytorch/blob/master/post_processing/seg_detector_representer.py
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......@@ -190,7 +193,8 @@ class DBPostProcess(object):
class DistillationDBPostProcess(object):
def __init__(self, model_name=["student"],
def __init__(self,
model_name=["student"],
key=None,
thresh=0.3,
box_thresh=0.6,
......@@ -201,7 +205,8 @@ class DistillationDBPostProcess(object):
**kwargs):
self.model_name = model_name
self.key = key
self.post_process = DBPostProcess(thresh=thresh,
self.post_process = DBPostProcess(
thresh=thresh,
box_thresh=box_thresh,
max_candidates=max_candidates,
unclip_ratio=unclip_ratio,
......
"""
Locality aware nms.
This code is refered from: https://github.com/songdejia/EAST/blob/master/locality_aware_nms.py
"""
import numpy as np
......
......@@ -11,7 +11,10 @@
# 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/WenmuZhou/PytorchOCR/blob/master/torchocr/utils/logging.py
"""
import os
import sys
import logging
......
......@@ -187,7 +187,7 @@ def create_predictor(args, mode, logger):
"nearest_interp_v2_0.tmp_0": [1, 256, 2, 2]
}
max_input_shape = {
"x": [1, 3, 2000, 2000],
"x": [1, 3, 1280, 1280],
"conv2d_92.tmp_0": [1, 120, 400, 400],
"conv2d_91.tmp_0": [1, 24, 200, 200],
"conv2d_59.tmp_0": [1, 96, 400, 400],
......@@ -237,16 +237,16 @@ def create_predictor(args, mode, logger):
opt_input_shape.update(opt_pact_shape)
elif mode == "rec":
min_input_shape = {"x": [1, 3, 32, 10]}
max_input_shape = {"x": [args.rec_batch_num, 3, 32, 2000]}
max_input_shape = {"x": [args.rec_batch_num, 3, 32, 1024]}
opt_input_shape = {"x": [args.rec_batch_num, 3, 32, 320]}
elif mode == "cls":
min_input_shape = {"x": [1, 3, 48, 10]}
max_input_shape = {"x": [args.rec_batch_num, 3, 48, 2000]}
max_input_shape = {"x": [args.rec_batch_num, 3, 48, 1024]}
opt_input_shape = {"x": [args.rec_batch_num, 3, 48, 320]}
else:
min_input_shape = {"x": [1, 3, 10, 10]}
max_input_shape = {"x": [1, 3, 1000, 1000]}
opt_input_shape = {"x": [1, 3, 500, 500]}
max_input_shape = {"x": [1, 3, 512, 512]}
opt_input_shape = {"x": [1, 3, 256, 256]}
config.set_trt_dynamic_shape_info(min_input_shape, max_input_shape,
opt_input_shape)
......
Markdown is supported
0% .
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