opensource/PaddleOCR
3
0
Fork 0
mirror of https://github.com/PaddlePaddle/PaddleOCR.git synced 2026-10-30 22:55:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'dygraph' into dygraph

Browse Source
This commit is contained in:
MissPenguin
2021-07-13 10:04:21 +08:00
committed by GitHub
parent b1d31f8ace 2062b5097c
commit 8ad4e3071f
36 changed files with 1570 additions and 191 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StyleText/engine/text_drawers.py
+1 -1
View File
@@ -66,6 +66,7 @@ class StdTextDrawer(object):
corpus_list.append(corpus[0:i])
text_input_list.append(text_input)
corpus = corpus[i:]
i = 0
break
draw.text((char_x, 2), char_i, fill=(0, 0, 0), font=font)
char_x += char_size
@@ -78,7 +79,6 @@ class StdTextDrawer(object):
corpus_list.append(corpus[0:i])
text_input_list.append(text_input)
corpus = corpus[i:]
break
return corpus_list, text_input_list
configs/det/ch_ppocr_v2.1/ch_det_lite_train_cml_v2.1.yml
+202
View File
@@ -0,0 +1,202 @@
Global:
use_gpu: true
epoch_num: 1200
log_smooth_window: 20
print_batch_step: 2
save_model_dir: ./output/ch_db_mv3/
save_epoch_step: 1200
# evaluation is run every 5000 iterations after the 4000th iteration
eval_batch_step: [3000, 2000]
cal_metric_during_train: False
pretrained_model: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
checkpoints:
save_inference_dir:
use_visualdl: False
infer_img: doc/imgs_en/img_10.jpg
save_res_path: ./output/det_db/predicts_db.txt
Architecture:
name: DistillationModel
algorithm: Distillation
Models:
Student:
pretrained: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
freeze_params: false
return_all_feats: false
model_type: det
algorithm: DB
Backbone:
name: MobileNetV3
scale: 0.5
model_name: large
disable_se: True
Neck:
name: DBFPN
out_channels: 96
Head:
name: DBHead
k: 50
Student2:
pretrained: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
freeze_params: false
return_all_feats: false
model_type: det
algorithm: DB
Transform:
Backbone:
name: MobileNetV3
scale: 0.5
model_name: large
disable_se: True
Neck:
name: DBFPN
out_channels: 96
Head:
name: DBHead
k: 50
Teacher:
pretrained: ./pretrain_models/ch_ppocr_server_v2.0_det_train/best_accuracy
freeze_params: true
return_all_feats: false
model_type: det
algorithm: DB
Transform:
Backbone:
name: ResNet
layers: 18
Neck:
name: DBFPN
out_channels: 256
Head:
name: DBHead
k: 50
Loss:
name: CombinedLoss
loss_config_list:
- DistillationDilaDBLoss:
weight: 1.0
model_name_pairs:
- ["Student", "Teacher"]
- ["Student2", "Teacher"]
key: maps
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
- DistillationDMLLoss:
model_name_pairs:
- ["Student", "Student2"]
maps_name: "thrink_maps"
weight: 1.0
# act: None
model_name_pairs: ["Student", "Student2"]
key: maps
- DistillationDBLoss:
weight: 1.0
model_name_list: ["Student", "Student2"]
# key: maps
# name: DBLoss
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
name: Cosine
learning_rate: 0.001
warmup_epoch: 2
regularizer:
name: 'L2'
factor: 0
PostProcess:
name: DistillationDBPostProcess
model_name: ["Student", "Student2", "Teacher"]
# key: maps
thresh: 0.3
box_thresh: 0.6
max_candidates: 1000
unclip_ratio: 1.5
Metric:
name: DistillationMetric
base_metric_name: DetMetric
main_indicator: hmean
key: "Student"
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: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- IaaAugment:
augmenter_args:
- { 'type': Fliplr, 'args': { 'p': 0.5 } }
- { 'type': Affine, 'args': { 'rotate': [-10, 10] } }
- { 'type': Resize, 'args': { 'size': [0.5, 3] } }
- EastRandomCropData:
size: [960, 960]
max_tries: 50
keep_ratio: true
- MakeBorderMap:
shrink_ratio: 0.4
thresh_min: 0.3
thresh_max: 0.7
- MakeShrinkMap:
shrink_ratio: 0.4
min_text_size: 8
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask'] # the order of the dataloader list
loader:
shuffle: True
drop_last: False
batch_size_per_card: 8
num_workers: 4
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: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- DetResizeForTest:
# image_shape: [736, 1280]
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
loader:
shuffle: False
drop_last: False
batch_size_per_card: 1 # must be 1
num_workers: 2
configs/det/ch_ppocr_v2.1/ch_det_lite_train_distill_v2.1.yml
+174
View File
@@ -0,0 +1,174 @@
Global:
use_gpu: true
epoch_num: 1200
log_smooth_window: 20
print_batch_step: 2
save_model_dir: ./output/ch_db_mv3/
save_epoch_step: 1200
# evaluation is run every 5000 iterations after the 4000th iteration
eval_batch_step: [3000, 2000]
cal_metric_during_train: False
pretrained_model: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
checkpoints:
save_inference_dir:
use_visualdl: False
infer_img: doc/imgs_en/img_10.jpg
save_res_path: ./output/det_db/predicts_db.txt
Architecture:
name: DistillationModel
algorithm: Distillation
Models:
Student:
pretrained: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
freeze_params: false
return_all_feats: false
model_type: det
algorithm: DB
Backbone:
name: MobileNetV3
scale: 0.5
model_name: large
disable_se: True
Neck:
name: DBFPN
out_channels: 96
Head:
name: DBHead
k: 50
Teacher:
pretrained: ./pretrain_models/ch_ppocr_server_v2.0_det_train/best_accuracy
freeze_params: true
return_all_feats: false
model_type: det
algorithm: DB
Transform:
Backbone:
name: ResNet
layers: 18
Neck:
name: DBFPN
out_channels: 256
Head:
name: DBHead
k: 50
Loss:
name: CombinedLoss
loss_config_list:
- DistillationDilaDBLoss:
weight: 1.0
model_name_pairs:
- ["Student", "Teacher"]
key: maps
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
- DistillationDBLoss:
weight: 1.0
model_name_list: ["Student", "Teacher"]
# key: maps
name: DBLoss
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
name: Cosine
learning_rate: 0.001
warmup_epoch: 2
regularizer:
name: 'L2'
factor: 0
PostProcess:
name: DistillationDBPostProcess
model_name: ["Student", "Student2"]
key: head_out
thresh: 0.3
box_thresh: 0.6
max_candidates: 1000
unclip_ratio: 1.5
Metric:
name: DistillationMetric
base_metric_name: DetMetric
main_indicator: hmean
key: "Student"
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: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- IaaAugment:
augmenter_args:
- { 'type': Fliplr, 'args': { 'p': 0.5 } }
- { 'type': Affine, 'args': { 'rotate': [-10, 10] } }
- { 'type': Resize, 'args': { 'size': [0.5, 3] } }
- EastRandomCropData:
size: [960, 960]
max_tries: 50
keep_ratio: true
- MakeBorderMap:
shrink_ratio: 0.4
thresh_min: 0.3
thresh_max: 0.7
- MakeShrinkMap:
shrink_ratio: 0.4
min_text_size: 8
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask'] # the order of the dataloader list
loader:
shuffle: True
drop_last: False
batch_size_per_card: 8
num_workers: 4
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: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- DetResizeForTest:
# image_shape: [736, 1280]
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
loader:
shuffle: False
drop_last: False
batch_size_per_card: 1 # must be 1
num_workers: 2
configs/det/ch_ppocr_v2.1/ch_det_lite_train_dml_v2.1.yml
+176
View File
@@ -0,0 +1,176 @@
Global:
use_gpu: true
epoch_num: 1200
log_smooth_window: 20
print_batch_step: 2
save_model_dir: ./output/ch_db_mv3/
save_epoch_step: 1200
# evaluation is run every 5000 iterations after the 4000th iteration
eval_batch_step: [3000, 2000]
cal_metric_during_train: False
pretrained_model: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
checkpoints:
save_inference_dir:
use_visualdl: False
infer_img: doc/imgs_en/img_10.jpg
save_res_path: ./output/det_db/predicts_db.txt
Architecture:
name: DistillationModel
algorithm: Distillation
Models:
Student:
pretrained: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
freeze_params: false
return_all_feats: false
model_type: det
algorithm: DB
Backbone:
name: MobileNetV3
scale: 0.5
model_name: large
disable_se: True
Neck:
name: DBFPN
out_channels: 96
Head:
name: DBHead
k: 50
Student2:
pretrained: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
freeze_params: false
return_all_feats: false
model_type: det
algorithm: DB
Transform:
Backbone:
name: MobileNetV3
scale: 0.5
model_name: large
disable_se: True
Neck:
name: DBFPN
out_channels: 96
Head:
name: DBHead
k: 50
Loss:
name: CombinedLoss
loss_config_list:
- DistillationDMLLoss:
model_name_pairs:
- ["Student", "Student2"]
maps_name: "thrink_maps"
weight: 1.0
act: "softmax"
model_name_pairs: ["Student", "Student2"]
key: maps
- DistillationDBLoss:
weight: 1.0
model_name_list: ["Student", "Student2"]
# key: maps
name: DBLoss
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
name: Cosine
learning_rate: 0.001
warmup_epoch: 2
regularizer:
name: 'L2'
factor: 0
PostProcess:
name: DistillationDBPostProcess
model_name: ["Student", "Student2"]
key: head_out
thresh: 0.3
box_thresh: 0.6
max_candidates: 1000
unclip_ratio: 1.5
Metric:
name: DistillationMetric
base_metric_name: DetMetric
main_indicator: hmean
key: "Student"
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: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- IaaAugment:
augmenter_args:
- { 'type': Fliplr, 'args': { 'p': 0.5 } }
- { 'type': Affine, 'args': { 'rotate': [-10, 10] } }
- { 'type': Resize, 'args': { 'size': [0.5, 3] } }
- EastRandomCropData:
size: [960, 960]
max_tries: 50
keep_ratio: true
- MakeBorderMap:
shrink_ratio: 0.4
thresh_min: 0.3
thresh_max: 0.7
- MakeShrinkMap:
shrink_ratio: 0.4
min_text_size: 8
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask'] # the order of the dataloader list
loader:
shuffle: True
drop_last: False
batch_size_per_card: 8
num_workers: 4
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: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- DetResizeForTest:
# image_shape: [736, 1280]
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
loader:
shuffle: False
drop_last: False
batch_size_per_card: 1 # must be 1
num_workers: 2
configs/rec/ch_ppocr_v2.1/rec_chinese_lite_train_distillation_v2.1.yml
+28 -29
View File
@@ -17,7 +17,7 @@ Global:
character_type: ch
max_text_length: 25
infer_mode: false
use_space_char: false
use_space_char: true
distributed: true
save_res_path: ./output/rec/predicts_chinese_lite_distillation_v2.1.txt
@@ -27,48 +27,29 @@ Optimizer:
beta1: 0.9
beta2: 0.999
lr:
name: Cosine
learning_rate: 0.0005
name: Piecewise
decay_epochs : [700, 800]
values : [0.001, 0.0001]
warmup_epoch: 5
regularizer:
name: L2
factor: 1.0e-05
factor: 2.0e-05
Architecture:
model_type: &model_type "rec"
name: DistillationModel
algorithm: Distillation
Models:
Student:
pretrained:
freeze_params: false
return_all_feats: true
model_type: rec
algorithm: CRNN
Transform:
Backbone:
name: MobileNetV3
scale: 0.5
model_name: small
small_stride: [1, 2, 2, 2]
Neck:
name: SequenceEncoder
encoder_type: rnn
hidden_size: 64
Head:
name: CTCHead
mid_channels: 96
fc_decay: 0.00001
Teacher:
pretrained:
freeze_params: false
return_all_feats: true
model_type: rec
model_type: *model_type
algorithm: CRNN
Transform:
Backbone:
name: MobileNetV3
name: MobileNetV1Enhance
scale: 0.5
model_name: small
small_stride: [1, 2, 2, 2]
Neck:
name: SequenceEncoder
encoder_type: rnn
@@ -76,7 +57,25 @@ Architecture:
Head:
name: CTCHead
mid_channels: 96
fc_decay: 0.00001
fc_decay: 0.00002
Student:
pretrained:
freeze_params: false
return_all_feats: true
model_type: *model_type
algorithm: CRNN
Transform:
Backbone:
name: MobileNetV1Enhance
scale: 0.5
Neck:
name: SequenceEncoder
encoder_type: rnn
hidden_size: 64
Head:
name: CTCHead
mid_channels: 96
fc_decay: 0.00002
Loss:
deploy/hubserving/readme.md
+1 -1
View File
@@ -29,7 +29,7 @@ deploy/hubserving/ocr_system/
### 1. 准备环境
```shell
# 安装paddlehub
pip3 install paddlehub==1.8.3 --upgrade -i https://pypi.tuna.tsinghua.edu.cn/simple
pip3 install paddlehub==2.1.0 --upgrade -i https://pypi.tuna.tsinghua.edu.cn/simple
```
### 2. 下载推理模型
deploy/hubserving/readme_en.md
+1 -1
View File
@@ -30,7 +30,7 @@ The following steps take the 2-stage series service as an example. If only the d
### 1. Prepare the environment
```shell
# Install paddlehub
pip3 install paddlehub==1.8.3 --upgrade -i https://pypi.tuna.tsinghua.edu.cn/simple
pip3 install paddlehub==2.1.0 --upgrade -i https://pypi.tuna.tsinghua.edu.cn/simple
```
### 2. Download inference model
deploy/slim/quantization/export_model.py
+37 -14
View File
@@ -37,6 +37,17 @@ from paddleslim.dygraph.quant import QAT
from ppocr.data import build_dataloader
def export_single_model(quanter, model, infer_shape, save_path, logger):
quanter.save_quantized_model(
model,
save_path,
input_spec=[
paddle.static.InputSpec(
shape=[None] + infer_shape, dtype='float32')
])
logger.info('inference QAT model is saved to {}'.format(save_path))
def main():
############################################################################################################
# 1. quantization configs
@@ -76,14 +87,21 @@ def main():
# for rec algorithm
if hasattr(post_process_class, 'character'):
char_num = len(getattr(post_process_class, 'character'))
config['Architecture']["Head"]['out_channels'] = char_num
if config['Architecture']["algorithm"] in ["Distillation",
]: # distillation model
for key in config['Architecture']["Models"]:
config['Architecture']["Models"][key]["Head"][
'out_channels'] = char_num
else: # base rec model
config['Architecture']["Head"]['out_channels'] = char_num
model = build_model(config['Architecture'])
# get QAT model
quanter = QAT(config=quant_config)
quanter.quantize(model)
init_model(config, model, logger)
init_model(config, model)
model.eval()
# build metric
@@ -92,25 +110,30 @@ def main():
# build dataloader
valid_dataloader = build_dataloader(config, 'Eval', device, logger)
use_srn = config['Architecture']['algorithm'] == "SRN"
model_type = config['Architecture']['model_type']
# start eval
metirc = program.eval(model, valid_dataloader, post_process_class,
eval_class)
metric = program.eval(model, valid_dataloader, post_process_class,
eval_class, model_type, use_srn)
logger.info('metric eval ***************')
for k, v in metirc.items():
for k, v in metric.items():
logger.info('{}:{}'.format(k, v))
save_path = '{}/inference'.format(config['Global']['save_inference_dir'])
infer_shape = [3, 32, 100] if config['Architecture'][
'model_type'] != "det" else [3, 640, 640]
quanter.save_quantized_model(
model,
save_path,
input_spec=[
paddle.static.InputSpec(
shape=[None] + infer_shape, dtype='float32')
])
logger.info('inference QAT model is saved to {}'.format(save_path))
save_path = config["Global"]["save_inference_dir"]
arch_config = config["Architecture"]
if arch_config["algorithm"] in ["Distillation", ]: # distillation model
for idx, name in enumerate(model.model_name_list):
sub_model_save_path = os.path.join(save_path, name, "inference")
export_single_model(quanter, model.model_list[idx], infer_shape,
sub_model_save_path, logger)
else:
save_path = os.path.join(save_path, "inference")
export_single_model(quanter, model, infer_shape, save_path, logger)
if __name__ == "__main__":
deploy/slim/quantization/quant.py
+10 -3
View File
@@ -109,9 +109,18 @@ def main(config, device, logger, vdl_writer):
# for rec algorithm
if hasattr(post_process_class, 'character'):
char_num = len(getattr(post_process_class, 'character'))
config['Architecture']["Head"]['out_channels'] = char_num
if config['Architecture']["algorithm"] in ["Distillation",
]: # distillation model
for key in config['Architecture']["Models"]:
config['Architecture']["Models"][key]["Head"][
'out_channels'] = char_num
else: # base rec model
config['Architecture']["Head"]['out_channels'] = char_num
model = build_model(config['Architecture'])
quanter = QAT(config=quant_config, act_preprocess=PACT)
quanter.quantize(model)
if config['Global']['distributed']:
model = paddle.DataParallel(model)
@@ -132,8 +141,6 @@ def main(config, device, logger, vdl_writer):
logger.info('train dataloader has {} iters, valid dataloader has {} iters'.
format(len(train_dataloader), len(valid_dataloader)))
quanter = QAT(config=quant_config, act_preprocess=PACT)
quanter.quantize(model)
# start train
program.train(config, train_dataloader, valid_dataloader, device, model,
doc/doc_ch/knowledge_distillation.md
+251
View File
@@ -0,0 +1,251 @@
# 知识蒸馏
## 1. 简介
### 1.1 知识蒸馏介绍
近年来，深度神经网络在计算机视觉、自然语言处理等领域被验证是一种极其有效的解决问题的方法。通过构建合适的神经网络，加以训练，最终网络模型的性能指标基本上都会超过传统算法。
在数据量足够大的情况下，通过合理构建网络模型的方式增加其参数量，可以显著改善模型性能，但是这又带来了模型复杂度急剧提升的问题。大模型在实际场景中使用的成本较高。
深度神经网络一般有较多的参数冗余，目前有几种主要的方法对模型进行压缩，减小其参数量。如裁剪、量化、知识蒸馏等，其中知识蒸馏是指使用教师模型(teacher model)去指导学生模型(student model)学习特定任务，保证小模型在参数量不变的情况下，得到比较大的性能提升。
此外，在知识蒸馏任务中，也衍生出了互学习的模型训练方法，论文[Deep Mutual Learning](https://arxiv.org/abs/1706.00384)中指出，使用两个完全相同的模型在训练的过程中互相监督，可以达到比单个模型训练更好的效果。
### 1.2 PaddleOCR知识蒸馏简介
无论是大模型蒸馏小模型，还是小模型之间互相学习，更新参数，他们本质上是都是不同模型之间输出或者特征图(feature map)之间的相互监督，区别仅在于 (1) 模型是否需要固定参数。(2) 模型是否需要加载预训练模型。
对于大模型蒸馏小模型的情况，大模型一般需要加载预训练模型并固定参数；对于小模型之间互相蒸馏的情况，小模型一般都不加载预训练模型，参数也都是可学习的状态。
在知识蒸馏任务中，不只有2个模型之间进行蒸馏的情况，多个模型之间互相学习的情况也非常普遍。因此在知识蒸馏代码框架中，也有必要支持该种类别的蒸馏方法。
PaddleOCR中集成了知识蒸馏的算法，具体地，有以下几个主要的特点：
- 支持任意网络的互相学习，不要求子网络结构完全一致或者具有预训练模型；同时子网络数量也没有任何限制，只需要在配置文件中添加即可。
- 支持loss函数通过配置文件任意配置，不仅可以使用某种loss，也可以使用多种loss的组合
- 支持知识蒸馏训练、预测、评估与导出等所有模型相关的环境，方便使用与部署。
通过知识蒸馏，在中英文通用文字识别任务中，不增加任何预测耗时的情况下，可以给模型带来3%以上的精度提升，结合学习率调整策略以及模型结构微调策略，最终提升提升超过5%。
## 2. 配置文件解析
在知识蒸馏训练的过程中，数据预处理、优化器、学习率、全局的一些属性没有任何变化。模型结构、损失函数、后处理、指标计算等模块的配置文件需要进行微调。
下面以识别与检测的知识蒸馏配置文件为例，对知识蒸馏的训练与配置进行解析。
### 2.1 识别配置文件解析
配置文件在[rec_chinese_lite_train_distillation_v2.1.yml](../../configs/rec/ch_ppocr_v2.1/rec_chinese_lite_train_distillation_v2.1.yml)。
#### 2.1.1 模型结构
知识蒸馏任务中，模型结构配置如下所示。
```yaml
Architecture:
model_type: &model_type "rec" # 模型类别，rec、det等，每个子网络的的模型类别都与
name: DistillationModel # 结构名称，蒸馏任务中，为DistillationModel，用于构建对应的结构
algorithm: Distillation # 算法名称
Models: # 模型，包含子网络的配置信息
Teacher: # 子网络名称，至少需要包含`pretrained`与`freeze_params`信息，其他的参数为子网络的构造参数
pretrained: # 该子网络是否需要加载预训练模型
freeze_params: false # 是否需要固定参数
return_all_feats: true # 子网络的参数，表示是否需要返回所有的features，如果为False，则只返回最后的输出
model_type: *model_type # 模型类别
algorithm: CRNN # 子网络的算法名称，该子网络剩余参与均为构造参数，与普通的模型训练配置一致
Transform:
Backbone:
name: MobileNetV1Enhance
scale: 0.5
Neck:
name: SequenceEncoder
encoder_type: rnn
hidden_size: 64
Head:
name: CTCHead
mid_channels: 96
fc_decay: 0.00002
Student: # 另外一个子网络，这里给的是DML的蒸馏示例，两个子网络结构相同，均需要学习参数
pretrained: # 下面的组网参数同上
freeze_params: false
return_all_feats: true
model_type: *model_type
algorithm: CRNN
Transform:
Backbone:
name: MobileNetV1Enhance
scale: 0.5
Neck:
name: SequenceEncoder
encoder_type: rnn
hidden_size: 64
Head:
name: CTCHead
mid_channels: 96
fc_decay: 0.00002
```
当然，这里如果希望添加更多的子网络进行训练，也可以按照`Student`与`Teacher`的添加方式，在配置文件中添加相应的字段。比如说如果希望有3个模型互相监督，共同训练，那么`Architecture`可以写为如下格式。
```yaml
Architecture:
model_type: &model_type "rec"
name: DistillationModel
algorithm: Distillation
Models:
Teacher:
pretrained:
freeze_params: false
return_all_feats: true
model_type: *model_type
algorithm: CRNN
Transform:
Backbone:
name: MobileNetV1Enhance
scale: 0.5
Neck:
name: SequenceEncoder
encoder_type: rnn
hidden_size: 64
Head:
name: CTCHead
mid_channels: 96
fc_decay: 0.00002
Student:
pretrained:
freeze_params: false
return_all_feats: true
model_type: *model_type
algorithm: CRNN
Transform:
Backbone:
name: MobileNetV1Enhance
scale: 0.5
Neck:
name: SequenceEncoder
encoder_type: rnn
hidden_size: 64
Head:
name: CTCHead
mid_channels: 96
fc_decay: 0.00002
Student2: # 知识蒸馏任务中引入的新的子网络，其他部分与上述配置相同
pretrained:
freeze_params: false
return_all_feats: true
model_type: *model_type
algorithm: CRNN
Transform:
Backbone:
name: MobileNetV1Enhance
scale: 0.5
Neck:
name: SequenceEncoder
encoder_type: rnn
hidden_size: 64
Head:
name: CTCHead
mid_channels: 96
fc_decay: 0.00002
```
最终该模型训练时，包含3个子网络：`Teacher`, `Student`, `Student2`。
蒸馏模型`DistillationModel`类的具体实现代码可以参考[distillation_model.py](../../ppocr/modeling/architectures/distillation_model.py)。
最终模型`forward`输出为一个字典，key为所有的子网络名称，例如这里为`Student`与`Teacher`，value为对应子网络的输出，可以为`Tensor`（只返回该网络的最后一层）和`dict`（也返回了中间的特征信息）。
在识别任务中，为了添加更多损失函数，保证蒸馏方法的可扩展性，将每个子网络的输出保存为`dict`，其中包含子模块输出。以该识别模型为例，每个子网络的输出结果均为`dict`，key包含`backbone_out`,`neck_out`, `head_out`，`value`为对应模块的tensor，最终对于上述配置文件，`DistillationModel`的输出格式如下。
```json
{
"Teacher": {
"backbone_out": tensor,
"neck_out": tensor,
"head_out": tensor,
},
"Student": {
"backbone_out": tensor,
"neck_out": tensor,
"head_out": tensor,
}
}
```
#### 2.1.2 损失函数
知识蒸馏任务中，损失函数配置如下所示。
```yaml
Loss:
name: CombinedLoss # 损失函数名称，基于改名称，构建用于损失函数的类
loss_config_list: # 损失函数配置文件列表，为CombinedLoss的必备函数
- DistillationCTCLoss: # 基于蒸馏的CTC损失函数，继承自标准的CTC loss
weight: 1.0 # 损失函数的权重，loss_config_list中，每个损失函数的配置都必须包含该字段
model_name_list: ["Student", "Teacher"] # 对于蒸馏模型的预测结果，提取这两个子网络的输出，与gt计算CTC loss
key: head_out # 取子网络输出dict中，该key对应的tensor
- DistillationDMLLoss: # 蒸馏的DML损失函数，继承自标准的DMLLoss
weight: 1.0 # 权重
act: "softmax" # 激活函数，对输入使用激活函数处理，可以为softmax, sigmoid或者为None，默认为None
model_name_pairs: # 用于计算DML loss的子网络名称对，如果希望计算其他子网络的DML loss，可以在列表下面继续填充
- ["Student", "Teacher"]
key: head_out # 取子网络输出dict中，该key对应的tensor
- DistillationDistanceLoss: # 蒸馏的距离损失函数
weight: 1.0 # 权重
mode: "l2" # 距离计算方法，目前支持l1, l2, smooth_l1
model_name_pairs: # 用于计算distance loss的子网络名称对
- ["Student", "Teacher"]
key: backbone_out # 取子网络输出dict中，该key对应的tensor
```
上述损失函数中，所有的蒸馏损失函数均继承自标准的损失函数类，主要功能为: 对蒸馏模型的输出进行解析，找到用于计算损失的中间节点(tensor)，再使用标准的损失函数类去计算。
以上述配置为例，最终蒸馏训练的损失函数包含下面3个部分。
- `Student`和`Teacher`的最终输出(`head_out`)与gt的CTC loss，权重为1。在这里因为2个子网络都需要更新参数，因此2者都需要计算与g的loss。
- `Student`和`Teacher`的最终输出(`head_out`)之间的DML loss，权重为1。
- `Student`和`Teacher`的骨干网络输出(`backbone_out`)之间的l2 loss，权重为1。
关于`CombinedLoss`更加具体的实现可以参考: [combined_loss.py](../../ppocr/losses/combined_loss.py#L23)。关于`DistillationCTCLoss`等蒸馏损失函数更加具体的实现可以参考[distillation_loss.py](../../ppocr/losses/distillation_loss.py)。
#### 2.1.3 后处理
知识蒸馏任务中，后处理配置如下所示。
```yaml
PostProcess:
name: DistillationCTCLabelDecode # 蒸馏任务的CTC解码后处理，继承自标准的CTCLabelDecode类
model_name: ["Student", "Teacher"] # 对于蒸馏模型的预测结果，提取这两个子网络的输出，进行解码
key: head_out # 取子网络输出dict中，该key对应的tensor
```
以上述配置为例，最终会同时计算`Student`和`Teahcer` 2个子网络的CTC解码输出，返回一个`dict`，`key`为用于处理的子网络名称，`value`为用于处理的子网络列表。
关于`DistillationCTCLabelDecode`更加具体的实现可以参考: [rec_postprocess.py](../../ppocr/postprocess/rec_postprocess.py#L128)
#### 2.1.4 指标计算
知识蒸馏任务中，指标计算配置如下所示。
```yaml
Metric:
name: DistillationMetric # 蒸馏任务的CTC解码后处理，继承自标准的CTCLabelDecode类
base_metric_name: RecMetric # 指标计算的基类，对于模型的输出，会基于该类，计算指标
main_indicator: acc # 指标的名称
key: "Student" # 选取该子网络的 main_indicator 作为作为保存保存best model的判断标准
```
以上述配置为例，最终会使用`Student`子网络的acc指标作为保存best model的判断指标，同时，日志中也会打印出所有子网络的acc指标。
关于`DistillationMetric`更加具体的实现可以参考: [distillation_metric.py](../../ppocr/metrics/distillation_metric.py#L24)。
### 2.2 检测配置文件解析
* coming soon!
doc/doc_ch/recognition.md
+3 -1
View File
@@ -375,7 +375,9 @@ PaddleOCR目前已支持80种（除中文外）语种识别，`configs/rec/multi
更多支持语种请参考: [多语言模型](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.1/doc/doc_ch/multi_languages.md#%E8%AF%AD%E7%A7%8D%E7%BC%A9%E5%86%99)
多语言模型训练方式与中文模型一致，训练数据集均为100w的合成数据，少量的字体可以在 [百度网盘](https://pan.baidu.com/s/1bS_u207Rm7YbY33wOECKDA) 上下载，提取码：frgi。
多语言模型训练方式与中文模型一致，训练数据集均为100w的合成数据，少量的字体可以通过下面两种方式下载。
* [百度网盘](https://pan.baidu.com/s/1bS_u207Rm7YbY33wOECKDA)。提取码：frgi。
* [google drive](https://drive.google.com/file/d/18cSWX7wXSy4G0tbKJ0d9PuIaiwRLHpjA/view)
如您希望在现有模型效果的基础上调优，请参考下列说明修改配置文件：
doc/doc_en/recognition_en.md
+3 -1
View File
@@ -375,7 +375,9 @@ Currently, the multi-language algorithms supported by PaddleOCR are:
For more supported languages, please refer to : [Multi-language model](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.1/doc/doc_en/multi_languages_en.md#4-support-languages-and-abbreviations)
The multi-language model training method is the same as the Chinese model. The training data set is 100w synthetic data. A small amount of fonts and test data can be downloaded on [Baidu Netdisk](https://pan.baidu.com/s/1bS_u207Rm7YbY33wOECKDA),Extraction code:frgi.
The multi-language model training method is the same as the Chinese model. The training data set is 100w synthetic data. A small amount of fonts and test data can be downloaded using the following two methods.
* [Baidu Netdisk](https://pan.baidu.com/s/1bS_u207Rm7YbY33wOECKDA),Extraction code:frgi.
* [Google drive](https://drive.google.com/file/d/18cSWX7wXSy4G0tbKJ0d9PuIaiwRLHpjA/view)
If you want to finetune on the basis of the existing model effect, please refer to the following instructions to modify the configuration file:
doc/joinus.PNG
BIN
View File
Binary file not shown.
Side by Side Swipe Overlay

Before

Width:  |  Height:  |  Size: 205 KiB

After

Width:  |  Height:  |  Size: 188 KiB

ppocr/data/imaug/copy_paste.py
+6 -4
View File
@@ -73,11 +73,14 @@ class CopyPaste(object):
box_img_pil = Image.fromarray(box_img).convert('RGBA')
src_w, src_h = src_img.size
box_w, box_h = box_img_pil.size
if box_w > src_w or box_h > src_h:
return src_img, None
angle = np.random.randint(0, 360)
box = np.array([[[0, 0], [box_w, 0], [box_w, box_h], [0, box_h]]])
box = rotate_bbox(box_img, box, angle)[0]
box_img_pil = box_img_pil.rotate(angle, expand=1)
box_w, box_h = box_img_pil.width, box_img_pil.height
if src_w - box_w < 0 or src_h - box_h < 0:
return src_img, None
paste_x, paste_y = self.select_coord(src_polys, box, src_w - box_w,
src_h - box_h)
@@ -85,7 +88,6 @@ class CopyPaste(object):
return src_img, None
box[:, 0] += paste_x
box[:, 1] += paste_y
box_img_pil = box_img_pil.rotate(angle, expand=1)
r, g, b, A = box_img_pil.split()
src_img.paste(box_img_pil, (paste_x, paste_y), mask=A)
@@ -105,7 +107,7 @@ class CopyPaste(object):
num_poly_in_rect = 0
for poly in src_polys:
if not is_poly_outside_rect(poly, xmax1, ymin1,
if not is_poly_outside_rect(poly, xmin1, ymin1,
xmax1 - xmin1, ymax1 - ymin1):
num_poly_in_rect += 1
break
ppocr/losses/basic_loss.py
+31 -6
View File
@@ -54,6 +54,27 @@ class CELoss(nn.Layer):
return loss
class KLJSLoss(object):
def __init__(self, mode='kl'):
assert mode in ['kl', 'js', 'KL', 'JS'], "mode can only be one of ['kl', 'js', 'KL', 'JS']"
self.mode = mode
def __call__(self, p1, p2, reduction="mean"):
loss = paddle.multiply(p2, paddle.log( (p2+1e-5)/(p1+1e-5) + 1e-5))
if self.mode.lower() == "js":
loss += paddle.multiply(p1, paddle.log((p1+1e-5)/(p2+1e-5) + 1e-5))
loss *= 0.5
if reduction == "mean":
loss = paddle.mean(loss, axis=[1,2])
elif reduction=="none" or reduction is None:
return loss
else:
loss = paddle.sum(loss, axis=[1,2])
return loss
class DMLLoss(nn.Layer):
"""
DMLLoss
@@ -69,17 +90,21 @@ class DMLLoss(nn.Layer):
self.act = nn.Sigmoid()
else:
self.act = None
self.jskl_loss = KLJSLoss(mode="js")
def forward(self, out1, out2):
if self.act is not None:
out1 = self.act(out1)
out2 = self.act(out2)
log_out1 = paddle.log(out1)
log_out2 = paddle.log(out2)
loss = (F.kl_div(
log_out1, out2, reduction='batchmean') + F.kl_div(
log_out2, out1, reduction='batchmean')) / 2.0
if len(out1.shape) < 2:
log_out1 = paddle.log(out1)
log_out2 = paddle.log(out2)
loss = (F.kl_div(
log_out1, out2, reduction='batchmean') + F.kl_div(
log_out2, out1, reduction='batchmean')) / 2.0
else:
loss = self.jskl_loss(out1, out2)
return loss
ppocr/losses/combined_loss.py
+8 -7
View File
@@ -17,7 +17,7 @@ import paddle.nn as nn
from .distillation_loss import DistillationCTCLoss
from .distillation_loss import DistillationDMLLoss
from .distillation_loss import DistillationDistanceLoss
from .distillation_loss import DistillationDistanceLoss, DistillationDBLoss, DistillationDilaDBLoss
class CombinedLoss(nn.Layer):
@@ -44,15 +44,16 @@ class CombinedLoss(nn.Layer):
def forward(self, input, batch, **kargs):
loss_dict = {}
loss_all = 0.
for idx, loss_func in enumerate(self.loss_func):
loss = loss_func(input, batch, **kargs)
if isinstance(loss, paddle.Tensor):
loss = {"loss_{}_{}".format(str(loss), idx): loss}
weight = self.loss_weight[idx]
loss = {
"{}_{}".format(key, idx): loss[key] * weight
for key in loss
}
loss_dict.update(loss)
loss_dict["loss"] = paddle.add_n(list(loss_dict.values()))
for key in loss.keys():
if key == "loss":
loss_all += loss[key] * weight
else:
loss_dict["{}_{}".format(key, idx)] = loss[key]
loss_dict["loss"] = loss_all
return loss_dict
ppocr/losses/distillation_loss.py
+171 -9
View File
@@ -14,23 +14,76 @@
import paddle
import paddle.nn as nn
import numpy as np
import cv2
from .rec_ctc_loss import CTCLoss
from .basic_loss import DMLLoss
from .basic_loss import DistanceLoss
from .det_db_loss import DBLoss
from .det_basic_loss import BalanceLoss, MaskL1Loss, DiceLoss
def _sum_loss(loss_dict):
if "loss" in loss_dict.keys():
return loss_dict
else:
loss_dict["loss"] = 0.
for k, value in loss_dict.items():
if k == "loss":
continue
else:
loss_dict["loss"] += value
return loss_dict
class DistillationDMLLoss(DMLLoss):
"""
"""
def __init__(self, model_name_pairs=[], act=None, key=None,
name="loss_dml"):
def __init__(self,
model_name_pairs=[],
act=None,
key=None,
maps_name=None,
name="dml"):
super().__init__(act=act)
assert isinstance(model_name_pairs, list)
self.key = key
self.model_name_pairs = model_name_pairs
self.model_name_pairs = self._check_model_name_pairs(model_name_pairs)
self.name = name
self.maps_name = self._check_maps_name(maps_name)
def _check_model_name_pairs(self, model_name_pairs):
if not isinstance(model_name_pairs, list):
return []
elif isinstance(model_name_pairs[0], list) and isinstance(model_name_pairs[0][0], str):
return model_name_pairs
else:
return [model_name_pairs]
def _check_maps_name(self, maps_name):
if maps_name is None:
return None
elif type(maps_name) == str:
return [maps_name]
elif type(maps_name) == list:
return [maps_name]
else:
return None
def _slice_out(self, outs):
new_outs = {}
for k in self.maps_name:
if k == "thrink_maps":
new_outs[k] = outs[:, 0, :, :]
elif k == "threshold_maps":
new_outs[k] = outs[:, 1, :, :]
elif k == "binary_maps":
new_outs[k] = outs[:, 2, :, :]
else:
continue
return new_outs
def forward(self, predicts, batch):
loss_dict = dict()
@@ -40,13 +93,30 @@ class DistillationDMLLoss(DMLLoss):
if self.key is not None:
out1 = out1[self.key]
out2 = out2[self.key]
loss = super().forward(out1, out2)
if isinstance(loss, dict):
for key in loss:
loss_dict["{}_{}_{}_{}".format(key, pair[0], pair[1],
idx)] = loss[key]
if self.maps_name is None:
loss = super().forward(out1, out2)
if isinstance(loss, dict):
for key in loss:
loss_dict["{}_{}_{}_{}".format(key, pair[0], pair[1],
idx)] = loss[key]
else:
loss_dict["{}_{}".format(self.name, idx)] = loss
else:
loss_dict["{}_{}".format(self.name, idx)] = loss
outs1 = self._slice_out(out1)
outs2 = self._slice_out(out2)
for _c, k in enumerate(outs1.keys()):
loss = super().forward(outs1[k], outs2[k])
if isinstance(loss, dict):
for key in loss:
loss_dict["{}_{}_{}_{}_{}".format(key, pair[
0], pair[1], map_name, idx)] = loss[key]
else:
loss_dict["{}_{}_{}".format(self.name, self.maps_name[_c],
idx)] = loss
loss_dict = _sum_loss(loss_dict)
return loss_dict
@@ -73,6 +143,98 @@ class DistillationCTCLoss(CTCLoss):
return loss_dict
class DistillationDBLoss(DBLoss):
def __init__(self,
model_name_list=[],
balance_loss=True,
main_loss_type='DiceLoss',
alpha=5,
beta=10,
ohem_ratio=3,
eps=1e-6,
name="db",
**kwargs):
super().__init__()
self.model_name_list = model_name_list
self.name = name
self.key = None
def forward(self, predicts, batch):
loss_dict = {}
for idx, model_name in enumerate(self.model_name_list):
out = predicts[model_name]
if self.key is not None:
out = out[self.key]
loss = super().forward(out, batch)
if isinstance(loss, dict):
for key in loss.keys():
if key == "loss":
continue
name = "{}_{}_{}".format(self.name, model_name, key)
loss_dict[name] = loss[key]
else:
loss_dict["{}_{}".format(self.name, model_name)] = loss
loss_dict = _sum_loss(loss_dict)
return loss_dict
class DistillationDilaDBLoss(DBLoss):
def __init__(self,
model_name_pairs=[],
key=None,
balance_loss=True,
main_loss_type='DiceLoss',
alpha=5,
beta=10,
ohem_ratio=3,
eps=1e-6,
name="dila_dbloss"):
super().__init__()
self.model_name_pairs = model_name_pairs
self.name = name
self.key = key
def forward(self, predicts, batch):
loss_dict = dict()
for idx, pair in enumerate(self.model_name_pairs):
stu_outs = predicts[pair[0]]
tch_outs = predicts[pair[1]]
if self.key is not None:
stu_preds = stu_outs[self.key]
tch_preds = tch_outs[self.key]
stu_shrink_maps = stu_preds[:, 0, :, :]
stu_binary_maps = stu_preds[:, 2, :, :]
# dilation to teacher prediction
dilation_w = np.array([[1, 1], [1, 1]])
th_shrink_maps = tch_preds[:, 0, :, :]
th_shrink_maps = th_shrink_maps.numpy() > 0.3 # thresh = 0.3
dilate_maps = np.zeros_like(th_shrink_maps).astype(np.float32)
for i in range(th_shrink_maps.shape[0]):
dilate_maps[i] = cv2.dilate(
th_shrink_maps[i, :, :].astype(np.uint8), dilation_w)
th_shrink_maps = paddle.to_tensor(dilate_maps)
label_threshold_map, label_threshold_mask, label_shrink_map, label_shrink_mask = batch[
1:]
# calculate the shrink map loss
bce_loss = self.alpha * self.bce_loss(
stu_shrink_maps, th_shrink_maps, label_shrink_mask)
loss_binary_maps = self.dice_loss(stu_binary_maps, th_shrink_maps,
label_shrink_mask)
# k = f"{self.name}_{pair[0]}_{pair[1]}"
k = "{}_{}_{}".format(self.name, pair[0], pair[1])
loss_dict[k] = bce_loss + loss_binary_maps
loss_dict = _sum_loss(loss_dict)
return loss_dict
class DistillationDistanceLoss(DistanceLoss):
"""
"""
ppocr/metrics/det_metric.py
+1
View File
@@ -55,6 +55,7 @@ class DetMetric(object):
result = self.evaluator.evaluate_image(gt_info_list, det_info_list)
self.results.append(result)
def get_metric(self):
"""
return metrics {
ppocr/metrics/distillation_metric.py
+4 -7
View File
@@ -24,8 +24,8 @@ from .cls_metric import ClsMetric
class DistillationMetric(object):
def __init__(self,
key=None,
base_metric_name="RecMetric",
main_indicator='acc',
base_metric_name=None,
main_indicator=None,
**kwargs):
self.main_indicator = main_indicator
self.key = key
@@ -42,16 +42,13 @@ class DistillationMetric(object):
main_indicator=self.main_indicator, **self.kwargs)
self.metrics[key].reset()
def __call__(self, preds, *args, **kwargs):
def __call__(self, preds, batch, **kwargs):
assert isinstance(preds, dict)
if self.metrics is None:
self._init_metrcis(preds)
output = dict()
for key in preds:
metric = self.metrics[key].__call__(preds[key], *args, **kwargs)
for sub_key in metric:
output["{}_{}".format(key, sub_key)] = metric[sub_key]
return output
self.metrics[key].__call__(preds[key], batch, **kwargs)
def get_metric(self):
"""
ppocr/modeling/architectures/base_model.py
+4 -1
View File
@@ -79,7 +79,10 @@ class BaseModel(nn.Layer):
x = self.neck(x)
y["neck_out"] = x
x = self.head(x, targets=data)
y["head_out"] = x
if isinstance(x, dict):
y.update(x)
else:
y["head_out"] = x
if self.return_all_feats:
return y
else:
ppocr/modeling/architectures/distillation_model.py
+2 -2
View File
@@ -21,7 +21,7 @@ from ppocr.modeling.backbones import build_backbone
from ppocr.modeling.necks import build_neck
from ppocr.modeling.heads import build_head
from .base_model import BaseModel
from ppocr.utils.save_load import init_model
from ppocr.utils.save_load import init_model, load_pretrained_params
__all__ = ['DistillationModel']
@@ -46,7 +46,7 @@ class DistillationModel(nn.Layer):
pretrained = model_config.pop("pretrained")
model = BaseModel(model_config)
if pretrained is not None:
init_model(model, path=pretrained)
model = load_pretrained_params(model, pretrained)
if freeze_params:
for param in model.parameters():
param.trainable = False
ppocr/modeling/backbones/__init__.py
+13 -10
View File
@@ -12,33 +12,36 @@
# See the License for the specific language governing permissions and
# limitations under the License.
__all__ = ['build_backbone']
__all__ = ["build_backbone"]
def build_backbone(config, model_type):
if model_type == 'det':
if model_type == "det":
from .det_mobilenet_v3 import MobileNetV3
from .det_resnet_vd import ResNet
from .det_resnet_vd_sast import ResNet_SAST
support_dict = ['MobileNetV3', 'ResNet', 'ResNet_SAST']
elif model_type == 'rec' or model_type == 'cls':
support_dict = ["MobileNetV3", "ResNet", "ResNet_SAST"]
elif model_type == "rec" or model_type == "cls":
from .rec_mobilenet_v3 import MobileNetV3
from .rec_resnet_vd import ResNet
from .rec_resnet_fpn import ResNetFPN
support_dict = ['MobileNetV3', 'ResNet', 'ResNetFPN']
elif model_type == 'e2e':
from .rec_mv1_enhance import MobileNetV1Enhance
support_dict = [
"MobileNetV1Enhance", "MobileNetV3", "ResNet", "ResNetFPN"
]
elif model_type == "e2e":
from .e2e_resnet_vd_pg import ResNet
support_dict = ['ResNet']
support_dict = ["ResNet"]
elif model_type == "table":
from .table_resnet_vd import ResNet
from .table_mobilenet_v3 import MobileNetV3
support_dict = ['ResNet', 'MobileNetV3']
support_dict = ["ResNet", "MobileNetV3"]
else:
raise NotImplementedError
module_name = config.pop('name')
module_name = config.pop("name")
assert module_name in support_dict, Exception(
'when model typs is {}, backbone only support {}'.format(model_type,
"when model typs is {}, backbone only support {}".format(model_type,
support_dict))
module_class = eval(module_name)(**config)
return module_class
ppocr/modeling/backbones/rec_mv1_enhance.py
+256
View File
@@ -0,0 +1,256 @@
# 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.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import paddle
from paddle import ParamAttr
import paddle.nn as nn
import paddle.nn.functional as F
from paddle.nn import Conv2D, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2D, MaxPool2D, AvgPool2D
from paddle.nn.initializer import KaimingNormal
import math
import numpy as np
import paddle
from paddle import ParamAttr, reshape, transpose, concat, split
import paddle.nn as nn
import paddle.nn.functional as F
from paddle.nn import Conv2D, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2D, MaxPool2D, AvgPool2D
from paddle.nn.initializer import KaimingNormal
import math
from paddle.nn.functional import hardswish, hardsigmoid
from paddle.regularizer import L2Decay
class ConvBNLayer(nn.Layer):
def __init__(self,
num_channels,
filter_size,
num_filters,
stride,
padding,
channels=None,
num_groups=1,
act='hard_swish'):
super(ConvBNLayer, self).__init__()
self._conv = Conv2D(
in_channels=num_channels,
out_channels=num_filters,
kernel_size=filter_size,
stride=stride,
padding=padding,
groups=num_groups,
weight_attr=ParamAttr(initializer=KaimingNormal()),
bias_attr=False)
self._batch_norm = BatchNorm(
num_filters,
act=act,
param_attr=ParamAttr(regularizer=L2Decay(0.0)),
bias_attr=ParamAttr(regularizer=L2Decay(0.0)))
def forward(self, inputs):
y = self._conv(inputs)
y = self._batch_norm(y)
return y
class DepthwiseSeparable(nn.Layer):
def __init__(self,
num_channels,
num_filters1,
num_filters2,
num_groups,
stride,
scale,
dw_size=3,
padding=1,
use_se=False):
super(DepthwiseSeparable, self).__init__()
self.use_se = use_se
self._depthwise_conv = ConvBNLayer(
num_channels=num_channels,
num_filters=int(num_filters1 * scale),
filter_size=dw_size,
stride=stride,
padding=padding,
num_groups=int(num_groups * scale))
if use_se:
self._se = SEModule(int(num_filters1 * scale))
self._pointwise_conv = ConvBNLayer(
num_channels=int(num_filters1 * scale),
filter_size=1,
num_filters=int(num_filters2 * scale),
stride=1,
padding=0)
def forward(self, inputs):
y = self._depthwise_conv(inputs)
if self.use_se:
y = self._se(y)
y = self._pointwise_conv(y)
return y
class MobileNetV1Enhance(nn.Layer):
def __init__(self, in_channels=3, scale=0.5, **kwargs):
super().__init__()
self.scale = scale
self.block_list = []
self.conv1 = ConvBNLayer(
num_channels=3,
filter_size=3,
channels=3,
num_filters=int(32 * scale),
stride=2,
padding=1)
conv2_1 = DepthwiseSeparable(
num_channels=int(32 * scale),
num_filters1=32,
num_filters2=64,
num_groups=32,
stride=1,
scale=scale)
self.block_list.append(conv2_1)
conv2_2 = DepthwiseSeparable(
num_channels=int(64 * scale),
num_filters1=64,
num_filters2=128,
num_groups=64,
stride=1,
scale=scale)
self.block_list.append(conv2_2)
conv3_1 = DepthwiseSeparable(
num_channels=int(128 * scale),
num_filters1=128,
num_filters2=128,
num_groups=128,
stride=1,
scale=scale)
self.block_list.append(conv3_1)
conv3_2 = DepthwiseSeparable(
num_channels=int(128 * scale),
num_filters1=128,
num_filters2=256,
num_groups=128,
stride=(2, 1),
scale=scale)
self.block_list.append(conv3_2)
conv4_1 = DepthwiseSeparable(
num_channels=int(256 * scale),
num_filters1=256,
num_filters2=256,
num_groups=256,
stride=1,
scale=scale)
self.block_list.append(conv4_1)
conv4_2 = DepthwiseSeparable(
num_channels=int(256 * scale),
num_filters1=256,
num_filters2=512,
num_groups=256,
stride=(2, 1),
scale=scale)
self.block_list.append(conv4_2)
for _ in range(5):
conv5 = DepthwiseSeparable(
num_channels=int(512 * scale),
num_filters1=512,
num_filters2=512,
num_groups=512,
stride=1,
dw_size=5,
padding=2,
scale=scale,
use_se=False)
self.block_list.append(conv5)
conv5_6 = DepthwiseSeparable(
num_channels=int(512 * scale),
num_filters1=512,
num_filters2=1024,
num_groups=512,
stride=(2, 1),
dw_size=5,
padding=2,
scale=scale,
use_se=True)
self.block_list.append(conv5_6)
conv6 = DepthwiseSeparable(
num_channels=int(1024 * scale),
num_filters1=1024,
num_filters2=1024,
num_groups=1024,
stride=1,
dw_size=5,
padding=2,
use_se=True,
scale=scale)
self.block_list.append(conv6)
self.block_list = nn.Sequential(*self.block_list)
self.pool = nn.MaxPool2D(kernel_size=2, stride=2, padding=0)
self.out_channels = int(1024 * scale)
def forward(self, inputs):
y = self.conv1(inputs)
y = self.block_list(y)
y = self.pool(y)
return y
class SEModule(nn.Layer):
def __init__(self, channel, reduction=4):
super(SEModule, self).__init__()
self.avg_pool = AdaptiveAvgPool2D(1)
self.conv1 = Conv2D(
in_channels=channel,
out_channels=channel // reduction,
kernel_size=1,
stride=1,
padding=0,
weight_attr=ParamAttr(),
bias_attr=ParamAttr())
self.conv2 = Conv2D(
in_channels=channel // reduction,
out_channels=channel,
kernel_size=1,
stride=1,
padding=0,
weight_attr=ParamAttr(),
bias_attr=ParamAttr())
def forward(self, inputs):
outputs = self.avg_pool(inputs)
outputs = self.conv1(outputs)
outputs = F.relu(outputs)
outputs = self.conv2(outputs)
outputs = hardsigmoid(outputs)
return paddle.multiply(x=inputs, y=outputs)
ppocr/postprocess/__init__.py
+3 -2
View File
@@ -21,7 +21,7 @@ import copy
__all__ = ['build_post_process']
from .db_postprocess import DBPostProcess
from .db_postprocess import DBPostProcess, DistillationDBPostProcess
from .east_postprocess import EASTPostProcess
from .sast_postprocess import SASTPostProcess
from .rec_postprocess import CTCLabelDecode, AttnLabelDecode, SRNLabelDecode, DistillationCTCLabelDecode, \
@@ -34,7 +34,8 @@ def build_post_process(config, global_config=None):
support_dict = [
'DBPostProcess', 'EASTPostProcess', 'SASTPostProcess', 'CTCLabelDecode',
'AttnLabelDecode', 'ClsPostProcess', 'SRNLabelDecode', 'PGPostProcess',
'DistillationCTCLabelDecode', 'TableLabelDecode'
'DistillationCTCLabelDecode', 'TableLabelDecode',
'DistillationDBPostProcess'
]
config = copy.deepcopy(config)
ppocr/postprocess/db_postprocess.py
+26
View File
@@ -187,3 +187,29 @@ class DBPostProcess(object):
boxes_batch.append({'points': boxes})
return boxes_batch
class DistillationDBPostProcess(object):
def __init__(self, model_name=["student"],
key=None,
thresh=0.3,
box_thresh=0.6,
max_candidates=1000,
unclip_ratio=1.5,
use_dilation=False,
score_mode="fast",
**kwargs):
self.model_name = model_name
self.key = key
self.post_process = DBPostProcess(thresh=thresh,
box_thresh=box_thresh,
max_candidates=max_candidates,
unclip_ratio=unclip_ratio,
use_dilation=use_dilation,
score_mode=score_mode)
def __call__(self, predicts, shape_list):
results = {}
for k in self.model_name:
results[k] = self.post_process(predicts[k], shape_list=shape_list)
return results
ppocr/utils/save_load.py
+23 -2
View File
@@ -91,14 +91,14 @@ def init_model(config, model, optimizer=None, lr_scheduler=None):
def load_dygraph_params(config, model, logger, optimizer):
ckp = config['Global']['checkpoints']
if ckp and os.path.exists(ckp):
if ckp and os.path.exists(ckp + ".pdparams"):
pre_best_model_dict = init_model(config, model, optimizer)
return pre_best_model_dict
else:
pm = config['Global']['pretrained_model']
if pm is None:
return {}
if not os.path.exists(pm) or not os.path.exists(pm + ".pdparams"):
if not os.path.exists(pm) and not os.path.exists(pm + ".pdparams"):
logger.info(f"The pretrained_model {pm} does not exists!")
return {}
pm = pm if pm.endswith('.pdparams') else pm + '.pdparams'
@@ -116,6 +116,27 @@ def load_dygraph_params(config, model, logger, optimizer):
logger.info(f"loaded pretrained_model successful from {pm}")
return {}
def load_pretrained_params(model, path):
if path is None:
return False
if not os.path.exists(path) and not os.path.exists(path + ".pdparams"):
print(f"The pretrained_model {path} does not exists!")
return False
path = path if path.endswith('.pdparams') else path + '.pdparams'
params = paddle.load(path)
state_dict = model.state_dict()
new_state_dict = {}
for k1, k2 in zip(state_dict.keys(), params.keys()):
if list(state_dict[k1].shape) == list(params[k2].shape):
new_state_dict[k1] = params[k2]
else:
print(
f"The shape of model params {k1} {state_dict[k1].shape} not matched with loaded params {k2} {params[k2].shape} !"
)
model.set_state_dict(new_state_dict)
print(f"load pretrain successful from {path}")
return model
def save_model(model,
optimizer,
test/ocr_det_params.txt
+10 -10
View File
@@ -1,15 +1,14 @@
model_name:ocr_det
python:python3.7
gpu_list:-1|0|0,1
Global.auto_cast:False|True
gpu_list:0|0,1
Global.auto_cast:null
Global.epoch_num:10
Global.save_model_dir:./output/
Global.save_inference_dir:./output/
Train.loader.batch_size_per_card:
Global.use_gpu
Global.pretrained_model
Global.use_gpu:
Global.pretrained_model:null
trainer:norm|pact|fpgm
trainer:norm|pact
norm_train:tools/train.py -c configs/det/det_mv3_db.yml -o Global.pretrained_model=./pretrain_models/MobileNetV3_large_x0_5_pretrained
quant_train:deploy/slim/quantization/quant.py -c configs/det/det_mv3_db.yml -o Global.pretrained_model=./pretrain_models/det_mv3_db_v2.0_train/best_accuracy
fpgm_train:null
@@ -17,6 +16,8 @@ distill_train:null
eval:tools/eval.py -c configs/det/det_mv3_db.yml -o
Global.save_inference_dir:./output/
Global.pretrained_model:
norm_export:tools/export_model.py -c configs/det/det_mv3_db.yml -o
quant_export:deploy/slim/quantization/export_model.py -c configs/det/det_mv3_db.yml -o
fpgm_export:deploy/slim/prune/export_prune_model.py
@@ -29,7 +30,6 @@ inference:tools/infer/predict_det.py
--rec_batch_num:1
--use_tensorrt:True|False
--precision:fp32|fp16|int8
--det_model_dir
--image_dir
--save_log_path
--det_model_dir:./inference/ch_ppocr_mobile_v2.0_det_infer/
--image_dir:./inference/ch_det_data_50/all-sum-510/
--save_log_path:./test/output/
test/prepare.sh
+4 -2
View File
@@ -26,8 +26,10 @@ IFS=$'\n'
# The training params
model_name=$(func_parser_value "${lines[0]}")
train_model_list=$(func_parser_value "${lines[0]}")
trainer_list=$(func_parser_value "${lines[10]}")
# MODE be one of ['lite_train_infer' 'whole_infer' 'whole_train_infer']
MODE=$2
# prepare pretrained weights and dataset
@@ -62,8 +64,8 @@ else
rm -rf ./train_data/icdar2015
wget -nc -P ./train_data https://paddleocr.bj.bcebos.com/dygraph_v2.0/test/ch_det_data_50.tar
if [ ${model_name} = "ocr_det" ]; then
eval_model_name="ch_ppocr_mobile_v2.0_det_train"
wget -nc -P ./inference https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_train.tar
eval_model_name="ch_ppocr_mobile_v2.0_det_infer"
wget -nc -P ./inference https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_infer.tar
cd ./inference && tar xf ${eval_model_name}.tar && cd ../
else
eval_model_name="ch_ppocr_mobile_v2.0_rec_train"
test/test.sh
+78 -62
View File
@@ -41,59 +41,51 @@ gpu_list=$(func_parser_value "${lines[2]}")
autocast_list=$(func_parser_value "${lines[3]}")
autocast_key=$(func_parser_key "${lines[3]}")
epoch_key=$(func_parser_key "${lines[4]}")
epoch_num=$(func_parser_value "${lines[4]}")
save_model_key=$(func_parser_key "${lines[5]}")
save_infer_key=$(func_parser_key "${lines[6]}")
train_batch_key=$(func_parser_key "${lines[7]}")
train_use_gpu_key=$(func_parser_key "${lines[8]}")
pretrain_model_key=$(func_parser_key "${lines[9]}")
train_batch_key=$(func_parser_key "${lines[6]}")
train_use_gpu_key=$(func_parser_key "${lines[7]}")
pretrain_model_key=$(func_parser_key "${lines[8]}")
pretrain_model_value=$(func_parser_value "${lines[8]}")
trainer_list=$(func_parser_value "${lines[10]}")
norm_trainer=$(func_parser_value "${lines[11]}")
pact_trainer=$(func_parser_value "${lines[12]}")
fpgm_trainer=$(func_parser_value "${lines[13]}")
distill_trainer=$(func_parser_value "${lines[14]}")
trainer_list=$(func_parser_value "${lines[9]}")
norm_trainer=$(func_parser_value "${lines[10]}")
pact_trainer=$(func_parser_value "${lines[11]}")
fpgm_trainer=$(func_parser_value "${lines[12]}")
distill_trainer=$(func_parser_value "${lines[13]}")
eval_py=$(func_parser_value "${lines[15]}")
norm_export=$(func_parser_value "${lines[16]}")
pact_export=$(func_parser_value "${lines[17]}")
fpgm_export=$(func_parser_value "${lines[18]}")
distill_export=$(func_parser_value "${lines[19]}")
eval_py=$(func_parser_value "${lines[14]}")
inference_py=$(func_parser_value "${lines[20]}")
use_gpu_key=$(func_parser_key "${lines[21]}")
use_gpu_list=$(func_parser_value "${lines[21]}")
use_mkldnn_key=$(func_parser_key "${lines[22]}")
use_mkldnn_list=$(func_parser_value "${lines[22]}")
cpu_threads_key=$(func_parser_key "${lines[23]}")
cpu_threads_list=$(func_parser_value "${lines[23]}")
batch_size_key=$(func_parser_key "${lines[24]}")
batch_size_list=$(func_parser_value "${lines[24]}")
use_trt_key=$(func_parser_key "${lines[25]}")
use_trt_list=$(func_parser_value "${lines[25]}")
precision_key=$(func_parser_key "${lines[26]}")
precision_list=$(func_parser_value "${lines[26]}")
model_dir_key=$(func_parser_key "${lines[27]}")
image_dir_key=$(func_parser_key "${lines[28]}")
save_log_key=$(func_parser_key "${lines[29]}")
save_infer_key=$(func_parser_key "${lines[15]}")
export_weight=$(func_parser_key "${lines[16]}")
norm_export=$(func_parser_value "${lines[17]}")
pact_export=$(func_parser_value "${lines[18]}")
fpgm_export=$(func_parser_value "${lines[19]}")
distill_export=$(func_parser_value "${lines[20]}")
inference_py=$(func_parser_value "${lines[21]}")
use_gpu_key=$(func_parser_key "${lines[22]}")
use_gpu_list=$(func_parser_value "${lines[22]}")
use_mkldnn_key=$(func_parser_key "${lines[23]}")
use_mkldnn_list=$(func_parser_value "${lines[23]}")
cpu_threads_key=$(func_parser_key "${lines[24]}")
cpu_threads_list=$(func_parser_value "${lines[24]}")
batch_size_key=$(func_parser_key "${lines[25]}")
batch_size_list=$(func_parser_value "${lines[25]}")
use_trt_key=$(func_parser_key "${lines[26]}")
use_trt_list=$(func_parser_value "${lines[26]}")
precision_key=$(func_parser_key "${lines[27]}")
precision_list=$(func_parser_value "${lines[27]}")
infer_model_key=$(func_parser_key "${lines[28]}")
infer_model=$(func_parser_value "${lines[28]}")
image_dir_key=$(func_parser_key "${lines[29]}")
infer_img_dir=$(func_parser_value "${lines[29]}")
save_log_key=$(func_parser_key "${lines[30]}")
LOG_PATH="./test/output"
mkdir -p ${LOG_PATH}
status_log="${LOG_PATH}/results.log"
if [ ${MODE} = "lite_train_infer" ]; then
export infer_img_dir="./train_data/icdar2015/text_localization/ch4_test_images/"
export epoch_num=10
elif [ ${MODE} = "whole_infer" ]; then
export infer_img_dir="./train_data/icdar2015/text_localization/ch4_test_images/"
export epoch_num=10
elif [ ${MODE} = "whole_train_infer" ]; then
export infer_img_dir="./train_data/icdar2015/text_localization/ch4_test_images/"
export epoch_num=300
else
export infer_img_dir="./inference/ch_det_data_50/all-sum-510"
export infer_model_dir="./inference/ch_ppocr_mobile_v2.0_det_train/best_accuracy"
fi
function func_inference(){
IFS='|'
@@ -109,8 +101,8 @@ function func_inference(){
for use_mkldnn in ${use_mkldnn_list[*]}; do
for threads in ${cpu_threads_list[*]}; do
for batch_size in ${batch_size_list[*]}; do
_save_log_path="${_log_path}/infer_cpu_usemkldnn_${use_mkldnn}_threads_${threads}_batchsize_${batch_size}"
command="${_python} ${_script} ${use_gpu_key}=${use_gpu} ${use_mkldnn_key}=${use_mkldnn} ${cpu_threads_key}=${threads} ${model_dir_key}=${_model_dir} ${batch_size_key}=${batch_size} ${image_dir_key}=${_img_dir} ${save_log_key}=${_save_log_path}"
_save_log_path="${_log_path}/infer_cpu_usemkldnn_${use_mkldnn}_threads_${threads}_batchsize_${batch_size}.log"
command="${_python} ${_script} ${use_gpu_key}=${use_gpu} ${use_mkldnn_key}=${use_mkldnn} ${cpu_threads_key}=${threads} ${infer_model_key}=${_model_dir} ${batch_size_key}=${batch_size} ${image_dir_key}=${_img_dir} ${save_log_key}=${_save_log_path} --benchmark=True"
eval $command
status_check $? "${command}" "${status_log}"
done
@@ -123,8 +115,8 @@ function func_inference(){
continue
fi
for batch_size in ${batch_size_list[*]}; do
_save_log_path="${_log_path}/infer_gpu_usetrt_${use_trt}_precision_${precision}_batchsize_${batch_size}"
command="${_python} ${_script} ${use_gpu_key}=${use_gpu} ${use_trt_key}=${use_trt} ${precision_key}=${precision} ${model_dir_key}=${_model_dir} ${batch_size_key}=${batch_size} ${image_dir_key}=${_img_dir} ${save_log_key}=${_save_log_path}"
_save_log_path="${_log_path}/infer_gpu_usetrt_${use_trt}_precision_${precision}_batchsize_${batch_size}.log"
command="${_python} ${_script} ${use_gpu_key}=${use_gpu} ${use_trt_key}=${use_trt} ${precision_key}=${precision} ${infer_model_key}=${_model_dir} ${batch_size_key}=${batch_size} ${image_dir_key}=${_img_dir} ${save_log_key}=${_save_log_path} --benchmark=True"
eval $command
status_check $? "${command}" "${status_log}"
done
@@ -144,6 +136,7 @@ for gpu in ${gpu_list[*]}; do
env=""
elif [ ${#gpu} -le 1 ];then
env="export CUDA_VISIBLE_DEVICES=${gpu}"
eval ${env}
elif [ ${#gpu} -le 15 ];then
IFS=","
array=(${gpu})
@@ -155,6 +148,7 @@ for gpu in ${gpu_list[*]}; do
ips=${array[0]}
gpu=${array[1]}
IFS="|"
env=" "
fi
for autocast in ${autocast_list[*]}; do
for trainer in ${trainer_list[*]}; do
@@ -179,13 +173,32 @@ for gpu in ${gpu_list[*]}; do
continue
fi
save_log="${LOG_PATH}/${trainer}_gpus_${gpu}_autocast_${autocast}"
if [ ${#gpu} -le 2 ];then # epoch_num #TODO
cmd="${python} ${run_train} ${train_use_gpu_key}=${use_gpu} ${autocast_key}=${autocast} ${epoch_key}=${epoch_num} ${save_model_key}=${save_log} "
elif [ ${#gpu} -le 15 ];then
cmd="${python} -m paddle.distributed.launch --gpus=${gpu} ${run_train} ${autocast_key}=${autocast} ${epoch_key}=${epoch_num} ${save_model_key}=${save_log}"
# not set autocast when autocast is null
if [ ${autocast} = "null" ]; then
set_autocast=" "
else
cmd="${python} -m paddle.distributed.launch --ips=${ips} --gpus=${gpu} ${run_train} ${autocast_key}=${autocast} ${epoch_key}=${epoch_num} ${save_model_key}=${save_log}"
set_autocast="${autocast_key}=${autocast}"
fi
# not set epoch when whole_train_infer
if [ ${MODE} != "whole_train_infer" ]; then
set_epoch="${epoch_key}=${epoch_num}"
else
set_epoch=" "
fi
# set pretrain
if [ ${pretrain_model_value} != "null" ]; then
set_pretrain="${pretrain_model_key}=${pretrain_model_value}"
else
set_pretrain=" "
fi
save_log="${LOG_PATH}/${trainer}_gpus_${gpu}_autocast_${autocast}"
if [ ${#gpu} -le 2 ];then # train with cpu or single gpu
cmd="${python} ${run_train} ${train_use_gpu_key}=${use_gpu} ${save_model_key}=${save_log} ${set_epoch} ${set_pretrain} ${set_autocast}"
elif [ ${#gpu} -le 15 ];then # train with multi-gpu
cmd="${python} -m paddle.distributed.launch --gpus=${gpu} ${run_train} ${save_model_key}=${save_log} ${set_epoch} ${set_pretrain} ${set_autocast}"
else # train with multi-machine
cmd="${python} -m paddle.distributed.launch --ips=${ips} --gpus=${gpu} ${run_train} ${save_model_key}=${save_log} ${set_pretrain} ${set_epoch} ${set_autocast}"
fi
# run train
eval $cmd
@@ -198,24 +211,27 @@ for gpu in ${gpu_list[*]}; do
# run export model
save_infer_path="${save_log}"
export_cmd="${python} ${run_export} ${save_model_key}=${save_log} ${pretrain_model_key}=${save_log}/latest ${save_infer_key}=${save_infer_path}"
export_cmd="${python} ${run_export} ${save_model_key}=${save_log} ${export_weight}=${save_log}/latest ${save_infer_key}=${save_infer_path}"
eval $export_cmd
status_check $? "${export_cmd}" "${status_log}"
#run inference
eval $env
save_infer_path="${save_log}"
func_inference "${python}" "${inference_py}" "${save_infer_path}" "${LOG_PATH}" "${infer_img_dir}"
eval "unset CUDA_VISIBLE_DEVICES"
done
done
done
else
save_infer_path="${LOG_PATH}/${MODE}"
run_export=${norm_export}
export_cmd="${python} ${run_export} ${save_model_key}=${save_infer_path} ${pretrain_model_key}=${infer_model_dir} ${save_infer_key}=${save_infer_path}"
eval $export_cmd
status_check $? "${export_cmd}" "${status_log}"
GPUID=$3
if [ ${#GPUID} -le 0 ];then
env=" "
else
env="export CUDA_VISIBLE_DEVICES=${GPUID}"
fi
echo $env
#run inference
func_inference "${python}" "${inference_py}" "${save_infer_path}" "${LOG_PATH}" "${infer_img_dir}"
func_inference "${python}" "${inference_py}" "${infer_model}" "${LOG_PATH}" "${infer_img_dir}"
fi
test1/table/README_ch.md
+23 -1
View File
@@ -19,7 +19,29 @@
### 2.1 训练
TBD
#### 数据准备
训练数据使用公开数据集[PubTabNet](https://arxiv.org/abs/1911.10683)，可以从[官网](https://github.com/ibm-aur-nlp/PubTabNet)下载。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
*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
python3 tools/train.py -c configs/table/table_mv3.yml
# 单机多卡训练，通过 --gpus 参数设置使用的GPU ID
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/table/table_mv3.yml
```
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](./config.md)。
#### 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
python3 tools/train.py -c configs/table/table_mv3.yml -o Global.checkpoints=./your/trained/model
```
**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。
### 2.2 评估
先cd到PaddleOCR/ppstructure目录下
tools/eval.py
+6 -3
View File
@@ -27,7 +27,7 @@ from ppocr.data import build_dataloader
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 init_model
from ppocr.utils.save_load import init_model, load_pretrained_params
from ppocr.utils.utility import print_dict
import tools.program as program
@@ -55,7 +55,10 @@ def main():
model = build_model(config['Architecture'])
use_srn = config['Architecture']['algorithm'] == "SRN"
model_type = config['Architecture']['model_type']
if "model_type" in config['Architecture'].keys():
model_type = config['Architecture']['model_type']
else:
model_type = None
best_model_dict = init_model(config, model)
if len(best_model_dict):
@@ -68,7 +71,7 @@ def main():
# start eval
metric = program.eval(model, valid_dataloader, post_process_class,
eval_class, model_type, use_srn)
eval_class, model_type, use_srn)
logger.info('metric eval ***************')
for k, v in metric.items():
logger.info('{}:{}'.format(k, v))
tools/infer/predict_det.py
+4 -5
View File
@@ -106,7 +106,7 @@ class TextDetector(object):
model_precision=args.precision,
batch_size=1,
data_shape="dynamic",
save_path="./output/auto_log.lpg",
save_path=args.save_log_path,
inference_config=self.config,
pids=pid,
process_name=None,
@@ -174,7 +174,7 @@ class TextDetector(object):
data = {'image': img}
st = time.time()
if self.args.benchmark:
self.autolog.times.start()
@@ -212,7 +212,7 @@ class TextDetector(object):
else:
raise NotImplementedError
self.predictor.try_shrink_memory()
#self.predictor.try_shrink_memory()
post_result = self.postprocess_op(preds, shape_list)
dt_boxes = post_result[0]['points']
if self.det_algorithm == "SAST" and self.det_sast_polygon:
@@ -262,7 +262,6 @@ if __name__ == "__main__":
"det_res_{}".format(img_name_pure))
cv2.imwrite(img_path, src_im)
logger.info("The visualized image saved in {}".format(img_path))
if args.benchmark:
text_detector.autolog.report()
tools/infer/predict_system.py
-2
View File
@@ -174,8 +174,6 @@ def main(args):
logger.info("The predict total time is {}".format(time.time() - _st))
logger.info("\nThe predict total time is {}".format(total_time))
img_num = text_sys.text_detector.det_times.img_num
if __name__ == "__main__":
args = utility.parse_args()
tools/infer/utility.py
+6 -3
View File
@@ -37,7 +37,7 @@ def init_args():
parser.add_argument("--use_gpu", type=str2bool, default=True)
parser.add_argument("--ir_optim", type=str2bool, default=True)
parser.add_argument("--use_tensorrt", type=str2bool, default=False)
parser.add_argument("--min_subgraph_size", type=int, default=3)
parser.add_argument("--min_subgraph_size", type=int, default=10)
parser.add_argument("--precision", type=str, default="fp32")
parser.add_argument("--gpu_mem", type=int, default=500)
@@ -164,7 +164,7 @@ def create_predictor(args, mode, logger):
config.enable_use_gpu(args.gpu_mem, 0)
if args.use_tensorrt:
config.enable_tensorrt_engine(
precision_mode=inference.PrecisionType.Float32,
precision_mode=precision,
max_batch_size=args.max_batch_size,
min_subgraph_size=args.min_subgraph_size)
# skip the minmum trt subgraph
@@ -176,6 +176,7 @@ def create_predictor(args, mode, logger):
"conv2d_59.tmp_0": [1, 96, 20, 20],
"nearest_interp_v2_1.tmp_0": [1, 96, 10, 10],
"nearest_interp_v2_2.tmp_0": [1, 96, 20, 20],
"conv2d_124.tmp_0": [1, 96, 20, 20],
"nearest_interp_v2_3.tmp_0": [1, 24, 20, 20],
"nearest_interp_v2_4.tmp_0": [1, 24, 20, 20],
"nearest_interp_v2_5.tmp_0": [1, 24, 20, 20],
@@ -188,6 +189,7 @@ def create_predictor(args, mode, logger):
"conv2d_91.tmp_0": [1, 96, 200, 200],
"conv2d_59.tmp_0": [1, 96, 400, 400],
"nearest_interp_v2_1.tmp_0": [1, 96, 200, 200],
"conv2d_124.tmp_0": [1, 256, 400, 400],
"nearest_interp_v2_2.tmp_0": [1, 96, 400, 400],
"nearest_interp_v2_3.tmp_0": [1, 24, 400, 400],
"nearest_interp_v2_4.tmp_0": [1, 24, 400, 400],
@@ -202,6 +204,7 @@ def create_predictor(args, mode, logger):
"conv2d_59.tmp_0": [1, 96, 160, 160],
"nearest_interp_v2_1.tmp_0": [1, 96, 80, 80],
"nearest_interp_v2_2.tmp_0": [1, 96, 160, 160],
"conv2d_124.tmp_0": [1, 256, 160, 160],
"nearest_interp_v2_3.tmp_0": [1, 24, 160, 160],
"nearest_interp_v2_4.tmp_0": [1, 24, 160, 160],
"nearest_interp_v2_5.tmp_0": [1, 24, 160, 160],
@@ -237,7 +240,7 @@ def create_predictor(args, mode, logger):
# enable memory optim
config.enable_memory_optim()
config.disable_glog_info()
#config.disable_glog_info()
config.delete_pass("conv_transpose_eltwiseadd_bn_fuse_pass")
if mode == 'table':
tools/program.py
+4 -1
View File
@@ -186,7 +186,10 @@ def train(config,
model.train()
use_srn = config['Architecture']['algorithm'] == "SRN"
model_type = config['Architecture']['model_type']
try:
model_type = config['Architecture']['model_type']
except:
model_type = None
if 'start_epoch' in best_model_dict:
start_epoch = best_model_dict['start_epoch']
tools/train.py
-1
View File
@@ -98,7 +98,6 @@ def main(config, device, logger, vdl_writer):
eval_class = build_metric(config['Metric'])
# load pretrain model
pre_best_model_dict = load_dygraph_params(config, model, logger, optimizer)
logger.info('train dataloader has {} iters'.format(len(train_dataloader)))
if valid_dataloader is not None:
logger.info('valid dataloader has {} iters'.format(