未验证 提交 32d533b3 编写于 作者: D Double_V 提交者: GitHub

Merge pull request #5079 from LDOUBLEV/kie_doc

add distill doc
...@@ -21,6 +21,7 @@ Architecture: ...@@ -21,6 +21,7 @@ Architecture:
model_type: det model_type: det
Models: Models:
Teacher: Teacher:
pretrained: ./pretrain_models/ch_ppocr_server_v2.0_det_train/best_accuracy
freeze_params: true freeze_params: true
return_all_feats: false return_all_feats: false
model_type: det model_type: det
...@@ -36,6 +37,7 @@ Architecture: ...@@ -36,6 +37,7 @@ Architecture:
name: DBHead name: DBHead
k: 50 k: 50
Student: Student:
pretrained:
freeze_params: false freeze_params: false
return_all_feats: false return_all_feats: false
model_type: det model_type: det
...@@ -52,6 +54,7 @@ Architecture: ...@@ -52,6 +54,7 @@ Architecture:
name: DBHead name: DBHead
k: 50 k: 50
Student2: Student2:
pretrained:
freeze_params: false freeze_params: false
return_all_feats: false return_all_feats: false
model_type: det model_type: det
......
...@@ -18,6 +18,7 @@ Global: ...@@ -18,6 +18,7 @@ Global:
Architecture: Architecture:
name: DistillationModel name: DistillationModel
algorithm: Distillation algorithm: Distillation
model_type: det
Models: Models:
Student: Student:
pretrained: ./pretrain_models/MobileNetV3_large_x0_5_pretrained pretrained: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
......
...@@ -18,6 +18,7 @@ Global: ...@@ -18,6 +18,7 @@ Global:
Architecture: Architecture:
name: DistillationModel name: DistillationModel
algorithm: Distillation algorithm: Distillation
model_type: det
Models: Models:
Student: Student:
pretrained: ./pretrain_models/MobileNetV3_large_x0_5_pretrained pretrained: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
......
...@@ -281,4 +281,274 @@ paddle.save(s_params, "ch_PP-OCRv2_rec_train/student.pdparams") ...@@ -281,4 +281,274 @@ paddle.save(s_params, "ch_PP-OCRv2_rec_train/student.pdparams")
### 2.2 检测配置文件解析 ### 2.2 检测配置文件解析
* coming soon!
检测模型蒸馏的配置文件在PaddleOCR/configs/det/ch_PP-OCRv2/目录下,包含三个蒸馏配置文件:
- ch_PP-OCRv2_det_cml.yml,采用cml蒸馏,采用一个大模型蒸馏两个小模型,且两个小模型互相学习的方法
- ch_PP-OCRv2_det_dml.yml,采用DML的蒸馏,两个Student模型互蒸馏的方法
- ch_PP-OCRv2_det_distill.yml,采用Teacher大模型蒸馏小模型Student的方法
#### 2.2.1 模型结构
知识蒸馏任务中,模型结构配置如下所示:
```
Architecture:
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
Student: # 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: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
freeze_params: false # 是否需要固定参数
return_all_feats: false # 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: 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: # 另外一个子网络,这里给的是普通大模型蒸小模型的蒸馏示例,
pretrained: ./pretrain_models/ch_ppocr_server_v2.0_det_train/best_accuracy
freeze_params: true # Teacher模型是训练好的,不需要参与训练,freeze_params设置为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
```
如果是采用DML,即两个小模型互相学习的方法,上述配置文件里的Teacher网络结构需要设置为Student模型一样的配置,具体参考配置文件[ch_PP-OCRv2_det_dml.yml](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.4/configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_dml.yml)

下面介绍[ch_PP-OCRv2_det_cml.yml](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.4/configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml)的配置文件参数:
```
Architecture:
name: DistillationModel
algorithm: Distillation
model_type: det
Models:
Teacher: # CML蒸馏的Teacher模型配置
pretrained: ./pretrain_models/ch_ppocr_server_v2.0_det_train/best_accuracy
freeze_params: true # Teacher 不训练
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
Student: # CML蒸馏的Student模型配置
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: # CML蒸馏的Student2模型配置
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
```
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``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)。
在蒸馏任务中,为了方便添加蒸馏损失函数,每个网络的输出保存为`dict`,其中包含子模块输出。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,检测ch_PP-OCRv2_det_distill.yml蒸馏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
- DistillationDilaDBLoss: # 基于蒸馏的DB损失函数,继承自标准的DBloss
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_pairs: # 对于蒸馏模型的预测结果,提取这两个子网络的输出,计算Teacher模型和Student模型输出的loss
- ["Student", "Teacher"]
key: maps # 取子网络输出dict中,该key对应的tensor
balance_loss: true # 以下几个参数为标准DBloss的配置参数
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
- DistillationDBLoss: # 基于蒸馏的DB损失函数,继承自标准的DBloss,用于计算Student和GT之间的loss
weight: 1.0
model_name_list: ["Student"] # 模型名字只有Student,表示计算Student和GT之间的loss
name: DBLoss
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
```
同理,检测ch_PP-OCRv2_det_cml.yml蒸馏损失函数配置如下所示。相比较于ch_PP-OCRv2_det_distill.yml的损失函数配置,cml蒸馏的损失函数配置做了3个改动:
```yaml
Loss:
name: CombinedLoss
loss_config_list:
- DistillationDilaDBLoss:
weight: 1.0
model_name_pairs:
- ["Student", "Teacher"]
- ["Student2", "Teacher"] # 改动1,计算两个Student和Teacher的损失
key: maps
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
- DistillationDMLLoss: # 改动2,增加计算两个Student之间的损失
model_name_pairs:
- ["Student", "Student2"]
maps_name: "thrink_maps"
weight: 1.0
# act: None
key: maps
- DistillationDBLoss:
weight: 1.0
model_name_list: ["Student", "Student2"] # 改动3,计算两个Student和GT之间的损失
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
```
关于`DistillationDilaDBLoss`更加具体的实现可以参考: [distillation_loss.py](https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.4/ppocr/losses/distillation_loss.py#L185)。关于`DistillationDBLoss`等蒸馏损失函数更加具体的实现可以参考[distillation_loss.py](https://github.com/PaddlePaddle/PaddleOCR/blob/04c44974b13163450dfb6bd2c327863f8a194b3c/ppocr/losses/distillation_loss.py?_pjax=%23js-repo-pjax-container%2C%20div%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L148)
#### 2.1.3 后处理
知识蒸馏任务中,检测蒸馏后处理配置如下所示。
```yaml
PostProcess:
name: DistillationDBPostProcess # DB检测蒸馏任务的CTC解码后处理,继承自标准的DBPostProcess类
model_name: ["Student", "Student2", "Teacher"] # 对于蒸馏模型的预测结果,提取多个子网络的输出,进行解码,不需要后处理的网络可以不在model_name中设置
thresh: 0.3
box_thresh: 0.6
max_candidates: 1000
unclip_ratio: 1.5
```
以上述配置为例,最终会同时计算`Student``Student2``Teacher` 3个子网络的输出做后处理计算。同时,由于有多个输入,后处理返回的输出也有多个,
关于`DistillationDBPostProcess`更加具体的实现可以参考: [db_postprocess.py](../../ppocr/postprocess/db_postprocess.py#L195)
#### 2.1.4 蒸馏指标计算
知识蒸馏任务中,检测蒸馏指标计算配置如下所示。
```yaml
Metric:
name: DistillationMetric
base_metric_name: DetMetric
main_indicator: hmean
key: "Student"
```
由于蒸馏需要包含多个网络,甚至多个Student网络,在计算指标的时候只需要计算一个Student网络的指标即可,`key`字段设置为`Student`则表示只计算`Student`网络的精度。
#### 2.1.5 检测蒸馏模型finetune
检测蒸馏有三种方式:
- 采用ch_PP-OCRv2_det_distill.yml,Teacher模型设置为PaddleOCR提供的模型或者您训练好的大模型
- 采用ch_PP-OCRv2_det_cml.yml,采用cml蒸馏,同样Teacher模型设置为PaddleOCR提供的模型或者您训练好的大模型
- 采用ch_PP-OCRv2_det_dml.yml,采用DML的蒸馏,两个Student模型互蒸馏的方法,在PaddleOCR采用的数据集上大约有1.7%的精度提升。
在具体finetune时,需要在网络结构的`pretrained`参数中设置要加载的预训练模型。
在精度提升方面,cml的精度>dml的精度>distill蒸馏方法的精度。当数据量不足或者Teacher模型精度与Student精度相差不大的时候,这个结论或许会改变。
另外,由于PaddleOCR提供的蒸馏预训练模型包含了多个模型的参数,如果您希望提取Student模型的参数,可以参考如下代码:
```
# 下载蒸馏训练模型的参数
wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_distill_train.tar
```
```python
import paddle
# 加载预训练模型
all_params = paddle.load("ch_PP-OCRv2_det_distill_train/best_accuracy.pdparams")
# 查看权重参数的keys
print(all_params.keys())
# 学生模型的权重提取
s_params = {key[len("Student."):]: all_params[key] for key in all_params if "Student." in key}
# 查看学生模型权重参数的keys
print(s_params.keys())
# 保存
paddle.save(s_params, "ch_PP-OCRv2_det_distill_train/student.pdparams")
```
最终`Student`模型的参数将会保存在`ch_PP-OCRv2_det_distill_train/student.pdparams`中,用于模型的fine-tune。
...@@ -259,7 +259,6 @@ else ...@@ -259,7 +259,6 @@ else
env="" env=""
elif [ ${#gpu} -le 1 ];then elif [ ${#gpu} -le 1 ];then
env="export CUDA_VISIBLE_DEVICES=${gpu}" env="export CUDA_VISIBLE_DEVICES=${gpu}"
eval ${env}
elif [ ${#gpu} -le 15 ];then elif [ ${#gpu} -le 15 ];then
IFS="," IFS=","
array=(${gpu}) array=(${gpu})
...@@ -280,6 +279,7 @@ else ...@@ -280,6 +279,7 @@ else
set_amp_config=" " set_amp_config=" "
fi fi
for trainer in ${trainer_list[*]}; do for trainer in ${trainer_list[*]}; do
eval ${env}
flag_quant=False flag_quant=False
if [ ${trainer} = ${pact_key} ]; then if [ ${trainer} = ${pact_key} ]; then
run_train=${pact_trainer} run_train=${pact_trainer}
...@@ -332,7 +332,6 @@ else ...@@ -332,7 +332,6 @@ else
cmd="${python} -m paddle.distributed.launch --ips=${ips} --gpus=${gpu} ${run_train} ${set_use_gpu} ${set_save_model} ${set_pretrain} ${set_epoch} ${set_autocast} ${set_batchsize} ${set_train_params1} ${set_amp_config}" cmd="${python} -m paddle.distributed.launch --ips=${ips} --gpus=${gpu} ${run_train} ${set_use_gpu} ${set_save_model} ${set_pretrain} ${set_epoch} ${set_autocast} ${set_batchsize} ${set_train_params1} ${set_amp_config}"
fi fi
# run train # run train
eval "unset CUDA_VISIBLE_DEVICES"
eval $cmd eval $cmd
status_check $? "${cmd}" "${status_log}" status_check $? "${cmd}" "${status_log}"
......
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