forgetest5
/
PaddleOCR
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #810 from yukavio/develop 

update bash of slim pruning
This commit is contained in:
Double_V 2020-09-23 20:20:21 +08:00 committed by GitHub
parent 115b517584 2d07a0bcb9
commit c320457d73
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 8 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
deploy/slim/prune/README.md
View File

@ -51,14 +51,14 @@ python setup.py install
进入PaddleOCR根目录，通过以下命令对模型进行敏感度分析训练： 进入PaddleOCR根目录，通过以下命令对模型进行敏感度分析训练：
```bash ```bash
python deploy/slim/prune/sensitivity_anal.py -c configs/det/det_mv3_db.yml -o Global.pretrain_weights="your trained model" Global.test_batch_size_per_card=1 python deploy/slim/prune/sensitivity_anal.py -c configs/det/det_mv3_db_v1.1.yml -o Global.pretrain_weights="your trained model" Global.test_batch_size_per_card=1
``` ```
### 4. 模型裁剪训练 ### 4. 模型裁剪训练
裁剪时通过之前的敏感度分析文件决定每个网络层的裁剪比例。在具体实现时，为了尽可能多的保留从图像中提取的低阶特征，我们跳过了backbone中靠近输入的4个卷积层。同样，为了减少由于裁剪导致的模型性能损失，我们通过之前敏感度分析所获得的敏感度表，人工挑选出了一些冗余较少，对裁剪较为敏感的[网络层](https://github.com/PaddlePaddle/PaddleOCR/blob/develop/deploy/slim/prune/pruning_and_finetune.py#L41)（指在较低的裁剪比例下就导致很高性能损失的网络层），并在之后的裁剪过程中选择避开这些网络层。裁剪过后finetune的过程沿用OCR检测模型原始的训练策略。 裁剪时通过之前的敏感度分析文件决定每个网络层的裁剪比例。在具体实现时，为了尽可能多的保留从图像中提取的低阶特征，我们跳过了backbone中靠近输入的4个卷积层。同样，为了减少由于裁剪导致的模型性能损失，我们通过之前敏感度分析所获得的敏感度表，人工挑选出了一些冗余较少，对裁剪较为敏感的[网络层](https://github.com/PaddlePaddle/PaddleOCR/blob/develop/deploy/slim/prune/pruning_and_finetune.py#L41)（指在较低的裁剪比例下就导致很高性能损失的网络层），并在之后的裁剪过程中选择避开这些网络层。裁剪过后finetune的过程沿用OCR检测模型原始的训练策略。
```bash ```bash
python deploy/slim/prune/pruning_and_finetune.py -c configs/det/det_mv3_db.yml -o Global.pretrain_weights=./deploy/slim/prune/pretrain_models/det_mv3_db/best_accuracy Global.test_batch_size_per_card=1 python deploy/slim/prune/pruning_and_finetune.py -c configs/det/det_mv3_db_v1.1.yml -o Global.pretrain_weights=./deploy/slim/prune/pretrain_models/det_mv3_db/best_accuracy Global.test_batch_size_per_card=1
``` ```
通过对比可以发现，经过裁剪训练保存的模型更小。 通过对比可以发现，经过裁剪训练保存的模型更小。
@ -66,7 +66,7 @@ python deploy/slim/prune/pruning_and_finetune.py -c configs/det/det_mv3_db.yml -
在得到裁剪训练保存的模型后，我们可以将其导出为inference_model： 在得到裁剪训练保存的模型后，我们可以将其导出为inference_model：
```bash ```bash
python deploy/slim/prune/export_prune_model.py -c configs/det/det_mv3_db.yml -o Global.pretrain_weights=./output/det_db/best_accuracy Global.test_batch_size_per_card=1 Global.save_inference_dir=inference_model python deploy/slim/prune/export_prune_model.py -c configs/det/det_mv3_db_v1.1.yml -o Global.pretrain_weights=./output/det_db/best_accuracy Global.test_batch_size_per_card=1 Global.save_inference_dir=inference_model
``` ```
inference model的预测和部署参考： inference model的预测和部署参考：

6
deploy/slim/prune/README_en.md
View File

@ -55,7 +55,7 @@ Enter the PaddleOCR root directory，perform sensitivity analysis on the model w
```bash ```bash
python deploy/slim/prune/sensitivity_anal.py -c configs/det/det_mv3_db.yml -o Global.pretrain_weights=./deploy/slim/prune/pretrain_models/det_mv3_db/best_accuracy Global.test_batch_size_per_card=1 python deploy/slim/prune/sensitivity_anal.py -c configs/det/det_mv3_db_v1.1.yml -o Global.pretrain_weights=./deploy/slim/prune/pretrain_models/det_mv3_db/best_accuracy Global.test_batch_size_per_card=1
``` ```
@ -67,7 +67,7 @@ python deploy/slim/prune/sensitivity_anal.py -c configs/det/det_mv3_db.yml -o Gl
```bash ```bash
python deploy/slim/prune/pruning_and_finetune.py -c configs/det/det_mv3_db.yml -o Global.pretrain_weights=./deploy/slim/prune/pretrain_models/det_mv3_db/best_accuracy Global.test_batch_size_per_card=1 python deploy/slim/prune/pruning_and_finetune.py -c configs/det/det_mv3_db_v1.1.yml -o Global.pretrain_weights=./deploy/slim/prune/pretrain_models/det_mv3_db/best_accuracy Global.test_batch_size_per_card=1
``` ```
@ -76,7 +76,7 @@ python deploy/slim/prune/pruning_and_finetune.py -c configs/det/det_mv3_db.yml -
We can export the pruned model as inference_model for deployment: We can export the pruned model as inference_model for deployment:
```bash ```bash
python deploy/slim/prune/export_prune_model.py -c configs/det/det_mv3_db.yml -o Global.pretrain_weights=./output/det_db/best_accuracy Global.test_batch_size_per_card=1 Global.save_inference_dir=inference_model python deploy/slim/prune/export_prune_model.py -c configs/det/det_mv3_db_v1.1.yml -o Global.pretrain_weights=./output/det_db/best_accuracy Global.test_batch_size_per_card=1 Global.save_inference_dir=inference_model
``` ```
Reference for prediction and deployment of inference model: Reference for prediction and deployment of inference model:

3
deploy/slim/prune/pruning_and_finetune.py
View File

@ -92,7 +92,8 @@ def main():
sen = load_sensitivities("sensitivities_0.data") sen = load_sensitivities("sensitivities_0.data")
for i in skip_list: for i in skip_list:
sen.pop(i) if i in sen.keys():
sen.pop(i)
back_bone_list = ['conv' + str(x) for x in range(1, 5)] back_bone_list = ['conv' + str(x) for x in range(1, 5)]
for i in back_bone_list: for i in back_bone_list:
for key in list(sen.keys()): for key in list(sen.keys()):