Skip to content

Latest commit

 

History

History
99 lines (81 loc) · 3.29 KB

README.md

File metadata and controls

99 lines (81 loc) · 3.29 KB
Raw

PaddleMVF模型精度对齐说明文档

所有的对齐代码以及结果均保存在alignment文件夹中。

依赖环境

PaddlePaddle 2.2.0

torch 1.9.0

repord_log 1.0

安装方法参考:https://github.com/WenmuZhou/reprod_log

1 模型结构对齐

第一步是最基本的模型结构对齐，首先创建数据。

python alignment/step1/create_fakedata.py

然后分别执行paddle版本和torch版本的前向运算。

python alignment/step1/paddle_forward.py 
python alignment/step1/torch_forward.py

最后检查两个版本的网络输出结果差异，检测结果通过。

python alignment/step1/check_step1.py
[2021/12/20 15:19:14] root INFO: logits: 
[2021/12/20 15:19:14] root INFO:        mean diff: check passed: True, value: 1.973553542811146e-09
[2021/12/20 15:19:14] root INFO: diff check passed

2 评估指标对齐

分别运行paddle版本和torch版本的评估程序。

python alignment/step2/paddle_val.py --dataset_root test_tipc/data/mini_ucf
python alignment/step2/toch_val.py --dataset_root test_tipc/data/mini_ucf

最后检测两个评估程序的输出差异。

$ python alignment/step2/check_step2.py 
[2021/12/20 15:52:45] root INFO: top1_acc: 
[2021/12/20 15:52:45] root INFO:        mean diff: check passed: True, value: 0.0
[2021/12/20 15:52:45] root INFO: mean_class_accuracy: 
[2021/12/20 15:52:45] root INFO:        mean diff: check passed: True, value: 0.0
[2021/12/20 15:52:45] root INFO: diff check passed

结果一致，测试通过。

这里使用了TIPC的小规模数据集，如果找不到数据集需要先解压TIPC测试链条的数据集。

bash test_tipc/prepare.sh test_tipc/configs/mvf/train_infer_python.txt 'lite_train_lite_infer'

3损失函数对齐

分别运行paddle版本和torch版本的程序。

python alignment/step3/paddle_loss.py
python alignment/step3/torch_loss.py

运行检测程序， 检测损失函数是否对齐。

python alignment/step3/check_step3.py
[2021/12/21 09:40:33] root INFO: loss: 
[2021/12/21 09:40:33] root INFO:        mean diff: check passed: True, value: 0.0
[2021/12/21 09:40:33] root INFO: diff check passed

结果检测通过。

4反向初次对齐

分别运行paddle版本和torch版本的程序，反向运算5次。

python alignment/step4/paddle_train.py
python alignment/step4/torch_train.py

运行检测程序， 检测反向运算是否对齐。

python alignment/step4/check_step4.py
[2021/12/21 09:49:46] root INFO: loss_0: 
[2021/12/21 09:49:46] root INFO:        mean diff: check passed: True, value: 0.0
[2021/12/21 09:49:46] root INFO: loss_1: 
[2021/12/21 09:49:46] root INFO:        mean diff: check passed: True, value: 0.0
[2021/12/21 09:49:46] root INFO: loss_2: 
[2021/12/21 09:49:46] root INFO:        mean diff: check passed: True, value: 0.0
[2021/12/21 09:49:46] root INFO: loss_3: 
[2021/12/21 09:49:46] root INFO:        mean diff: check passed: True, value: 0.0
[2021/12/21 09:49:46] root INFO: loss_4: 
[2021/12/21 09:49:46] root INFO:        mean diff: check passed: True, value: 0.0
[2021/12/21 09:49:46] root INFO: diff check passed

5次反向运算结果检测通过。

经过上面的对齐检测，可以证明模型复现正确。