search

tengteng95 / Pose-Transfer

Code for the paper Progressive Pose Attention for Person Image Generation in CVPR19 (Oral).
MIT License
734 stars 160 forks source link

Bad IS and SSIM score #51

Closed dawn7564 closed 4 years ago

dawn7564 commented 4 years ago

Really really thanks for your awesome works.

However, using your code and hyper-parameter, I got bad evaluation results after training by myself: SSIM=0.268 IS=3.652 mask-SSIM=0.792 mask-IS=3.733

Meanwhile, using your provided pre-trained model, I got almost right results: SSIM=0.311 IS=3.329 mask-SSIM=0.811 mask-IS=3.778

My training environment is pytorch0.3.1, I don't know why the results are different.

Thanks again~

tengteng95 commented 4 years ago

Hi, it is really strange to obtain such an unsatisfying result. I could not tell where is wrong at the moment since too few information is provided. Do you make any modifications to the original code? Does the training process converge well?

dawn7564 commented 4 years ago

Hi, thanks for replying : ) I didn't modify any code, I just download the vgg19-dcbb9e9d.pth by myself and put it on the .torch/models/ because it's slow downloading by program.

and training process is here pytorch0-3-1train0-3-1

My training and testing environment(conda): pytorch0.3.1 cuda90; torchvision0.2.0; python3.6.9;

By the way, I also use pytorch1.3 to run codes pytorch_v1.0 and get almost the same bad results...

chengaopro commented 4 years ago

Really really thanks for your awesome works.

However, using your code and hyper-parameter, I got bad evaluation results after training by myself: SSIM=0.268 IS=3.652 mask-SSIM=0.792 mask-IS=3.733

Meanwhile, using your provided pre-trained model, I got almost right results: SSIM=0.311 IS=3.329 mask-SSIM=0.811 mask-IS=3.778

My training environment is pytorch0.3.1, I don't know why the results are different.

Thanks again~

Hi, I strictly followed their instruction and environment, use their provided pre-trained model, but I only got IS=3.159, mask-IS=3.735, SSIM=0.2815, mask-SSIM=0.799 on Market-1501 dataset, which is worse than those in original paper. I saw you reproduced their performance, so what do you think is the cause of this gap? @dawn7564 @tengteng95 By the way, I have reproduced their performance on DeepFashion dataset. But the issue in Market-1501 dataset has troubled me for a long time... , thanks

dawn7564 commented 4 years ago

Hi, i have got the almost right results now by redownloading the dataset, files, and building new conda environment. Frankly, i dont know why the results come right after that. Maybe you can try my way.

I use pytorch v1.0.1, python3.6 for training and testing, if it helps.