Couldn't replicate results

[Ibrahimatef]

Hi Again, I was trying again your code to make sure everything is right, I found that issue was closed :( Unfortunately, I couldn't solve the problem I got following results with saving person image, garment, flow, warped garment, and try-on image :

Do you know what mistake could lead to this poor output or you could check the performance of uploaded pretrained models ? Here is command i used to replicate the results :

!python test.py --name demo \
                --resize_or_crop None --batchSize 1 --gpu_ids 0 \
                --warp_checkpoint ckp/non_aug/PFAFN_warp_epoch_101.pth \
                --gen_checkpoint ckp/non_aug/PFAFN_gen_epoch_101.pth \
                --dataroot 'VITON_test'

Also I think results are saved in our_t_results or im_gar_flow_wg folders which i have to create them manually

[SenHe]

I replied to you. Please check this link: https://github.com/SenHe/Flow-Style-VTON/issues/2#issuecomment-1092735565

[Ibrahimatef]

I have read your reply and couldn't get your results though, that is why I opened this issue to check if something may be wrong

[SenHe]

I don't know what's wrong in your side - as I can reproduce the results with the GitHub code and the downloaded checkpoint. What's your PyTorch version?

[Ibrahimatef]

I am using pytorch version 1.10 with cuda 11.2

[SenHe]

OK.

I checked my environment for training and inference: Python: 3.6.13 Torch: 1.1.0 Torchvision: 0.3.0

[Ibrahimatef]

Thank your for your help, I will check results with this environment

[Ibrahimatef]

I tried with your environment versions and results were very good, Thanks for your efforts and great work!

[taruntiwarihp]

Hi, I tried your environment version and I'm getting the very best results :(, Please guide me Python: 3.6.13 Torch: 1.1.0 Torchvision: 0.3.0 Cuda 11.4

[tuallen]

Add align_corners=True to all occurrences of F.grid_sample. In newer versions of PyTorch, F.grid_sample now has align_corners set to False by default.

I obtained the same results as @SenHe on a RTX 3090 and the latest versions of Torch, Torchvision, and CUDA after adding the above change.

[caip1299920300]

After version 1.2.0 ： Add ”align_corners=True“ parameter to all “F.grid_sample(）” functions in the code to solve the problem。 such as：flow = F.grid_sample(last_flow, flow, mode='bilinear', padding_mode='border',align_corners=True) Get good results：

[89lixx]

The FID is quite different from a RTX 3090. I got 9.37. Did you get the FID as the paper mentioned? @tuallen @caip1299920300

[caip1299920300]

yes,8.89  

☆·．快~~乐·╮ @.***

 

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2023年2月21日(星期二) 下午4:50 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [SenHe/Flow-Style-VTON] Couldn't replicate results (Issue #3)

The FID is quite different from a RTX 3090. I got 9.37. Did you get the FID as the paper mentioned? @tuallen @caip1299920300

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

[89lixx]

thanks!