search

NVlabs / few-shot-vid2vid

Pytorch implementation for few-shot photorealistic video-to-video translation.
Other
1.79k stars 276 forks source link

Identical pictures after inference. #9

Open paniquex opened 4 years ago

paniquex commented 4 years ago

I have trained your model for 5 epoches on pose dataset with script ./scripts/pose/train_g1.sh. After this I ran ./scripts/pose/test.sh and got 30 identical pictures in ./results/pose/test_lates_finetune/02-0_01/synthesized/. Is it ok and I have just need to wait more? How many epoches I need to wait, before I could see any changes?

tcwang0509 commented 4 years ago

Hmm that's strange. Does the input image change in the results? It might be a potential bug.

paniquex commented 4 years ago

I have been waiting for 100 epoches and got results, which aren't similar to reference images. But on 101 and after I've got only black picture as result. Is it ok?

FreedomGu commented 4 years ago

I meet same problem as @paniquex meet, after i trained 17 epochs, it occurs wired squares and didnt coverage to tgt image ( i mean the pose) , hope to hear your update, thanks!

songyn95 commented 2 years ago

@tcwang0509 @paniquex @FreedomGu Sorry to bother you. I encountered the same problem when implementing this method. When the training iteration reached 14/70, the composite image turned black. I want to know what the reason is and how to solve it?