search

NVlabs / few-shot-vid2vid

Pytorch implementation for few-shot photorealistic video-to-video translation.
Other
1.8k stars 275 forks source link

Validation Results Are Good But Test Results Are Unsatisfactory (Pose) #63

Closed erkankaracakan closed 3 years ago

erkankaracakan commented 4 years ago

Thanks for this excellent works.

I am trying to train for pose. My dataset has 3 sequences and nearly 40000 frames. Train is at 99 epochs now. Validation results are very good. But when i take a test, results are unsatisfactory. Hands like ghost and face is very bad. Shall I wait or is there a problem?

Train code that i used is here. python train.py --name pose --dataset_mode fewshot_pose --adaptive_spade --warp_ref --spade_combine --niter_single 150 --niter 200 --batchSize 1 --lr 0.00004

Thanks a lot.

cszy98 commented 4 years ago

maybe "--remove_face_labels" and "--add_face_D" will be useful.

erkankaracakan commented 4 years ago

maybe "--remove_face_labels" and "--add_face_D" will be useful.

I trained until 60 epoch using them. But faces were terrible in validation results, like the following issue. So i have started a new train without them. Now, validation results are good but test results not.

https://github.com/NVlabs/few-shot-vid2vid/issues/24

deucalionAlpha commented 4 years ago

Have you trained the face model? Have you meet the problem at this link https://github.com/NVlabs/few-shot-vid2vid/issues/64? or can you share the face-trained model with me? thank you!

tcwang0509 commented 3 years ago

This repo is now deprecated. Please refer to the new Imaginaire repo: https://github.com/NVlabs/imaginaire.