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

How is the pretrained model obtained? #64

Closed nalzok closed 4 years ago

nalzok commented 4 years ago

Could you please provide the parameters so that we can reproduce it?

tengteng95 commented 4 years ago

In the ReadMe file, we have provided the running script which contains the hyper-parameters we adopt in our experiments and should achieve similar results as our provided model.

nalzok commented 4 years ago

I see. Thanks for the reply!

When running train.py with the provided arguments, we replaced --niter 500 --niter_decay 200 with --niter 100 --niter_decay 40 to reduce running time. However, even with one-fifth of the epochs, there appears to be a raise in multiple loss matrices towards the end (see log files). For example, pair_L1loss went up from ~8 to ~9, and perceptual went up from ~5 to ~6. Also, there doesn't appear to be an obvious decrease in other loss matrices. Initially, I was wondering whether the rest of the epochs are unnecessary.

opt.txt loss_log.txt

Your pre-trained model does yield visually better results than the one we trained with fewer epochs, though, which means they are likely significant. This is kinda interesting.