Unsup_Recycle_GAN
Unsup_Recycle_GAN copied to clipboard
wangkaihong
Training Results
I tried to replicate your trained weights on Viper to CityScapes. I am wondering if getting those reconstructed images after 8000 iterations is expected, which does not seem right to me.
where the first image is fake Viper and the second one is fake CityScapes.
If this is not normal, is there any step where I might go wrong? I used the same training command you provided, which is
python train.py --dataroot path/to/data/ --model unsup_single --dataset_mode unaligned_scale --name v2c_experiment --loadSizeW 542 --loadSizeH 286 --resize_mode rectangle --fineSizeW 512 --fineSizeH 256 --crop_mode rectangle --which_model_netG resnet_6blocks --no_dropout --pool_size 0 --lambda_spa_unsup_A 10 --lambda_spa_unsup_B 10 --lambda_unsup_cycle_A 10 --lambda_unsup_cycle_B 10 --lambda_cycle_A 0 --lambda_cycle_B 0 --lambda_content_A 1 --lambda_content_B 1 --batchSize 1 --noise_level 0.001 --niter_decay 0 --niter 2
This doesn't seem right to me too. However, I am not able to pinpoint the issue based on this either. Given that the experiments had been successfully reproduced by other users, maybe you can try to focus more on the data and the environment side? Or maybe re-runing the model with a different random seed may simply help.
Is there a particular reason why you turned off cycle loss in the provided training command? I saw in Table 11 turning on cycle loss will give lower warping error.
There isn't a strong reason to turn that off; you can see that both having or not the cycle loss didn't change the performance significantly, so we didn't use it for the sake of simplicity. Feel free to turn it on if you feel like it might be helpful in your case.
