Model does not train - Githubissues

UBCDingXin / improved_CcGAN

Continuous Conditional Generative Adversarial Networks (CcGAN)

https://arxiv.org/abs/2011.07466

MIT License

121 stars 35 forks source link

Model does not train #5

Closed haoming-codes closed 3 years ago

haoming-codes commented 3 years ago

Dear authors of CcGAN, thank you for sharing the code! I'm running CcGAN Hard in improved_CcGAN/UTKFace/UTKFace_64x64/CcGAN/scripts/run_train.sh (which hopefully is the setting used in the paper) directly in Google Colab, and observe that g_loss and d_loss (logged here) stays about constant (at about .8 and 1.3, respectively) throughout the 40,000 iterations. Is this the expected behavior?

UBCDingXin commented 3 years ago

Hi, Thanks a lot for your interest in our work. Regarding your question, I suggest printing some generated images during training (e.g. print some images for some ages every 1000 iterations). If the image quality gradually improves, the training should be okay. Sometimes the d loss and g loss are not very informative. Hopefully, it can help. Please let me know if there are any other issues. Best,

Xin