Giters

ap229997 / DRNET

PyTorch implementation of the NIPS 2017 paper - Unsupervised Learning of Disentangled Representations from Video

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Why the rec_loss returned from function train_main_network() is divided by args.batch_size in main.py?

Dawn90 opened this issue · comments

commented

  1. Why the rec_loss returned from function train_main_network() is divided by args.batch_size in main.py? As we know, nn.MSELoss() already returns the mean loss by default.
    line 183: return sim_loss.data/args.batch_size, rec_loss.data/args.batch_size

  2. Are the default parameters just right for the MNIST dataset? I run python main.py for 4500 iterations with the default parameters, and the reconstruction loss had converged to 0.000126. But the visualization of the reconstructed image is almost a black picture. I'm really confused where I make a mistake. Could you give me any instructions?

Thank you verrrrrry much!
Best,
Dong