glow-pytorch icon indicating copy to clipboard operation
glow-pytorch copied to clipboard

Published 20 hours ago • verified publisher icon rosinality

Loss value

Open viettmab opened this issue 3 years ago • 2 comments

I have a question about the loss value.

The loss value is the negative log likelihood (bits per dim) in paper, isn't it? When I trained the model, the loss value was much smaller than the best value in paper.

Thank you

viettmab avatar Dec 13 '21 13:12 viettmab

I remember it is due to train-test gap.

rosinality avatar Dec 15 '21 00:12 rosinality

Use $n_bits=8$.

mahaichuan avatar Dec 23 '21 13:12 mahaichuan