pytorch/examples

Gradient vanishing of G in the DCGAN example

Open

#822 opened on Sep 11, 2020

View on GitHub
 (0 comments) (0 reactions) (0 assignees)Python (9,429 forks)batch import
help wanted

Repository metrics

Stars
 (21,634 stars)
PR merge metrics
 (No merged PRs in 30d)

Description

Hello,

I have trained the DCGAN with the default hyper-parameter settings on the downloaded "img_align_celeba" dataset (recommended in the tutorial). However, the results reveal strong gradient vanishing of G. While Loss_D keeps decreasing towards 0, Loss_G grows high (towards 100).

It seems that D is trained so well, preventing a good training on G. I didn't do any modifications on the code. Do you know what happened?

Thanks!

Contributor guide

Research direction
Investigate the default hyperparameters in the DCGAN example. Check the learning rates for generator and discriminator, and consider adjusting them. Look at the loss function implementation to ensure it matches the original DCGAN paper. Additionally, try using different GAN training techniques like label smoothing or adding noise to the discriminator inputs.
Tech stack
python
Domain
machine learning
Issue type
Research
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
3-5 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Needs investigation
Prerequisites
Understanding of GANsFamiliarity with PyTorchBasic knowledge of DCGAN
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
20