search

junyanz / pytorch-CycleGAN-and-pix2pix

Image-to-Image Translation in PyTorch
Other
22.8k stars 6.29k forks source link

wandb & visdom question #1608

Open xia-QQ opened 11 months ago

xia-QQ commented 11 months ago

I used visdom and wandb to record the training process.But steps is not the same,For visdom ,it's xaxis have 200 epochs,but in wandb have 1200 steps.Why is there such a problem.

JustinasLekavicius commented 9 months ago

Sorry for the late response, but I believe the amount of steps depends on your batch size and number of input samples. I think it may also be dependant on training options such as display_freq and print_freq.

If for example you have 200 epochs, but you have batch size of 1 and 6 input samples, then you'll have 1200 steps. Unless you have more samples than that, then the wandb steps value may be impacted by display_freq and print_freq values (how often the results are logged to Weights and Biases).