Results seem a little off when trained from scratch

[varunjain3]

Hi, I trained the model using the given model script and the hyperparameters for IFRnet(base model) on 8xV100 and 4xV100s. I am getting a PSNR of around 34.5 on the 8GPUs trained model and 35 on the 4 GPUs trained model.

Are there some hyperparameters that I should change to reproduce the results given in the paper? Also, any intuition on the results varies. I m using the vimeo90k dataset for training.

I am using a batchsize of 55, my V100 are 32gigs

[ltkong218]

Thanks for your question. I think the total training iterations should keep the same whether you use 4 or 8 GPUs.

You can try to double the training epochs when using double GPUs. In this situation, the training iterations can be the same.