search

twtygqyy / pytorch-SRResNet

pytorch implementation for Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network arXiv:1609.04802
MIT License
481 stars 130 forks source link

Tips for training #15

Open leviethung2103 opened 6 years ago

leviethung2103 commented 6 years ago

Hello, I would like to ask you about your experience in training the dataset. About choosing the iteration 500 (training will take about one week), size_average in MSE_Loss ? Why you set up size_average = False ? Thank you

twtygqyy commented 6 years ago

Hi, if you set size_average = True, then the loss would become very small, and when you use the loss for gradient calculation, the gradient will also be very small. No good for the convergence speed.