pytorch_MELM icon indicating copy to clipboard operation
pytorch_MELM copied to clipboard

refined_loss_1 is nan

Open fanglinpu opened this issue 5 years ago • 3 comments

The refined_loss_1 is nan when I train the model using vgg16 backbone.

fanglinpu avatar Apr 23 '19 07:04 fanglinpu

Maybe the learning rate is larger. And you can try learning rate 0.001 which is the same as my settings'.

vasgaowei avatar Apr 23 '19 07:04 vasgaowei

I use the learning rate 0.001, but the refined_loss_1 is nan.

fanglinpu avatar Apr 23 '19 09:04 fanglinpu

Make sure you download the right version of pretrained 'vgg16.pth' file. And I give the link of the version of pretrained 'vgg16.pth' file which woks fine.

vasgaowei avatar May 24 '19 01:05 vasgaowei