Ask for Train Dataloader Composer about dctnet

[kaix90]

Hi Feng,
Thanks very much for your interest. The preprocessing pipeline for validation is defined in the function valloader_upscaled_static in datasets/dataloader_imagenet_dct.py. Please ignore transform4 and transform5 because they are used for debugging only. I will delete these code in the next commit. Thanks.

[HoracceFeng]

Hello, it is me again. I use this code to train my own dataset last night, but no matter how I adjust the hyperparameters, the loss will go to NaN in less than 10 steps.

My dataset is a traffic sign dataset, images are pretty small (around 39 to 96px). I have tried setting different batch size (64, 128 ,156), different image size (64, 128, 256, 512), and learning rate ( 1e1 to 1e4), but all of them fail to train.

Any suggestions?
Thank you.