search

LiuLei95 / PyTorch-Learned-Image-Compression-with-GMM-and-Attention

This repo is implementation for Learned Image Compression with Discretized Gaussian Mixture Likelihoods and Attention Modules in pytorch.
Apache License 2.0
65 stars 3 forks source link

R-D performance #4

Closed lilala0 closed 2 years ago

lilala0 commented 2 years ago

Hi~Thanks for your sharing code. When i train this network with Flickr_2W datasets as your description, that is: A、Crop the Flicker_2W images randomly and get an new dataset with 82G B、Train the 4096_256 modle i encountered two problems as follow： 1) I cannot achieve the excellent R-D loss as you, So do you think what i should adjust to achieve the same R-D loss as you? 2) Sometimes, the R-D loss is nan. I do not know why this happen. 3) In datasets.py, the methods of training data augmentation only includes RandomHorizontalFlip and RandomVerticalFlip? Did you use RandomResizedCrop and Normalize in your modle training?

LiuLei95 commented 2 years ago
  1. You can train the network used the pre-trained model continue, which is saved before R-D loss is nan. Sometimes, the R-D loss is nan, espectly in high bit-rate.
  2. RandomResizeCrop is used in the A、Crop the Flicker_2W images randomly and get an new dataset with 82G. And Normalize we don't used.
lilala0 commented 2 years ago

Ok, i will try your advices. Again, Thanks for your reply.