aRTist

It may be a problem of distributed training. After I choose a single GPU training, there is no abnormality. I hope you guys can provide a solution. I noticed that...

I did check that the number of images was 27000, and I downloaded the data set from [here](https://seungjunnah.github.io/Datasets/reds.html) many times, and then processed it through the script `pre_process_reds_dataset.py` you provided....

> Hi, based on your description, an idea to debug is to dump the img_bytes variable when error occurs. You can print the length or/and save the file with open('rb')....

batch size changed from 4x2 to 1x6