Train epochs

[llxllu]

Thank you for your excellent work! The paper mentions a total of 50 training epochs, but I used two 4090 graphics cards and the same batch size according to the experimental setup. Why are there 100 epochs for training the same number of steps?

[jjlinghu]

I also noticed the issue regarding the number of training epochs. However, I believe that 50 epochs can approximately achieve the performance reported in the author's paper. The aim of training for 100 epochs probably ensures that the training log becomes stable.