samples_per_epoch

[JiangThea]

I appreciate the insightful work done by the authors on this paper. I have a question regarding the data processing described. According to the paper, during the training phase: there are 250,000 iterations, 50 epochs, and a batch size of 32, which implies that each epoch consists of 8,000,000 images. For the pre-training phase: there are 60,000 iterations, 300 epochs, and a batch size of 32, suggesting each epoch consists of 1,920,000 images. However, in the code, samples_per_epoch is set to 50,000, meaning each epoch only contains 50,000/32 iterations. Could you clarify which scenario reflects the actual basis for your training process?