I noticed that there are two batch size args for evaluation in training args. What’s the difference between the two? It seems eval_batch_size is still used as per-device evaluation batch size in the trainer source codes?
1 Like