question about coco train dataset
TianyuLee opened this issue · comments
Dear author
according to the source code and train log, the batch size is 4x4, so I use the batch size to train the coco and found that there are 7393(118287/16) iters in one epoch, but there are 7329 iters in train log, so the train set is preprocessd by some operations?
I have not process the original instances_train2017.json