About batch size (1024)

[shawnpark07]

Hello, thanks for your impressive work :)

While comprehending your work, here comes one ambiguity.

As you know, H36M train dataset has total 1,559,752 frames, and these would be grouped into 6,716 sequences (243 frames / sequence). But if we choose batch size of 1,024 as mentioned in the paper, there would be only 6~7 batches per one epoch.

Did I understand right? If so, I think the training would be very unstable due to small number of batches per epoch. Will looking forward to your answer. Thanks!

[JinluZhang1126]

Hi, 1024 means frame numbers of each batch, we apply stride sample to separate 1024 into 5 periods, therefore the overall iterations will be about 1300 in our code.

[flyyyyer]

Your paper said the bachsize is 1024, but actually batchsize is 1024//243=4 in the tranning. It seems to reduce the batchsize to achive large iterations.

[JinluZhang1126]

Your paper said the bachsize is 1024, but actually batchsize is 1024//243=4 in the tranning. It seems to reduce the batchsize to achive large iterations.

Yes, the batch size is actually 4, and the input sequence length is 1024, which may be confused.

[Aerikervalid]

If I increase batchsize does that mean the effect is increased and the model is more stable，But the large bachsize was worse in several experiments. Why do the authors think this is?