Did you experiment with different batch sizes in training? about stylegestures HOT 2 CLOSED

Hey @simonalexanderson,
Thank you for your answer. Yes, I saw Gesticulator 👍 Accumulating Gradients and Gradient Checkpointing both seem like a good idea, thanks for the hint.
I trained your model with batch size of 10 to see if this gives the same effect as with the input feature space which I extended by BERT features (the one I called "fast-forwarding").
The clear answer: No, it does not. The motion of the generated gestures look as smooth as with a batch size of 80. Of course, they most likely differ in other aspects in comparison to the training with a batch size of 80, I didn't check on these. But they definitely do not show the same behavior as with the integration of BERT features (i.e., "fast-fowarding").

from stylegestures.

Hi @stefan-kohnen ,
Adding BERT features sounds very interesting and I would love to see what the results will show. I guess you have seen our work with Gesticulator (Kucherenko et al 2020), which has a different architecture but also uses semantic features. We have not experimented so much with batch sizes, but tend to use the maximum for our gpus, which is typically 50-100. 10 seems low to me. You may experiment with accumulating gradients over several batches (search for "accumulating gradients") or checkpointing the rnn in the coupling layer: https://github.com/prigoyal/pytorch_memonger/blob/master/tutorial/Checkpointing_for_PyTorch_models.ipynb. The later should give a large memory improvement with only a few lines of code. Please share if you get any insights on how batch size effects training.

from stylegestures.