search

auspicious3000 / autovc

AutoVC: Zero-Shot Voice Style Transfer with Only Autoencoder Loss
https://arxiv.org/abs/1905.05879
MIT License
987 stars 205 forks source link

About the batch size #11

Closed nkcdy closed 5 years ago

nkcdy commented 5 years ago

The paper said the batch size is 2 for 100k steps. Would it be too small for batch normalization? are you sure the batch size is 2 in your training?

auspicious3000 commented 5 years ago

Yes, the batch size is 2.

barby1138 commented 5 years ago

nkcdy - think why there is x.squeeze!1) in input and x.unsqueeze!1) in output Also BatchNorm in small batch size seems to give lower noise.

nkcdy commented 5 years ago

I can't figure out why there is a squeeze at the input and an unaqueeze at the output. I also can't see any difference in the data before and after squeeze except the transpose operation during training. Could you please provide some prompt?

auspicious3000 commented 5 years ago

squeeze and unsqueeze only add or remove dimensions, please refer to the doc of pytorch for more details

nkcdy commented 5 years ago

yes， I know exactly the meaning of squeeze and unsqueeze. But Barby1138 implied there seems to exist some relationhsip between these two operations and the small size batchmorm. I just got confused.

auspicious3000 commented 5 years ago

They are not related