New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
https://github.com/soroushmehr/sampleRNN_ICLR2017/blob/master/models/three_tier/three_tier.py#L496 #2
Comments
We use convolution instead. It works the same in this case. |
yes. I figured that out after a little looking around. |
I am still confused by this. In the paper, they say they use "linear projections," so I thought you could use |
You are absolutely right, we could have used The thing is, we need to take the time axis into account. We have different layers of RNNs operating at different time scales. If we have two layers of RNNs, the upper one 4x slower than the lower one, the upsampling layer has to transform a For the MLP, the reason we use |
I see. Thanks a lot. I did a SampleRNN, but it runs more than twice as slow as yours. I found it really confusing to try to figure out why exactly my model was so slow using actual profiling tools, but given what you write, it seems super likely that the difference comes down to two things:
|
Yeah, that's probably the reason. Although intuitively, the MLP should have the largest impact on speed, because it gets executed most often (depending on |
Yes, you are right. With the sliding MLP, but still using custom RNN cells, yours is only about 25% faster than mine (for a config similar to the three tier network in the SampleRNN paper but with no weight norm). |
Is it correct that the code does not implement images2neibs? I think its unfold in pytorch?
This line in the original code: https://github.com/soroushmehr/sampleRNN_ICLR2017/blob/master/models/three_tier/three_tier.py#L496
The text was updated successfully, but these errors were encountered: