Hi everyone, I'm attempting to code the Tacotron speech synthesis system from scratch to make sure I understand it. I'm done implementing the first convolutional filterbank layer and have implemented the max pooling layer, but I don't understand why the authors of chose a max-pooling over time with stride 1. They claim it's to keep the temporal resolution, but my problem is that I think using a stride of 1 is equivalent to just doing nothing and keeping the data as is.
As an example, say we have a matrix in which every time step corresponds to one column:
A= [1,2,3,4;
5,6,7,8;
1,2,3,4];
If we max pool over time with stride 2, we'll have:
B = [2,4;
6,8;
2,4]
Max-pooling with stride one will keep the time resolution but also result in B=A (keep every column). So what's the point of even saying that max-pooling was applied?
I hope my question was clear enough, thank you for reading.
Typically, kernel size (the area over which we look for the max value) is equal to stride (the step with which we move kernel), but it's not always the case. If the kernel is 1x1 and stride is 1, then it is indeed an identity transformation. But if the kernel size is not 1x1, then it's not an identity transformation for stride=1. Such transformation when applied to an image makes it brighter and less detailed while keeping roughly the same image size (depending on padding).
I see, I think the Tacotron paper doesn't provide the kernel size hence my confusion.
Yes it does, in table 1 it mentions Max-pooling: stride=1, width=2. Obviously the width corresponds to the kernel size.
Thank you for your help.
- Badges
- Science topic
- Similar topics
- Physical Sciences
- Materials Science
- Materials
- Nanomaterials