-
Notifications
You must be signed in to change notification settings - Fork 51
Open
Description
Hi, thank you for the implementation of WaveGrad. It helps me a lot for implementing WaveFit from Google in the following repository.
https://github.com/yukara-ikemiya/wavefit-pytorch
I happened to find the dilation order of upsampling blocks might be wrong.
The original paper says the dilation factor 1, 2, 4, 8 should be used for the first 3 blocks, but in your implementation, it is used for the last 3 blocks.
I believe this should be modified as below.
wavegrad/src/wavegrad/model.py
Lines 169 to 175 in c3d54c8
| self.upsample = nn.ModuleList([ | |
| UBlock(768, 512, 5, [1, 2, 1, 2]), | |
| UBlock(512, 512, 5, [1, 2, 1, 2]), | |
| UBlock(512, 256, 3, [1, 2, 4, 8]), | |
| UBlock(256, 128, 2, [1, 2, 4, 8]), | |
| UBlock(128, 128, 2, [1, 2, 4, 8]), | |
| ]) |
self.upsample = nn.ModuleList([
UBlock(768, 512, 5, [1, 2, 4, 8]),
UBlock(512, 512, 5, [1, 2, 4, 8]),
UBlock(512, 256, 3, [1, 2, 4, 8]),
UBlock(256, 128, 2, [1, 2, 1, 2]),
UBlock(128, 128, 2, [1, 2, 1, 2]),
])I conducted small scale experiments and confirmed that this modification actually improved the performance.
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
No labels