Support arbitrary number of feature dimensions in latent to discrete transformation #26
Conversation
There was a problem hiding this comment.
Thanks @VolodyaCO. Don't forget to add a an entry to the release-log.
|
@thisac I don't understand very well the reno process for release notes. Is every PR intended to create a new release? |
|
@thisac nvm, I just read "All of the list items in this section are combined when the release notes are rendered" |
VolodyaCO
force-pushed
the
feature/flexible-features-in-latent-to-discrete
branch
2 times, most recently
from
7dd9d16 to
4119f2d
Compare
VolodyaCO
force-pushed
the
feature/flexible-features-in-latent-to-discrete
branch
from
4119f2d to
a3ec9ec
Compare
tests/test_dvae_winci2020.py
Outdated
| if __name__ == "__main__": | ||
| unittest.main() | ||
| unittest.main() | ||
| unittest.main() |
There was a problem hiding this comment.
Is this redundant? If so, please remove it.
There was a problem hiding this comment.
Ah, yes, it is! thanks for noticing.
tests/test_dvae_winci2020.py
Outdated
|
|
||
| def forward(self, x: torch.Tensor) -> torch.Tensor: | ||
| return x * 20 - 10 | ||
| # x is of shape (batch_size, replica_size, l1, l2, ...) |
There was a problem hiding this comment.
What's replica_size here?
There was a problem hiding this comment.
The gumbel softmax sampling is non-deterministic (depends on the random seed). The DVAE can be told how many times to sample from the logits. This number of times is the replica size. This is documented in the forward method of the DVAE, but it's called n_samples instead of replica_size. I will change this accordingly.
There was a problem hiding this comment.
Changed this to n_samples.
VolodyaCO
force-pushed
the
feature/flexible-features-in-latent-to-discrete
branch
from
5225ee0 to
673a581
Compare
fix number of feature dimensions [docs] consistent notation release ignore files in aim testing different number of latent dimensions extend tests to more feature dimensions Update tests/test_dvae_winci2020.py Co-authored-by: Anahita Mansouri Bigvand <[email protected]> Update tests/test_dvae_winci2020.py Co-authored-by: Anahita Mansouri Bigvand <[email protected]> Update tests/test_dvae_winci2020.py Co-authored-by: Anahita Mansouri Bigvand <[email protected]>
VolodyaCO
force-pushed
the
feature/flexible-features-in-latent-to-discrete
branch
from
673a581 to
3b81ea8
Compare
The encoder might not encode a datapoint x to a 1D spin string. In the case of autoencoders where vector quantisation layers are added, encoders might encode a datapoint x to a HxWxC array. In such a case, the default latent to discrete transformation would be wrong, as it expected only a single feature dimension, whereas in this case there are three such feature dimensions.
This PR fixes that by automatically finding the number of feature dimensions in the encoded data representation, and performs the gumbel softmax transformation accordingly, keeping the same number of feature dimensions.