Input filter dimension for the first convolution layer.

[wwang2]

Hi!
Thanks for this implementation!

I have a question about the conv1d layer setting in your implementation here:

molecular-VAE/models.py

Line 12 in 99c3ec7

self.conv1d1 = nn.Conv1d(120, 9, kernel_size=9)

where you specify an input filter size = 120. However, 120 is the length of the padded SMILES strings. The one-hot encoded SMILES string has a dimension of 120 x 35, the filter input filter should be 35 instead of 120 because you are convolving along the SMILES sequence.

This also means that you need to transpose your bached sequence data for the convolution operation to correctly operate on the sequence along the right dimension.

Just want to confirm this detail. If I am right, I can submit a PR to correct this after benchmarking.

[wwang2]

update: i have been consistently getting lower loss after this fix.

[topazape]

Hi, @wwang2.

Thank you for your suggestion. Yes, you are correct. I misunderstood that line, please submit a PR.

Best regards.

[wwang2]

@topazape PR submitted

[topazape]

@wwang2
Thank you for your contribution!!

[wwang2]

no problem. That is my pleasure