minor api changes for ferminet layers #3551
Conversation
shaipranesh2
force-pushed
the
ferminet_pretraining_stable
branch
from
b9ed827
to
e45e97e
Compare
| #filling the weights with 1e-9 for faster convergence | ||
| self.v[0].weight.data.fill_(1e-9) | ||
| self.v[0].bias.data.fill_(1e-9) | ||
| #filling the weights with 2.5e-7 for faster convergence |
There was a problem hiding this comment.
Why are we making this change? What is the purpose of the doubling?
There was a problem hiding this comment.
Removed double datatypes as per our previous conversation. (Thought double datatypes will improve precision, but slowed down the model)
| if l == 0 or (self.n_one[l] != self.n_one[l - 1]) or ( | ||
| self.n_two[l] != self.n_two[l - 1]): | ||
| one_electron_tmp[:, i, :] = torch.tanh(self.v[l](f.to( | ||
| torch.float32))) | ||
| one_electron_tmp[:, i, :] = torch.tanh(self.v[l](f)) |
There was a problem hiding this comment.
Can you explain the changes in this section?
There was a problem hiding this comment.
Here the tensor was already in float32, so ".to(torch.float32)" adds redundancy
| @@ -5156,7 +5159,11 @@ def forward(self, one_electron: torch.Tensor, | |||
| dim=2))) * self.pi[one_d_index].T, | |||
| dim=1) | |||
|
|
|||
| return psi_up, psi_down | |||
| d_down = torch.det(psi_down[:, k, :, :].clone()) | |||
There was a problem hiding this comment.
What is the purpose of these changes? Can you add a comment?
There was a problem hiding this comment.
I changed the layer to return the determinant value of orbital matrix and the wavefunction's value. (the determinant calculation when clubbed with this layer, avoids redundant loops in the code)
shaipranesh2
changed the title
Description
Fix #(issue)
Type of change
Please check the option that is related to your PR.
Checklist
yapf -i <modified file>and check no errors (yapf version must be 0.32.0)mypy -p deepchemand check no errorsflake8 <modified file> --countand check no errorspython -m doctest <modified file>and check no errors