Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add support for charged molecules in TensorNet (total charge) #238

Merged
merged 9 commits into from
Nov 3, 2023
Merged

Add support for charged molecules in TensorNet (total charge) #238

Changes from 4 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
e66b3b1
Update tensornet.py for support of total charge q
guillemsimeon Nov 2, 2023
f832d2b
Update tensornet.py for q support
guillemsimeon Nov 2, 2023
20e0bd3
fix
guillemsimeon Nov 2, 2023
88f22f3
fix
guillemsimeon Nov 2, 2023
fbb94e8
fix comment
guillemsimeon Nov 3, 2023
fa26101
initialize zero charge as tensor, move charge broadcasting to interac…
guillemsimeon Nov 3, 2023
9c037da
add clarification comment
guillemsimeon Nov 3, 2023
c030457
trying fix
guillemsimeon Nov 3, 2023
eafd134
try fix
guillemsimeon Nov 3, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 14 additions & 4 deletions torchmdnet/models/tensornet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -217,6 +217,12 @@ def forward(
mask = (edge_index[0] < 0).unsqueeze(0).expand_as(edge_index)
zp = torch.cat((z, torch.zeros(1, device=z.device, dtype=z.dtype)), dim=0)
# I trick the model into thinking that the masked edges pertain to the extra atom
if q is None:
q = 0
else:
batchp = torch.cat((batch, batch[-1] + 1), dim=0)
qp = torch.cat((q, torch.zeros(1, device=q.device, dtype=q.dtype)), dim=0)
q = qp[batchp][...,None,None,None]
RaulPPelaez marked this conversation as resolved.
Show resolved Hide resolved
# WARNING: This can hurt performance if max_num_pairs >> actual_num_pairs
edge_index = edge_index.masked_fill(mask, z.shape[0])
edge_weight = edge_weight.masked_fill(mask[0], 0)
Expand All @@ -227,8 +233,12 @@ def forward(
# I avoid dividing by zero by setting the weight of self edges and self loops to 1
edge_vec = edge_vec / edge_weight.masked_fill(mask, 1).unsqueeze(1)
X = self.tensor_embedding(zp, edge_index, edge_weight, edge_vec, edge_attr)
if q is None:
q = 0
else:
q = q[batch][...,None,None,None]
for layer in self.layers:
X = layer(X, edge_index, edge_weight, edge_attr)
X = layer(X, edge_index, edge_weight, edge_attr, q)
I, A, S = decompose_tensor(X)
x = torch.cat((tensor_norm(I), tensor_norm(A), tensor_norm(S)), dim=-1)
x = self.out_norm(x)
Expand Down Expand Up @@ -379,7 +389,7 @@ def reset_parameters(self):
linear.reset_parameters()

def forward(
self, X: Tensor, edge_index: Tensor, edge_weight: Tensor, edge_attr: Tensor
self, X: Tensor, edge_index: Tensor, edge_weight: Tensor, edge_attr: Tensor, q: Tensor
) -> Tensor:

C = self.cutoff(edge_weight)
Expand All @@ -401,7 +411,7 @@ def forward(
if self.equivariance_invariance_group == "O(3)":
A = torch.matmul(msg, Y)
B = torch.matmul(Y, msg)
I, A, S = decompose_tensor(A + B)
I, A, S = decompose_tensor((1 + 0.1*q)*(A + B))
if self.equivariance_invariance_group == "SO(3)":
B = torch.matmul(Y, msg)
I, A, S = decompose_tensor(2 * B)
Expand All @@ -411,5 +421,5 @@ def forward(
A = self.linears_tensor[4](A.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
S = self.linears_tensor[5](S.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
dX = I + A + S
X = X + dX + torch.matrix_power(dX, 2)
X = X + dX + (1 + 0.1*q) * torch.matrix_power(dX, 2)
return X