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

rewrite the mace equivariant head to avoid a model explosion #52

Merged
merged 2 commits into from
Jun 5, 2024
Merged

rewrite the mace equivariant head to avoid a model explosion #52

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c560292
rewrite the mace equivariant head to avoid a model explosion
sblackburn-mila Jun 3, 2024
3f55d17
whiteline and useless variable deleted
sblackburn-mila Jun 3, 2024
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
20 changes: 13 additions & 7 deletions crystal_diffusion/models/score_prediction_head.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -91,6 +91,7 @@ def __init__(self, output_node_features_irreps: o3.Irreps,
"""Init method."""
super().__init__(output_node_features_irreps, hyper_params)

# raise the dimensionality of the time scalar
time_irreps_in = o3.Irreps("1x0e") # time is a scalar
time_irreps_out = o3.Irreps(hyper_params.time_embedding_irreps)

Expand All @@ -103,9 +104,18 @@ def __init__(self, output_node_features_irreps: o3.Irreps,
# by the various subsequent layers.
head_input_irreps = time_irreps_out + output_node_features_irreps

sorted_irreps, self.column_permutation_indices = (
sorted_irreps, _ = (
get_normalized_irreps_permutation_indices(head_input_irreps))

# mix the time embedding 0e irrep with the different components of the MACE output and avoid a dimensionality
# explosion with FullyConnectedTensorProduct that only computes the specified channels in the output and applies
# linear layers to control the dimensionality across each irrep
self.time_mixing_layer = o3.FullyConnectedTensorProduct(
irreps_in1=time_irreps_out,
irreps_in2=output_node_features_irreps,
irreps_out=sorted_irreps
)

self.head = torch.nn.Sequential()

for _ in range(hyper_params.number_of_layers):
Expand All @@ -115,10 +125,6 @@ def __init__(self, output_node_features_irreps: o3.Irreps,
batch_norm = e3nn.nn.BatchNorm(irreps=sorted_irreps)
self.head.append(batch_norm)

# Some sort of tensor product is necessary to mix the scalar time with the vector output.
product_layer = o3.TensorSquare(irreps_in=sorted_irreps, irreps_out=sorted_irreps)
self.head.append(product_layer)

gate = gate_dict[hyper_params.gate]
# There must be one activation per L-channel, and that activation must be 'None' for l !=0.
# The first channel is l = 0.
Expand All @@ -144,8 +150,8 @@ def forward(self, flat_node_features: torch.Tensor, flat_times: torch.Tensor) ->
flat_scores: scores computed using the head. Dimension [batch_size * number_of_atoms, spatial_dimension]
"""
embedded_times = self.time_embedding_linear_layer(flat_times)
head_input = torch.cat([embedded_times, flat_node_features], dim=1)[:, self.column_permutation_indices]
flat_scores = self.head(head_input)
mixed_features = self.time_mixing_layer(embedded_times, flat_node_features)
flat_scores = self.head(mixed_features)

return flat_scores

Expand Down
Loading