Fix TensorProductConv test and improve docs (#4480)

Closes #4459

Authors:
  - Tingyu Wang (https://github.com/tingyu66)
  - Ralph Liu (https://github.com/nv-rliu)

Approvers:
  - https://github.com/DejunL
  - Rick Ratzel (https://github.com/rlratzel)

URL: #4480

Author: tingyu66

python/cugraph-equivariant/cugraph_equivariant/nn/tensor_product_conv.py

@@ -11,7 +11,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Optional, Sequence, Union
+from typing import Optional, Sequence, Union, NamedTuple
 
 import torch
 from torch import nn
@@ -31,6 +31,11 @@
     ) from exc
 
 
+class Graph(NamedTuple):
+    edge_index: torch.Tensor
+    size: tuple[int, int]
+
+
 class FullyConnectedTensorProductConv(nn.Module):
     r"""Message passing layer for tensor products in DiffDock-like architectures.
     The left operand of tensor product is the spherical harmonic representation
@@ -81,27 +86,35 @@ class FullyConnectedTensorProductConv(nn.Module):
 
     Examples
     --------
-    >>> # Case 1: MLP with the input layer having 6 channels and 2 hidden layers
-    >>> #         having 16 channels. edge_emb.size(1) must match the size of
-    >>> #         the input layer: 6
-    >>>
+    Case 1: MLP with the input layer having 6 channels and 2 hidden layers
+    having 16 channels. edge_emb.size(1) must match the size of the input layer: 6
+
     >>> conv1 = FullyConnectedTensorProductConv(in_irreps, sh_irreps, out_irreps,
     >>>     mlp_channels=[6, 16, 16], mlp_activation=nn.ReLU()).cuda()
     >>> out = conv1(src_features, edge_sh, edge_emb, graph)
-    >>>
-    >>> # Case 2: Same as case 1 but with the scalar features from edges, sources
-    >>> #         and destinations passed in separately.
-    >>>
+
+    Case 2: If `edge_emb` is constructed by concatenating scalar features from
+    edges, sources and destinations, as in DiffDock, the layer can accept each
+    scalar component separately:
+
     >>> conv2 = FullyConnectedTensorProductConv(in_irreps, sh_irreps, out_irreps,
     >>>     mlp_channels=[6, 16, 16], mlp_activation=nn.ReLU()).cuda()
-    >>> out = conv3(src_features, edge_sh, edge_scalars, graph,
+    >>> out = conv2(src_features, edge_sh, edge_scalars, graph,
     >>>     src_scalars=src_scalars, dst_scalars=dst_scalars)
-    >>>
-    >>> # Case 3: No MLP, edge_emb will be directly used as the tensor product weights
-    >>>
+
+    This allows a smaller GEMM in the first MLP layer by performing GEMM on each
+    component before indexing. The first-layer weights are split into sections
+    for edges, sources and destinations, in that order.This is equivalent to
+
+    >>> src, dst = graph.edge_index
+    >>> edge_emb = torch.hstack((edge_scalars, src_scalars[src], dst_scalars[dst]))
+    >>> out = conv2(src_features, edge_sh, edge_emb, graph)
+
+    Case 3: No MLP, `edge_emb` will be directly used as the tensor product weights:
+
     >>> conv3 = FullyConnectedTensorProductConv(in_irreps, sh_irreps, out_irreps,
     >>>     mlp_channels=None).cuda()
-    >>> out = conv2(src_features, edge_sh, edge_emb, graph)
+    >>> out = conv3(src_features, edge_sh, edge_emb, graph)
 
     """
 
@@ -174,20 +187,20 @@ def forward(
             Edge embeddings that are fed into MLPs to generate tensor product weights.
             Shape: (num_edges, dim), where `dim` should be:
             - `tp.weight_numel` when the layer does not contain MLPs.
-            - num_edge_scalars, with the sum of num_[edge/src/dst]_scalars being
-              mlp_channels[0]
+            - num_edge_scalars, when scalar features from edges, sources and
+              destinations are passed in separately.
 
         graph : tuple
             A tuple that stores the graph information, with the first element being
             the adjacency matrix in COO, and the second element being its shape:
             (num_src_nodes, num_dst_nodes).
 
         src_scalars: torch.Tensor, optional
-            Scalar features of source nodes.
+            Scalar features of source nodes. See examples for usage.
             Shape: (num_src_nodes, num_src_scalars)
 
         dst_scalars: torch.Tensor, optional
-            Scalar features of destination nodes.
+            Scalar features of destination nodes. See examples for usage.
             Shape: (num_dst_nodes, num_dst_scalars)
 
         reduce : str, optional (default="mean")

python/cugraph-equivariant/cugraph_equivariant/tests/test_tensor_product_conv.py

@@ -13,10 +13,6 @@
 
 import pytest
 
-import torch
-from torch import nn
-from e3nn import o3
-
 try:
     from cugraph_equivariant.nn import FullyConnectedTensorProductConv
 except RuntimeError:
@@ -25,9 +21,29 @@
         allow_module_level=True,
     )
 
-device = torch.device("cuda:0")
+import torch
+from torch import nn
+from e3nn import o3
+from cugraph_equivariant.nn.tensor_product_conv import Graph
+
+device = torch.device("cuda")
 
 
+def create_random_graph(
+    num_src_nodes,
+    num_dst_nodes,
+    num_edges,
+    dtype=None,
+    device=None,
+):
+    row = torch.randint(num_src_nodes, (num_edges,), dtype=dtype, device=device)
+    col = torch.randint(num_dst_nodes, (num_edges,), dtype=dtype, device=device)
+    edge_index = torch.stack([row, col], dim=0)
+
+    return Graph(edge_index, (num_src_nodes, num_dst_nodes))
+
+
+@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
 @pytest.mark.parametrize("e3nn_compat_mode", [True, False])
 @pytest.mark.parametrize("batch_norm", [True, False])
 @pytest.mark.parametrize(
@@ -39,9 +55,10 @@
     ],
 )
 def test_tensor_product_conv_equivariance(
-    mlp_channels, mlp_activation, scalar_sizes, batch_norm, e3nn_compat_mode
+    mlp_channels, mlp_activation, scalar_sizes, batch_norm, e3nn_compat_mode, dtype
 ):
     torch.manual_seed(12345)
+    to_kwargs = {"device": device, "dtype": dtype}
 
     in_irreps = o3.Irreps("10x0e + 10x1e")
     out_irreps = o3.Irreps("20x0e + 10x1e")
@@ -55,68 +72,65 @@ def test_tensor_product_conv_equivariance(
         mlp_activation=mlp_activation,
         batch_norm=batch_norm,
         e3nn_compat_mode=e3nn_compat_mode,
-    ).to(device)
+    ).to(**to_kwargs)
 
     num_src_nodes, num_dst_nodes = 9, 7
     num_edges = 40
-    src = torch.randint(num_src_nodes, (num_edges,), device=device)
-    dst = torch.randint(num_dst_nodes, (num_edges,), device=device)
-    edge_index = torch.vstack((src, dst))
-
-    src_pos = torch.randn(num_src_nodes, 3, device=device)
-    dst_pos = torch.randn(num_dst_nodes, 3, device=device)
-    edge_vec = dst_pos[dst] - src_pos[src]
-    edge_sh = o3.spherical_harmonics(
-        tp_conv.sh_irreps, edge_vec, normalize=True, normalization="component"
-    ).to(device)
-    src_features = torch.randn(num_src_nodes, in_irreps.dim, device=device)
+    graph = create_random_graph(num_src_nodes, num_dst_nodes, num_edges, device=device)
+
+    edge_sh = torch.randn(num_edges, sh_irreps.dim, **to_kwargs)
+    src_features = torch.randn(num_src_nodes, in_irreps.dim, **to_kwargs)
 
     rot = o3.rand_matrix()
-    D_in = tp_conv.in_irreps.D_from_matrix(rot).to(device)
-    D_sh = tp_conv.sh_irreps.D_from_matrix(rot).to(device)
-    D_out = tp_conv.out_irreps.D_from_matrix(rot).to(device)
+    D_in = tp_conv.in_irreps.D_from_matrix(rot).to(**to_kwargs)
+    D_sh = tp_conv.sh_irreps.D_from_matrix(rot).to(**to_kwargs)
+    D_out = tp_conv.out_irreps.D_from_matrix(rot).to(**to_kwargs)
 
     if mlp_channels is None:
-        edge_emb = torch.randn(num_edges, tp_conv.tp.weight_numel, device=device)
+        edge_emb = torch.randn(num_edges, tp_conv.tp.weight_numel, **to_kwargs)
         src_scalars = dst_scalars = None
     else:
         if scalar_sizes:
-            edge_emb = torch.randn(num_edges, scalar_sizes[0], device=device)
+            edge_emb = torch.randn(num_edges, scalar_sizes[0], **to_kwargs)
             src_scalars = (
                 None
                 if scalar_sizes[1] == 0
-                else torch.randn(num_src_nodes, scalar_sizes[1], device=device)
+                else torch.randn(num_src_nodes, scalar_sizes[1], **to_kwargs)
             )
             dst_scalars = (
                 None
                 if scalar_sizes[2] == 0
-                else torch.randn(num_dst_nodes, scalar_sizes[2], device=device)
+                else torch.randn(num_dst_nodes, scalar_sizes[2], **to_kwargs)
             )
         else:
-            edge_emb = torch.randn(num_edges, tp_conv.mlp[0].in_features, device=device)
+            edge_emb = torch.randn(num_edges, tp_conv.mlp[0].in_features, **to_kwargs)
             src_scalars = dst_scalars = None
 
     # rotate before
+    torch.manual_seed(12345)
     out_before = tp_conv(
         src_features=src_features @ D_in.T,
         edge_sh=edge_sh @ D_sh.T,
         edge_emb=edge_emb,
-        graph=(edge_index, (num_src_nodes, num_dst_nodes)),
+        graph=graph,
         src_scalars=src_scalars,
         dst_scalars=dst_scalars,
     )
 
     # rotate after
+    torch.manual_seed(12345)
     out_after = (
         tp_conv(
             src_features=src_features,
             edge_sh=edge_sh,
             edge_emb=edge_emb,
-            graph=(edge_index, (num_src_nodes, num_dst_nodes)),
+            graph=graph,
             src_scalars=src_scalars,
             dst_scalars=dst_scalars,
         )
         @ D_out.T
     )
 
-    torch.allclose(out_before, out_after, rtol=1e-4, atol=1e-4)
+    atol = 1e-3 if dtype == torch.float32 else 1e-1
+    if e3nn_compat_mode:
+        assert torch.allclose(out_before, out_after, rtol=1e-4, atol=atol)