move self interaction logic into TP

Author: lucidrains

README.md

@@ -65,9 +65,9 @@ $ python setup.py test
 ## Todo
 
 - [x] move xi and xj separate project and sum logic into Conv class
+- [x] move self interacting key / value production into Conv, fix no pooling in conv with self interaction
 
 - [ ] figure out DTP heuristic
-- [ ] move self interacting key / value production into Conv, fix no pooling in conv with self interaction
 - [ ] start moving some spherical harmonic stuff to cpp or nim
 
 ## Citations

equiformer_pytorch/equiformer_pytorch.py

@@ -212,7 +212,6 @@ def __init__(
             self.kernel_unary[f'({di},{do})'] = PairwiseTP(di, mi, do, mo, edge_dim = edge_dim)
 
         if self_interaction:
-            assert self.pool, 'must pool edges if followed with self interaction'
             self.self_interact = Linear(fiber_in, fiber_out)
 
         self.project_out = project_out
@@ -290,7 +289,12 @@ def forward(
 
         if self.self_interaction:
             self_interact_out = self.self_interact(inp)
-            outputs = residual_fn(outputs, self_interact_out)
+
+            if self.pool:
+                outputs = residual_fn(outputs, self_interact_out)
+            else:
+                self_interact_out = {k: rearrange(v, '... d m -> ... 1 d m') for k, v in self_interact_out.items()}
+                outputs = {degree: torch.cat(tensors, dim = -3) for degree, tensors in enumerate(zip(self_interact_out.values(), outputs.values()))}
 
         if self.project_out:
             outputs = self.to_out(outputs)
@@ -445,8 +449,7 @@ def __init__(
         self.prenorm = Norm(fiber)
 
         self.to_q = Linear(fiber, hidden_fiber)
-        self.to_kv = TP(fiber, kv_hidden_fiber, edge_dim = edge_dim, pool = False, self_interaction = False, splits = splits)
-        self.to_self_kv = Linear(fiber, kv_hidden_fiber) if attend_self else None
+        self.to_kv = TP(fiber, kv_hidden_fiber, edge_dim = edge_dim, pool = False, self_interaction = attend_self, splits = splits)
 
         self.to_out = Linear(hidden_fiber, fiber)
 
@@ -458,7 +461,7 @@ def forward(
         basis,
         mask = None
     ):
-        attend_self, one_head_kv = exists(self.to_self_kv), self.single_headed_kv
+        one_head_kv = self.single_headed_kv
 
         device, dtype = get_tensor_device_and_dtype(features)
         neighbor_indices, neighbor_mask, edges = edge_info
@@ -479,9 +482,6 @@ def forward(
 
         kv_einsum_eq = 'b h i j d m' if not one_head_kv else 'b i j d m'
 
-        if attend_self:
-            self_keyvalues = self.to_self_kv(features)
-
         outputs = {}
 
         for degree, h, scale in zip(features.keys(), self.heads, self.scale):
@@ -492,16 +492,6 @@ def forward(
             if not one_head_kv:
                 kv = rearrange(kv, f'b i j (h d) m -> b h i j d m', h = h)
 
-            if attend_self:
-                self_kv = self_keyvalues[degree]
-
-                if not one_head_kv:
-                    self_kv = rearrange(self_kv, 'b n (h d) m -> b h n 1 d m', h = h)
-                else:
-                    self_kv = rearrange(self_kv, 'b n d m -> b n 1 d m')
-
-                kv = torch.cat((self_kv, kv), dim = -3)
-
             k, v = kv.chunk(2, dim = -2)
 
             sim = einsum(f'b h i d m, {kv_einsum_eq} -> b h i j', q, k) * scale
@@ -567,7 +557,7 @@ def __init__(
 
         # main branch tensor product
 
-        self.to_attn_and_v = TP(fiber, intermediate_fiber, edge_dim = edge_dim, pool = False, self_interaction = False, splits = splits)
+        self.to_attn_and_v = TP(fiber, intermediate_fiber, edge_dim = edge_dim, pool = False, self_interaction = attend_self, splits = splits)
 
         # non-linear projection of attention branch into the attention logits
 

equiformer_pytorch/version.py

@@ -1,3 +1,3 @@
-__version__ = '0.0.21'
+__version__ = '0.0.22'
 
 __cuda_pkg_name__ = f'equiformer_pytorch_cuda_{__version__.replace(".", "_")}'