Merge pull request #399 from bigict/model

feat: predict with conditional coords

Author: chungongyu

profold2/command/trainer.py

@@ -214,6 +214,9 @@ def create_cycling_data(
       accept_msa_attn=args.model_evoformer_accept_msa_attn,
       accept_frame_attn=args.model_evoformer_accept_frame_attn,
       accept_frame_update=args.model_evoformer_accept_frame_update,
+      conditional_pos=args.model_conditional_pos,
+      recycling_frames=args.model_recycling_frames,
+      recycling_pos=args.model_recycling_pos,
       headers=headers
   )
   ####
@@ -743,6 +746,11 @@ def add_arguments(parser):  # pylint: disable=redefined-outer-name
       default='model_headers_main.json',
       help='json format headers of model.'
   )
+  parser.add_argument(
+      '--model_conditional_pos',
+      action='store_true',
+      help='enable backbone atom position conditional.'
+  )
   parser.add_argument(
       '--model_recycles', type=int, default=2, help='number of recycles in model.'
   )

profold2/model/alphafold2.py

@@ -144,6 +144,10 @@ def __init__(
       accept_msa_attn=True,
       accept_frame_attn=False,
       accept_frame_update=False,
+      conditional_pos=False,
+      conditional_pos_min_bin=3.25,
+      conditional_pos_max_bin=20.75,
+      conditional_pos_num_bin=15,
       recycling_single_repr=True,
       recycling_frames=False,
       recycling_pos=False,
@@ -186,6 +190,18 @@ def __init__(
     # msa to single activations
     self.to_single_repr = nn.Linear(dim_msa, dim_single)
 
+    # conditional params
+    self.conditional_pos_linear = nn.Linear(
+        recycling_pos_num_bin, dim_pairwise
+    ) if conditional_pos else None
+    if conditional_pos:
+      conditional_pos_breaks = torch.linspace(
+          conditional_pos_min_bin,
+          conditional_pos_max_bin,
+          steps=conditional_pos_num_bin
+      )
+      self.register_buffer('conditional_pos_breaks', conditional_pos_breaks)
+
     # recycling params
     self.recycling_to_msa_repr = nn.Linear(
         dim_single, dim_msa
@@ -227,7 +243,48 @@ def forward(
     else:
       msa, msa_mask, msa_embed = None, None, None  # msa as features disabled
     del seq_embed, msa_embed
-    recyclables, = map(batch.get, ('recyclables', ))
+    # FIXME: fake recyclables
+    if 'recyclables' not in batch:
+      b, n, device = seq.shape[:-1], seq.shape[-1], seq.device
+      _, dim_msa, dim_pairwise = self.dim  # embedd_dim_get(self.dim)
+      if exists(self.conditional_pos_linear):
+        # assert all(key in batch for key in ('cond_mask', 'coord', 'coord_mask'))
+        if 'cond_mask' in batch:
+          cond_mask = batch['cond_mask']
+        else:
+          cond_mask = torch.zeros(b + (n, ), device=device)
+        cond_mask = cond_mask[..., :, None] * cond_mask[..., None, :]
+
+        if 'coord' in batch:
+          coord = batch['coord']
+        else:
+          coord = torch.zeros(
+              b + (n, residue_constants.atom14_type_num, 3), device=device
+          )
+        if 'coord_mask' in batch:
+          coord_mask = batch['coord_mask']
+        else:
+          coord_mask = torch.zeros(
+              b + (n, residue_constants.atom14_type_num), device=device
+          )
+        pseudo_beta, pseudo_beta_mask = functional.pseudo_beta_fn(
+            seq, coord, coord_mask
+        )
+        dgram = functional.distogram_from_positions(
+            self.conditional_pos_breaks, pseudo_beta
+        )
+        pairwise_repr = self.conditional_pos_linear(dgram) * pseudo_beta_mask[..., None]
+
+        pairwise_repr = pairwise_repr * cond_mask[..., None]
+      else:
+        pairwise_repr = torch.zeros(b + (n, n, dim_pairwise), device=device)
+
+      batch['recyclables'] = Recyclables(
+          msa_first_row_repr=torch.zeros(b + (n, dim_msa), device=device),
+          pairwise_repr=pairwise_repr,
+          coords=torch.zeros(b + (n, residue_constants.atom_type_num, 3), device=device)
+      )
+    recyclables = batch['recyclables']
 
     representations = {'recycling': return_recyclables}
 
@@ -350,6 +407,8 @@ def from_config(config):
         'template_depth',
         'num_tokens',
         'num_msa_tokens',
+        'conditional_pos',
+        'recycling_frames',
         'recycling_single_repr',
         'recycling_pos',
     ):
@@ -367,18 +426,6 @@ def embeddings(self):
   def forward(self, batch, *, num_recycle=0, **kwargs):
     assert num_recycle >= 0
 
-    # variables
-    seq = batch['seq']
-    b, n, device = seq.shape[:-1], seq.shape[-1], seq.device
-    # FIXME: fake recyclables
-    if 'recyclables' not in batch:
-      _, dim_msa, dim_pairwise = self.impl.dim  # embedd_dim_get(self.impl.dim)
-      batch['recyclables'] = Recyclables(
-          msa_first_row_repr=torch.zeros(b + (n, dim_msa), device=device),
-          pairwise_repr=torch.zeros(b + (n, n, dim_pairwise), device=device),
-          coords=torch.zeros(b + (n, residue_constants.atom_type_num, 3), device=device)
-      )
-
     if self.training:
       num_recycle = random.randint(0, num_recycle)
     cycling_function = functools.partial(