Fix loss bugs

config.py

@@ -211,7 +211,7 @@ config = mlc.ConfigDict({
             "min_resolution": 0.1,
             "max_resolution": 3.0,
             "cutoff": 15.,
-            "num_bins": 50,
+            "no_bins": 50,
             "eps": 1e-10,
             "weight": 0.01,
         },

openfold/model/heads.py

@@ -49,7 +49,7 @@ class AuxiliaryHeads(nn.Module):
 
     def forward(self, outputs):
         aux_out = {}
-        lddt_logits = self.plddt(outputs["single"])
+        lddt_logits = self.plddt(outputs["sm"]["single"])
         aux_out["lddt_logits"] = lddt_logits
 
         # Required for relaxation later on

openfold/model/structure_module.py

@@ -751,7 +751,7 @@ class StructureModule(nn.Module):
                 t = t.stop_rot_gradient()
 
         outputs = stack_tensor_dicts(outputs)
-        outputs["single_act"] = s
+        outputs["single"] = s
 
         return outputs
 

openfold/utils/feats.py

@@ -74,6 +74,7 @@ def get_chi_atom_indices():
 
 
 def compute_residx(batch):
+    float_type = batch["seq_mask"].dtype
     aatype = batch["aatype"]
 
     restype_atom14_to_atom37 = []  # mapping (restype, atom37) --> atom14
@@ -104,34 +105,28 @@ def compute_residx(batch):
     restype_atom37_to_atom14.append([0] * 37)
     restype_atom14_mask.append([0.] * 14)
 
-    restype_atom14_to_atom37 = np.array(restype_atom14_to_atom37, dtype=np.int32)
-    restype_atom37_to_atom14 = np.array(restype_atom37_to_atom14, dtype=np.int32)
-    restype_atom14_mask = np.array(restype_atom14_mask, dtype=np.float32)
- 
-    residx_atom14_to_atom37 = np.take_along_axis(
-        restype_atom14_to_atom37,
-        aatype[..., None],
-        axis=0
+    restype_atom14_to_atom37 = aatype.new_tensor(
+        restype_atom14_to_atom37
+    )
+    restype_atom37_to_atom14 = aatype.new_tensor(
+        restype_atom37_to_atom14
     )
-    residx_atom14_mask = np.take_along_axis(
-        restype_atom14_mask,
-        aatype[..., None],
-        axis=0,
+    restype_atom14_mask = aatype.new_tensor(
+        restype_atom14_mask, dtype=float_type
     )
  
+    residx_atom14_to_atom37 = restype_atom14_to_atom37[aatype]
+    residx_atom14_mask = restype_atom14_mask[aatype]
+
     batch['atom14_atom_exists'] = residx_atom14_mask
-    batch['residx_atom14_to_atom37'] = residx_atom14_to_atom37.long()
+    batch['residx_atom14_to_atom37'] = residx_atom14_to_atom37
 
     # create the gather indices for mapping back
-    residx_atom37_to_atom14 = np.take_along_axis(
-        restype_atom37_to_atom14,
-        aatype[..., None],
-        axis=0,
-    )
-    batch['residx_atom37_to_atom14'] = residx_atom37_to_atom14.long()
+    residx_atom37_to_atom14 = restype_atom37_to_atom14[aatype]
+    batch['residx_atom37_to_atom14'] = residx_atom37_to_atom14
 
     # create the corresponding mask
-    restype_atom37_mask = np.zeros([21, 37], dtype=np.float32)
+    restype_atom37_mask = torch.zeros([21, 37], dtype=float_type)
     for restype, restype_letter in enumerate(residue_constants.restypes):
       restype_name = residue_constants.restype_1to3[restype_letter]
       atom_names = residue_constants.residue_atoms[restype_name]
@@ -139,11 +134,7 @@ def compute_residx(batch):
         atom_type = residue_constants.atom_order[atom_name]
         restype_atom37_mask[restype, atom_type] = 1
 
-    residx_atom37_mask = np.take_along_axis(
-        restype_atom37_mask,
-        aatype[..., None],
-        axis=0,
-    )
+    residx_atom37_mask = restype_atom37_mask[aatype]
     batch['atom37_atom_exists'] = residx_atom37_mask
 
 

openfold/utils/loss.py

@@ -27,6 +27,7 @@ from openfold.utils.tensor_utils import (
     tree_map, 
     tensor_tree_map, 
     masked_mean,
+    permute_final_dims,
 )
 
 
@@ -289,28 +290,25 @@ def lddt_loss(
     all_atom_mask: torch.Tensor,
     resolution: torch.Tensor,
     cutoff: float = 15.,
-    num_bins: int = 50,
+    no_bins: int = 50,
     min_resolution: float = 0.1,
     max_resolution: float = 3.0,
     eps: float = 1e-10,
     **kwargs,
 ) -> torch.Tensor:
-    all_atom_positions = batch["all_atom_positions"]
-    all_atom_mask = batch["all_atom_mask"]
-
-    n = all_atom_mask.shape[-1]
+    n = all_atom_mask.shape[-2]
    
     ca_pos = residue_constants.atom_order["CA"]
-    all_atom_pred_pos = all_atom_pred_pos[..., :, ca_pos, :]
-    all_atom_positions = all_atom_positions[..., :, ca_pos, :]
-    all_atom_mask = all_atom_mask[..., :, ca_pos:(ca_pos + 1)] # keep dim
+    all_atom_pred_pos = all_atom_pred_pos[..., ca_pos, :]
+    all_atom_positions = all_atom_positions[..., ca_pos, :]
+    all_atom_mask = all_atom_mask[..., ca_pos:(ca_pos + 1)] # keep dim
 
     dmat_true = torch.sqrt(
         eps +
         torch.sum(
             (
-                all_atom_positions[..., None] - 
-                all_atom_positions[..., None, :]
+                all_atom_positions[..., None, :] - 
+                all_atom_positions[..., None, :, :]
             )**2,
             dim=-1,
         )
@@ -320,14 +318,13 @@ def lddt_loss(
         eps +
         torch.sum(
             (
-                all_atom_pred_pos[..., None] - 
-                all_atom_pred_pos[..., None, :]
+                all_atom_pred_pos[..., None, :] - 
+                all_atom_pred_pos[..., None, :, :]
 
             )**2,
             dim=-1,
         )
     )
-
     dists_to_score = (
         (dmat_true < cutoff) * all_atom_mask *
         permute_final_dims(all_atom_mask, 1, 0) *
@@ -337,28 +334,30 @@ def lddt_loss(
     dist_l1 = torch.abs(dmat_true - dmat_pred)
 
     score = (
-        (dist_l1 < 0.5) + 
-        (dist_l1 < 1.0) +
-        (dist_l1 < 2.0) +
-        (dist_l1 < 4.0)
+        (dist_l1 < 0.5).type(dist_l1.dtype) + 
+        (dist_l1 < 1.0).type(dist_l1.dtype) +
+        (dist_l1 < 2.0).type(dist_l1.dtype) +
+        (dist_l1 < 4.0).type(dist_l1.dtype)
     )
     score *= 0.25
 
     norm = 1. / (eps + torch.sum(dists_to_score, dim=-1))
     score = norm * (eps + torch.sum(dists_to_score * score, dim=-1))
 
-    # TODO: this feels a bit weird, but it's in the source
     score = score.detach() 
   
-    bin_index = torch.floor(lddt_ca * num_bins)
-
+    bin_index = torch.floor(score * no_bins).long()
+    bin_index = torch.clamp(bin_index, max=(no_bins - 1))
     lddt_ca_one_hot = torch.nn.functional.one_hot(
-        bin_index, num_classes=num_bins
+        bin_index, num_classes=no_bins
     )
 
     errors = softmax_cross_entropy(logits, lddt_ca_one_hot)
 
-    loss = torch.sum(errors * all_atom_mask) / (torch.sum(mask_ca) + eps)
+    all_atom_mask = all_atom_mask.squeeze(-1)
+    loss = (
+        torch.sum(errors * all_atom_mask) / (torch.sum(all_atom_mask) + 1e-8)
+    )
    
     loss *= (
         (resolution >= min_resolution) &
@@ -917,17 +916,16 @@ def find_structural_violations(
         overlap_tolerance=clash_overlap_tolerance,
         bond_length_tolerance_factor=violation_tolerance_factor
     )
-    atom14_dists_lower_bound = restype_atom14_bounds["lower_bound"][
+    atom14_atom_exists = batch["atom14_atom_exists"]
+    atom14_dists_lower_bound = (
+        atom14_pred_positions.new_tensor(restype_atom14_bounds["lower_bound"])[
             batch["aatype"]
         ]
-    atom14_dists_upper_bound = restype_atom14_bounds["upper_bound"][
+    )
+    atom14_dists_upper_bound = (
+        atom14_pred_positions.new_tensor(restype_atom14_bounds["upper_bound"])[
             batch["aatype"]
         ]
-    atom14_dists_lower_bound = atom14_pred_positions.new_tensor(
-        atom14_dists_lower_bound
-    )
-    atom14_dists_upper_bound = atom14_pred_positions.new_tensor(
-        atom14_dists_upper_bound
     )
     residue_violations = within_residue_violations(
         atom14_pred_positions=atom14_pred_positions,
@@ -1102,6 +1100,7 @@ def violation_loss(
     violations: Dict[str, torch.Tensor],
     atom14_atom_exists: torch.Tensor,
     eps=1e-6,
+    **kwargs,
 ) -> torch.Tensor:
     num_atoms = torch.sum(atom14_atom_exists)
     l_clash = torch.sum(

openfold/utils/tensor_utils.py

@@ -49,7 +49,7 @@ def dict_multimap(fn, dicts):
     for k, v in first.items():
         all_v = [d[k] for d in dicts]
         if(type(v) is dict):
-            new_dict[k] = dict_multimap(all_v)
+            new_dict[k] = dict_multimap(fn, all_v)
         else:
             new_dict[k] = fn(all_v)
     
@@ -122,9 +122,9 @@ def chunk_layer(layer, inputs, chunk_size, no_batch_dims):
     """
         Implements the "chunking" procedure described in section 1.11.8.
 
-        Layer outputs and inputs are interpreted as simplified "pytrees," 
-        consisting only of (nested) lists, tuples, and dicts with tensor
-        leaves.
+        Layer outputs and inputs are assumed to be simple "pytrees," 
+        consisting only of (arbitrarily nested) lists, tuples, and dicts with 
+        torch.Tensor leaves.
 
         Args:
             layer: