Fix orientation batching (#58)

* Fix orientation batching

* Vectorize orientations feature computation

* Comment out debugging code

* Simplify code

* update changelog

---------

Co-authored-by: Arian Jamasb <[email protected]>

Author: amorehead

CHANGELOG.md

@@ -10,6 +10,7 @@
 ### Features
 
 * Improves positional encoding performance by adding a `seq_pos` attribute on `Data/Protein` objects in the base dataset getter. [#53](https://github.com/a-r-j/ProteinWorkshop/pull/53/)
+* Ensure correct batched computation of orientation features. [#58](https://github.com/a-r-j/ProteinWorkshop/pull/58/)
 
 ### Models
 

proteinworkshop/config/visualise.yaml

@@ -2,7 +2,7 @@
 
 # === 1. Set config parameters ===
 name: "" # default name for the experiment, "" means logger (eg. wandb) will generate a unique name
-seed: 52 # seed for random number generators in pytorch, numpy and python.random
+seed: 52 # seed for random number generators in pytorch, numpy and python.random (as well as in UMAP)
 num_workers: 16 # number of subprocesses to use for data loading.
 
 # === 2. Specify defaults here. Defaults will be overwritten by equivalently named options in this file ===
@@ -29,7 +29,6 @@ compile: True
 # simply provide checkpoint path and plot filepath to embed dataset and plot its UMAP embeddings
 ckpt_path: null # path to checkpoint to load
 plot_filepath: null # path to which to save embeddings plot
-seed: 42 # random seed to be used by the UMAP algorithm
 use_cuda_device: True # if True, use an available CUDA device for embedding generation
 cuda_device_index: 0 # if CUDA devices are targeted and available, which available CUDA device to use for embedding generation
 

proteinworkshop/datasets/utils.py

@@ -137,9 +137,9 @@ def create_example_batch(n: int = 4) -> ProteinBatch:
     batch.pos = batch.coords[:, 1, :]
     batch.x = F.one_hot(batch.residue_type, num_classes=23).float()
 
+    batch.x_vector_attr = orientations(batch.pos, batch._slice_dict["coords"])
     batch.graph_y = torch.randint(0, 2, (n, 1))
 
-    batch.x_vector_attr = orientations(batch.pos)
     batch.edge_attr = pos_emb(batch.edge_index, 9)
     batch.edge_vector_attr = _normalize(
         batch.pos[batch.edge_index[0]] - batch.pos[batch.edge_index[1]]

proteinworkshop/features/node_features.py

@@ -103,7 +103,7 @@ def compute_vector_node_features(
     vector_node_features = []
     for feature in vector_features:
         if feature == "orientation":
-            vector_node_features.append(orientations(x.coords))
+            vector_node_features.append(orientations(x.coords, x._slice_dict["coords"]))
         elif feature == "virtual_cb_vector":
             raise NotImplementedError("Virtual CB vector not implemented yet.")
         else:
@@ -149,12 +149,46 @@ def compute_surface_feat(
 
 @jaxtyped(typechecker=typechecker)
 def orientations(
-    X: Union[CoordTensor, AtomTensor], ca_idx: int = 1
+    X: Union[CoordTensor, AtomTensor], coords_slice_index: torch.Tensor, ca_idx: int = 1
 ) -> OrientationTensor:
     if X.ndim == 3:
         X = X[:, ca_idx, :]
-    forward = _normalize(X[1:] - X[:-1])
-    backward = _normalize(X[:-1] - X[1:])
+
+    # NOTE: the first item in the coordinates slice index is always 0,
+    # and the last item is always the node count of the batch
+    batch_num_nodes = X.shape[0]
+    slice_index = coords_slice_index[1:] - 1
+    last_node_index = slice_index[:-1]
+    first_node_index = slice_index[:-1] + 1
+    slice_mask = torch.zeros(batch_num_nodes - 1, dtype=torch.bool)
+    last_node_forward_slice_mask = slice_mask.clone()
+    first_node_backward_slice_mask = slice_mask.clone()
+
+    # NOTE: all of the last (first) nodes in a subgraph have their
+    # forward (backward) vectors set to a padding value (i.e., 0.0)
+    # to mimic feature construction behavior with single input graphs
+    forward_slice = X[1:] - X[:-1]
+    backward_slice = X[:-1] - X[1:]
+    last_node_forward_slice_mask[last_node_index] = True
+    first_node_backward_slice_mask[first_node_index - 1] = True  # NOTE: for the backward slices, our indexing defaults to node index `1`
+    forward_slice[last_node_forward_slice_mask] = 0.0 # NOTE: this handles all but the last node in the last subgraph
+    backward_slice[first_node_backward_slice_mask] = 0.0 # NOTE: this handles all but the first node in the first subgraph
+
+    # NOTE: padding first and last nodes with zero vectors does not impact feature normalization
+    forward = _normalize(forward_slice)
+    backward = _normalize(backward_slice)
     forward = F.pad(forward, [0, 0, 0, 1])
     backward = F.pad(backward, [0, 0, 1, 0])
-    return torch.cat((forward.unsqueeze(-2), backward.unsqueeze(-2)), dim=-2)
+    orientations = torch.cat((forward.unsqueeze(-2), backward.unsqueeze(-2)), dim=-2)
+
+    # optionally debug/verify the orientations
+    # last_node_indices = torch.cat((last_node_index, torch.tensor([batch_num_nodes - 1])), dim=0)
+    # first_node_indices = torch.cat((torch.tensor([0]), first_node_index), dim=0)
+    # intermediate_node_indices_mask = torch.ones(batch_num_nodes, device=X.device, dtype=torch.bool)
+    # intermediate_node_indices_mask[last_node_indices] = False
+    # intermediate_node_indices_mask[first_node_indices] = False
+    # assert not orientations[last_node_indices][:, 0].any() and orientations[last_node_indices][:, 1].any()
+    # assert orientations[first_node_indices][:, 0].any() and not orientations[first_node_indices][:, 1].any()
+    # assert orientations[intermediate_node_indices_mask][:, 0].any() and orientations[intermediate_node_indices_mask][:, 1].any()
+
+    return orientations

proteinworkshop/train.py

@@ -14,7 +14,6 @@
 import torch.nn as nn
 import torch_geometric
 from graphein.protein.tensor.dataloader import ProteinDataLoader
-from graphein.ml.datasets.foldcomp_dataset import FoldCompLightningDataModule
 from lightning.pytorch.callbacks import Callback
 from lightning.pytorch.loggers import Logger
 from loguru import logger as log