makes more sense the other way around, and also make sure valid and test pdbs can be converted to atom inputs on the fly

Author: lucidrains

alphafold3_pytorch/configs.py

@@ -224,7 +224,7 @@ def create_instance(
             convert_pdb_to_atom = dataset_config.convert_pdb_to_atom
 
             if convert_pdb_to_atom:
-                assert dataset_type == 'atom', 'must be `atom` dataset_type if `convert_pdb_to_atom` is set to True'
+                assert dataset_type == 'pdb', 'must be `pdb` dataset_type if `convert_pdb_to_atom` is set to True'
 
             if dataset_type == 'pdb':
                 dataset_klass = PDBDataset
@@ -238,21 +238,29 @@ def create_instance(
             if exists(train_folder):
                 assert 'dataset' not in trainer_kwargs
 
+                dataset = dataset_klass(train_folder, **dataset_kwargs)
+
                 if convert_pdb_to_atom:
-                    pdb_dataset = PDBDataset(train_folder, **dataset_kwargs)
-                    train_folder = pdb_dataset_to_atom_inputs(pdb_dataset)
+                    dataset = pdb_dataset_to_atom_inputs(dataset, return_atom_dataset = True)
 
-                dataset = dataset_klass(train_folder, **dataset_kwargs)
                 trainer_kwargs.update(dataset = dataset)
 
             if exists(valid_folder):
                 assert 'valid_dataset' not in trainer_kwargs
                 dataset = dataset_klass(valid_folder, **dataset_kwargs)
+
+                if convert_pdb_to_atom:
+                    dataset = pdb_dataset_to_atom_inputs(dataset, return_atom_dataset = True)
+
                 trainer_kwargs.update(valid_dataset = dataset)
 
             if exists(test_folder):
                 assert 'test_dataset' not in trainer_kwargs
                 dataset = dataset_klass(test_folder, **dataset_kwargs)
+
+                if convert_pdb_to_atom:
+                    dataset = pdb_dataset_to_atom_inputs(dataset, return_atom_dataset = True)
+
                 trainer_kwargs.update(test_dataset = dataset)
 
             # handle weighted pdb sampling

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "alphafold3-pytorch"
-version = "0.2.44"
+version = "0.2.45"
 description = "Alphafold 3 - Pytorch"
 authors = [
     { name = "Phil Wang", email = "[email protected]" }

tests/configs/trainer_with_atom_dataset_created_from_pdb.yaml

@@ -56,6 +56,6 @@ checkpoint_every: 1
 checkpoint_folder: ./checkpoints
 overwrite_checkpoints: false
 dataset_config:
-    dataset_type: atom
+    dataset_type: pdb
     convert_pdb_to_atom: true
     train_folder: ./test-folder/data/train