feat: add runs n' poses val set (#210)

* feat: add runs n' poses val set

* chore: ruff

* chore: remove from defaults

configs/datasets/val/runs_and_poses.yaml

@@ -0,0 +1,11 @@
+defaults:
+  - base
+
+dataset:
+  dataset_parser:
+    _target_: datahub.datasets.parsers.ValidationDFParserLikeAF3
+  dataset:
+    _target_: datahub.datasets.datasets.PandasDataset
+    data: ${paths.data.pdb_data_dir}/runs_n_poses_entry_level_df.parquet
+    filters:
+      - "n_tokens_total < 1000" # Subset to reasonably-sized examples for efficiency

configs/experiment/pretrained/af3.yaml

@@ -12,4 +12,7 @@ tags:
 
 project: af3
 
-ckpt_path: /net/software/containers/versions/modelhub_inference/ckpts/modelhub_af3_with_confidence_latest.ckpt
+ckpt_config:
+  _target_: modelhub.utils.weights.CheckpointConfig
+  path: /net/software/containers/versions/modelhub_inference/ckpts/modelhub_af3_with_confidence_latest.ckpt
+  reset_optimizer: true

configs/validate.yaml

@@ -42,8 +42,4 @@ callbacks:
   dump_validation_structures_callback:
     dump_predictions: True
     one_model_per_file: False
-    dump_trajectories: False
-
-# passing checkpoint path required for validation
-# DO NOT set here; set in the experiment config file
-ckpt_path: ???
+    dump_trajectories: False

src/modelhub/validate.py

@@ -118,12 +118,20 @@ def validate(cfg: DictConfig) -> None:
         loader_cfg=cfg.dataloader["val"],
     )
 
+    # ... load the checkpoint configuration, regardless of whether it's a path or a config
+    if "ckpt_path" in cfg and cfg.ckpt_path:
+        ckpt_path = cfg.ckpt_path
+    elif "ckpt_config" in cfg and cfg.ckpt_config:
+        assert (
+            "path" in cfg.ckpt_config
+        ), "No checkpoint path provided in `ckpt_config`!"
+        ckpt_path = cfg.ckpt_config.path
+
     # ... validate the model
     ranked_logger.info("Validating model...")
-    assert cfg.ckpt_path is not None, "No checkpoint path provided for validation!"
     trainer.validate(
         val_loaders=val_loaders,
-        ckpt_path=cfg.ckpt_path,
+        ckpt_path=ckpt_path,
     )
 
     ranked_logger.info("Validation complete!")