use best model if val/test/predict after train

Author: cw-tan

nequip/scripts/train.py

@@ -33,6 +33,11 @@ def main(config: DictConfig) -> None:
         runs = [config.run]
     assert all([run_type in ["train", "val", "test", "predict"] for run_type in runs])
 
+    # ensure only single train at most, to protect restart and checkpointing logic later
+    assert (
+        sum([run_type == "train" for run_type in runs]) == 1
+    ), "only a single `train` instance can be present in `run`"
+
     if "train" in runs:
         assert (
             "loss" in config.train
@@ -164,14 +169,17 @@ def training_data_stats(stat_name: str):
     trainer = instantiate(trainer_cfg, inference_mode=False)
 
     # === loop of run types ===
-    ckpt_path = config.get("ckpt_path", None)
+    # restart behavior is such that
+    # - train from ckpt uses the correct ckpt file to restore training state (so it is given a specific `ckpt_path`)
+    # - val/test/predict from ckpt would use the `nequip_module` from the ckpt (and so uses `ckpt_path=None`)
+    # - if we train, then val/test/predict, we set `ckpt_path="best"` after training so val/test/predict tasks after that will use the "best" model
+    ckpt_path = None
     for run_type in runs:
         if run_type == "train":
+            ckpt_path = config.get("ckpt_path", None)
             logger.info("TRAIN RUN START")
             trainer.fit(nequip_module, datamodule=datamodule, ckpt_path=ckpt_path)
-            # `fit` is the only task that changes the model, so we remove the ckpt_path if train ever gets called
-            # this means that the latest model information is used for other tasks (val, test, predict)
-            ckpt_path = None
+            ckpt_path = "best"
             logger.info("TRAIN RUN END")
         elif run_type == "val":
             logger.info("VAL RUN START")