Merge pull request #280 from kAIto47802/pytorch-checkpoint-artifact

Introduce `optuna.artifacts` to the PyTorch checkpoint example

pytorch/pytorch_checkpoint.py

@@ -14,9 +14,11 @@
 """
 
 import os
-import shutil
 
 import optuna
+from optuna.artifacts import download_artifact
+from optuna.artifacts import FileSystemArtifactStore
+from optuna.artifacts import upload_artifact
 from optuna.storages import RetryFailedTrialCallback
 import torch
 import torch.nn as nn
@@ -37,6 +39,10 @@
 N_VALID_EXAMPLES = BATCHSIZE * 10
 CHECKPOINT_DIR = "pytorch_checkpoint"
 
+base_path = "./artifacts"
+os.makedirs(base_path, exist_ok=True)
+artifact_store = FileSystemArtifactStore(base_path=base_path)
+
 
 def define_model(trial):
     # We optimize the number of layers, hidden units and dropout ratio in each layer.
@@ -83,36 +89,36 @@ def objective(trial):
     lr = trial.suggest_float("lr", 1e-5, 1e-1, log=True)
     optimizer = getattr(optim, optimizer_name)(model.parameters(), lr=lr)
 
-    trial_number = RetryFailedTrialCallback.retried_trial_number(trial)
-    trial_checkpoint_dir = os.path.join(CHECKPOINT_DIR, str(trial_number))
-    checkpoint_path = os.path.join(trial_checkpoint_dir, "model.pt")
-    checkpoint_exists = os.path.isfile(checkpoint_path)
-
-    if trial_number is not None and checkpoint_exists:
-        checkpoint = torch.load(checkpoint_path)
+    artifact_id = None
+    retry_history = RetryFailedTrialCallback.retry_history(trial)
+    for trial_number in reversed(retry_history):
+        artifact_id = trial.study.trials[trial_number].user_attrs.get("artifact_id")
+        if artifact_id is not None:
+            retry_trial_number = trial_number
+            break
+
+    if artifact_id is not None:
+        download_artifact(
+            artifact_store=artifact_store,
+            file_path=f"./tmp_model_{trial.number}.pt",
+            artifact_id=artifact_id,
+        )
+        checkpoint = torch.load(f"./tmp_model_{trial.number}.pt")
+        os.remove(f"./tmp_model_{trial.number}.pt")
         epoch = checkpoint["epoch"]
         epoch_begin = epoch + 1
 
-        print(f"Loading a checkpoint from trial {trial_number} in epoch {epoch}.")
+        print(f"Loading a checkpoint from trial {retry_trial_number} in epoch {epoch}.")
 
         model.load_state_dict(checkpoint["model_state_dict"])
         optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
         accuracy = checkpoint["accuracy"]
     else:
-        trial_checkpoint_dir = os.path.join(CHECKPOINT_DIR, str(trial.number))
-        checkpoint_path = os.path.join(trial_checkpoint_dir, "model.pt")
         epoch_begin = 0
 
     # Get the FashionMNIST dataset.
     train_loader, valid_loader = get_mnist()
 
-    os.makedirs(trial_checkpoint_dir, exist_ok=True)
-    # A checkpoint may be corrupted when the process is killed during `torch.save`.
-    # Reduce the risk by first calling `torch.save` to a temporary file, then copy.
-    tmp_checkpoint_path = os.path.join(trial_checkpoint_dir, "tmp_model.pt")
-
-    print(f"Checkpoint path for trial is '{checkpoint_path}'.")
-
     # Training of the model.
     for epoch in range(epoch_begin, EPOCHS):
         model.train()
@@ -159,9 +165,15 @@ def objective(trial):
                 "optimizer_state_dict": optimizer.state_dict(),
                 "accuracy": accuracy,
             },
-            tmp_checkpoint_path,
+            f"./tmp_model_{trial.number}.pt",
+        )
+        artifact_id = upload_artifact(
+            artifact_store=artifact_store,
+            file_path=f"./tmp_model_{trial.number}.pt",
+            study_or_trial=trial,
         )
-        shutil.move(tmp_checkpoint_path, checkpoint_path)
+        trial.set_user_attr("artifact_id", artifact_id)
+        os.remove(f"./tmp_model_{trial.number}.pt")
 
         # Handle pruning based on the intermediate value.
         if trial.should_prune():