feat(utils): add hashing a checkpoint utility

ignite/utils.py

@@ -1,13 +1,24 @@
 import collections.abc as collections
 import functools
+import hashlib
 import logging
 import random
+import shutil
 import warnings
+from pathlib import Path
 from typing import Any, Callable, Dict, Optional, TextIO, Tuple, Type, TypeVar, Union, cast
 
 import torch
 
-__all__ = ["convert_tensor", "apply_to_tensor", "apply_to_type", "to_onehot", "setup_logger", "manual_seed"]
+__all__ = [
+    "convert_tensor",
+    "apply_to_tensor",
+    "apply_to_type",
+    "to_onehot",
+    "setup_logger",
+    "manual_seed",
+    "hash_checkpoint",
+]
 
 
 def convert_tensor(
@@ -272,3 +283,34 @@ def wrapper(*args: Any, **kwargs: Dict[str, Any]) -> Callable:
         return cast(F, wrapper)
 
     return decorator
+
+
+def hash_checkpoint(checkpoint_path: Union[str, Path], output_dir: Union[str, Path],) -> Tuple[Path, str]:
+    """
+    Hash the checkpoint file in the format of ``<filename>-<hash>.<ext>``
+    to be used with ``check_hash`` of :func:`torch.hub.load_state_dict_from_url`.
+
+    Args:
+        checkpoint_path: Path to the checkpoint file.
+        output_dir: Output directory to store the hashed checkpoint file.
+
+    Returns:
+        Path to the hashed checkpoint file, The 8 digits of SHA256 hash.
+
+    .. versionadded:: 0.5.0
+    """
+
+    if isinstance(checkpoint_path, str):
+        checkpoint_path = Path(checkpoint_path)
+
+    if isinstance(output_dir, str):
+        output_dir = Path(output_dir)
+
+    sha_hash = hashlib.sha256(checkpoint_path.read_bytes()).hexdigest()
+    old_filename = checkpoint_path.stem
+    new_filename = "-".join((old_filename, sha_hash[:8])) + ".pt"
+
+    hash_checkpoint_path = output_dir / new_filename
+    shutil.move(str(checkpoint_path), hash_checkpoint_path)
+
+    return hash_checkpoint_path, sha_hash

tests/ignite/test_utils.py

@@ -7,7 +7,7 @@
 import torch
 
 from ignite.engine import Engine, Events
-from ignite.utils import convert_tensor, deprecated, setup_logger, to_onehot
+from ignite.utils import convert_tensor, deprecated, hash_checkpoint, setup_logger, to_onehot
 
 
 def test_convert_tensor():
@@ -242,3 +242,17 @@ def func_with_everything():
 
 def test_smoke__utils():
     from ignite._utils import apply_to_tensor, apply_to_type, convert_tensor, to_onehot  # noqa: F401
+
+
+def test_hash_checkpoint(tmp_path):
+    # download lightweight model
+    from torchvision.models import squeezenet1_0
+
+    model = squeezenet1_0()
+    torch.hub.download_url_to_file(
+        "https://download.pytorch.org/models/squeezenet1_0-b66bff10.pth", f"{tmp_path}/squeezenet1_0.pt",
+    )
+    hash_checkpoint_path, sha_hash = hash_checkpoint(f"{tmp_path}/squeezenet1_0.pt", str(tmp_path))
+    model.load_state_dict(torch.load(hash_checkpoint_path), True)
+    assert sha_hash[:8] == "b66bff10"
+    assert hash_checkpoint_path.name == f"squeezenet1_0-{sha_hash[:8]}.pt"