Introduction

On Windows, TensorFlow checkpoint restore errors that say "Access is denied" are usually file-system problems rather than model problems. The most common causes are an unwritable directory, a locked checkpoint file, an antivirus or sync tool interfering with the files, or a path issue that makes TensorFlow touch the wrong location.

Check that the checkpoint path is real and writable

The first thing to verify is the directory you are restoring from. Use an explicit path and confirm that Python can list the files there:

1from pathlib import Path
2
3ckpt_dir = Path(r"C:\models\run1")
4print(ckpt_dir.exists())
5print(list(ckpt_dir.glob("*")))

Raw strings such as r"C:\models\run1" help avoid backslash-escape mistakes on Windows.

If the directory is inside a protected location such as Program Files or a restricted user folder, move the checkpoints to a normal writable working directory.

Restore from the checkpoint prefix, not a random file

TensorFlow checkpoints are not one single file. Depending on the API, you typically restore using the checkpoint prefix rather than the .index file directly.

For example:

1import tensorflow as tf
2
3model = tf.keras.Sequential([tf.keras.layers.Dense(4)])
4checkpoint = tf.train.Checkpoint(model=model)
5
6checkpoint.restore(r"C:\models\run1\ckpt-5").expect_partial()

If you point TensorFlow at the wrong file name or partial path, the resulting error messages can be misleading and look like general file-access problems.

File locks and external tools are common on Windows

Windows is more likely than Unix-like systems to produce "access denied" behavior when another process has a file open. Common culprits include:

• antivirus scanners
• cloud-sync clients such as OneDrive
• editors or tooling that index files aggressively
• another training process still writing the checkpoint

If the checkpoint directory is actively syncing or scanning, try moving it to a simpler local path and rerun the restore.

Create and test the directory explicitly

If the workflow also saves checkpoints, verify write access with a tiny test:

1from pathlib import Path
2
3ckpt_dir = Path(r"C:\temp\tf-checkpoints")
4ckpt_dir.mkdir(parents=True, exist_ok=True)
5
6test_file = ckpt_dir / "write_test.txt"
7test_file.write_text("ok", encoding="utf-8")
8print(test_file.read_text(encoding="utf-8"))

If this fails, the problem is Windows file access, not TensorFlow restore logic.

Use shorter, simpler paths when debugging

Long paths, unusual characters, or nested synced directories can complicate Windows checkpoint access. When debugging, prefer something like:

• 'C:\temp\tf-checkpoints'
• 'C:\models\experiment1'

That removes a lot of environmental ambiguity.

Avoid restoring from protected Windows folders

Directories under locations such as Program Files or heavily managed corporate folders can introduce permission surprises. For debugging, move checkpoints to a normal user-writable folder before chasing deeper TensorFlow explanations.

Checkpoint restore depends on matching model objects too

Not every restore failure is purely a permissions issue. If the restore path is correct but the model or checkpoint objects do not line up, TensorFlow may surface confusing errors. Still, when the message explicitly says "Access is denied," start with file access and path hygiene first.

Common Pitfalls

• Using a relative or incorrectly escaped Windows path.
• Pointing restore at the wrong checkpoint file instead of the checkpoint prefix.
• Storing checkpoints in protected, synced, or aggressively scanned directories.
• Trying to restore while another process is still writing the checkpoint files.
• Assuming every restore failure is a TensorFlow graph problem instead of a file-system problem.

Summary

• On Windows, "Access is denied" during checkpoint restore is usually a file or path issue.
• Use explicit raw-string paths and verify the directory is readable and writable.
• Restore from the correct checkpoint prefix, not an arbitrary checkpoint component file.
• Watch for file locks from antivirus, sync tools, or concurrent training processes.
• Debug with a short local path before investigating deeper model-level issues.