🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/tutorials/3866

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 59c2dc4 with merge base 5efb99a ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@claude

Claude finished @svekars's task in 1m 26s —— View job

Review of PR #3866

• Gather context and read changed files
• Review changes against origin/main
• Post review feedback

Overall this PR correctly migrates the two tutorial files to the torchvision v2 transforms API and makes several good modernization changes. A few items to address:

Issues

1. Stale prose reference to v1 API — beginner_source/introyt/trainingyt.py:60 still reads:

   We use ``torchvision.transforms.Normalize()`` to
   zero-center and normalize the distribution of the image tile content
   

   This should be updated to reference torchvision.transforms.v2 (or just v2.Normalize) to match the new code.

   Fix this →

2. Missing torch.no_grad() in validation loop — In tensorboardyt_tutorial.py:215-221, the validation loop switches to net.eval() but does not wrap inference in torch.no_grad(). The companion file trainingyt.py:302 correctly uses with torch.no_grad():. This is a best practice that should be consistent across tutorials teaching the same pattern.

   Fix this →

3. Inconsistent formatting in v2.Compose — In tensorboardyt_tutorial.py:79-82, the list items inside v2.Compose use the old indentation style (bracket on same line, items oddly aligned). In trainingyt.py:75-79, it's formatted more cleanly with items on their own lines and a trailing closing bracket. These two tutorials are closely related and should be consistent.

Minor / Nits

• tensorboardyt_tutorial.py:210: f'Batch {i+1}' — the spacing around + is inconsistent with trainingyt.py:255 which uses f' batch {i + 1} loss: {last_loss}'. Consistent spacing around operators is preferred (i.e., i + 1).

What looks good

• v2 transforms migration (ToImage() + ToDtype(torch.float32, scale=True) replacing ToTensor()) is correct.
• net.eval() / net.train() replacing net.train(False) / net.train(True) is the idiomatic modern approach.
• Removal of the stale TensorFlow/TensorBoard compatibility comment block is appropriate.
• super().__init__() replacing super(Net, self).__init__() is correct Python 3 style.
• f-string modernization is clean and consistent.

@saurabhkthakur good job on this, just a minor fix:

Stale prose reference to v1 API — beginner_source/introyt/trainingyt.py:60 still reads:

We use torchvision.transforms.Normalize() to
zero-center and normalize the distribution of the image tile content
This should be updated to reference torchvision.transforms.v2 (or just v2.Normalize) to match the new code.