Add optional precision-preserving preprocessing for examples/unconditional_image_generation/train_unconditional.py #12596
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
turian
commented
Nov 5, 2025
| - you can either provide your own folder as `--train_data_dir` | ||
| - or you can upload your dataset to the hub (possibly as a private repo, if you prefer so), and simply pass the `--dataset_name` argument. | ||
|
|
||
| If your dataset contains 16 or 32-bit channels (for example, medical TIFFs), add the `--preserve_input_precision` flag so the preprocessing keeps the original precision while still training a 3-channel model. Precision still depends on the decoder: Pillow keeps 16-bit grayscale and float inputs, but many 16-bit RGB files are decoded as 8-bit RGB, and the flag cannot recover precision lost at load time. |
Contributor
Author
There was a problem hiding this comment.
It occurs to me that there is a possible footgun here for people using 16-bit RGB or RGBA images, but I wanted to keep the diff minimal.
sayakpaul
approved these changes
Nov 6, 2025
Member
sayakpaul
left a comment
sayakpaul
left a comment
There was a problem hiding this comment.
Very cool work! Thank you! If you have some extended reading to add to the README, feel free to!
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
6 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Adds an opt-in
--preserve_input_precisionflag toexamples/unconditional_image_generation/train_unconditional.pyso users can keep 16/32-bit channel data (e.g. medical TIFFs) in full precision during preprocessing while still emitting 3-channel float32 tensors normalized to [-1, 1]. By default nothing changes: we still hitimage.convert("RGB") → ToTensor() → Normalize, preserving byte-for-byte parity with the current pipeline.With the flag enabled, the script now:
transforms.PILToTensor()+ConvertImageDtype(torch.float32)to avoid 8-bit quantization._ensure_three_channelshelper (repeat, pad, or slice as needed).README now documents the flag for users with high-bit-depth datasets.
Fixes # (issue)
Before submitting
Who can review?