Repository navigation
Conversation
Document the existing eight-bit image path for GPU training and cover its value range for multi-frame and mixed-depth windows. Configure the transform when constructing the dataset, preserving checkpoint identity. Co-Authored-By: Codex <noreply@anthropic.com> AI-Model: GPT-6 Co-Authored-By: Codex <noreply@ai-tool.com> AI-Contributed/Feature: 23/23 AI-Contributed/UT: 13/13
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Contiguous window training examples currently convert eight-bit images to float32 in DataLoader workers. This documents the existing
images_to_tensor(..., return_uint8=True)path and GPU-side normalization so image tensors use one byte per pixel through worker collation and host-to-device transfer. Tests cover multi-frame value equivalence and mixed bit depth, where the uint8 path must not silently normalize high-bit-depth frames.The change is deliberately limited to the documented training configuration; it does not introduce a new sampler or decoder backend. Real-image A10 measurements show the configuration improves Paimon’s own small CNN training proxy, but the same optimization applies to HDF5, so it is not evidence of a Paimon format advantage or the 30% end-to-end target.
Validation: the affected PyPaimon image/window tests passed before the final documentation wording adjustment;
git diff --checkpasses after rebasing onto current master. The final wording change only moves transform configuration into Dataset construction.