Apply the same augmentation chain to all camera images#970
Open
0xadvait wants to merge 1 commit into
Open
Conversation
augmax.Chain splits its rng once per sub-transform, so with 4 transforms on the base camera and 1 on wrist cameras, ColorJitter drew different parameters for base vs wrist views of the same frame even though the same rng is passed for every camera. The pi0.5 paper (Appendix E) applies the full crop/resize/rotate/jitter chain to all input images, so this change applies the full chain to every camera, which also makes augmentation parameters consistent across cameras within a frame. Adds a regression test for cross-camera consistency. Fixes Physical-Intelligence#859
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Problem
preprocess_observationpasses the samerngfor every camera, which suggests augmentation parameters are meant to be consistent across cameras within a frame. They are not:augmax.Chainsplits its key once per sub-transform, and the base camera chain has 4 transforms while wrist chains have 1, soColorJitterdraws different subkeys for base vs wrist. The same physical frame gets visibly different hue/brightness/contrast on the base camera vs the wrist cameras (#859 has a visual repro).Why this divergence matters
The pi0.5 paper (arXiv 2504.16054, Appendix E) describes the training recipe as applying the full chain to every input image:
followed by exactly the
RandomCrop(0.95) -> Resize -> Rotate(+-5) -> ColorJitter(0.3, 0.4, 0.5)chain this file uses. The wrist special case deviates from the recipe used to train the released checkpoints, and as a side effect breaks cross-camera color consistency through the Chain key-splitting described above.Fix
Apply the full augmentation chain to every camera image. With identical chains and the shared per-frame rng, all cameras receive identical augmentation parameters for a given frame, restoring cross-camera color consistency and matching the published recipe.
If the wrist exclusion from geometric augmentation was intentional, the minimal alternative is to keep the wrist chain geometric-free but give
ColorJittera dedicated key shared across cameras. Happy to switch this PR to that variant.Test
Adds
test_preprocess_observation_train_augmentations_consistent_across_cameras: identical images on all three cameras, fixed key, asserts all augmented outputs are bitwise identical across cameras, actually differ from the input, stay within [-1, 1], and thattrain=Falsepasses images through unchanged. Runs on CPU.Fixes #859