Add random block wise mask from Pixio#1919
Open
gabrielfruet wants to merge 6 commits into
Open
Conversation
…n masking Based on Pixio's MAE implementation. Masks patches in coherent blocks rather than individually for spatial coherence in image patch masking.
Tests cover: - Basic functionality and index coverage - Class token preservation - Block alignment verification - Various mask ratios and block sizes - Device handling (CPU/CUDA) - Error cases (non-square, invalid block size) - Determinism and batch consistency
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #1919 +/- ##
==========================================
+ Coverage 86.19% 86.25% +0.06%
==========================================
Files 169 169
Lines 7071 7110 +39
==========================================
+ Hits 6095 6133 +38
- Misses 976 977 +1 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
Contributor
There was a problem hiding this comment.
Pull request overview
Adds a new token-index masking utility to support Pixio-style block-wise masking (masking contiguous patch blocks instead of individual tokens), along with unit tests to validate expected behavior and prefix-token handling (CLS/REG tokens).
Changes:
- Introduce
random_block_wise_mask(...)inlightly/models/utils.pyto generate keep/mask indices using block granularity andnum_prefix_tokens. - Add a dedicated test suite in
tests/models/test_ModelUtils.pycovering prefix tokens, block sizes, mask ratios, determinism, and device placement.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 5 comments.
| File | Description |
|---|---|
lightly/models/utils.py |
Adds the new block-wise masking function returning (idx_keep, idx_mask) indices. |
tests/models/test_ModelUtils.py |
Adds unit tests for random_block_wise_mask behavior, error cases, determinism, and device handling. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
|
|
||
| if num_patches <= 0: | ||
| raise ValueError( | ||
| "Sequence length must be greater than prefix tokens (1 if not masking class token)." |
Comment on lines
+689
to
+691
| if num_prefix_tokens > 0: | ||
| noise[:, 0] = -1 | ||
|
|
| .expand(batch_size, -1) | ||
| ) | ||
| idx_keep = torch.cat([prefix_indices, patch_indices], dim=1) | ||
| idx_keep = torch.sort(torch.unique(idx_keep, dim=1), dim=1)[0] |
Comment on lines
+1183
to
+1184
| assert idx_keep.shape[1] == idx_keep[0].shape[0] | ||
| assert idx_mask.shape[1] == idx_mask[0].shape[0] |
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Open
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.
#1894
This is part of a series of PR's to implement the Pixio paper.
The only difference is the
num_prefix_tokensvsmask_class_token. Since Pixio relies on register tokens, themask_class_tokenwas not suitable for choosing whether to mask or not the REG/CLS tokens.