Repository navigation
Add MIRepNet model - #1146
Merged
bruAristimunha merged 10 commits intoSep 29, 2026
Merged
Add MIRepNet model#1146
Conversation
bruAristimunha
force-pushed
the
feat/mirepnet-1126
branch
from
August 31, 2026 08:49
c974a12 to
45146ac
Compare
Contributor
There was a problem hiding this comment.
Warning
Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.
Copilot review overview
Review effort: Lite
Findings: None
What changed in this PR
Adds the released MIRepNet convolutional-Transformer encoder and classification head, including pretrained-checkpoint integration and documentation.
Changes:
- Adds the
MIRepNetmodel with validation, feature extraction, head resetting, and checkpoint key mapping. - Extends shared attention with configurable score scaling while retaining PyTorch 2.0 compatibility.
- Adds focused tests, pretrained Hub integration, registry metadata, citation, licensing, and release documentation.
| File | Description |
|---|---|
test/unit_tests/models/test_return_features.py |
Adds MIRepNet to unified feature-return tests. |
test/unit_tests/models/test_modules.py |
Tests explicit attention scaling and SDPA compatibility. |
test/unit_tests/models/test_mirepnet.py |
Adds MIRepNet parameter-validation tests. |
test/unit_tests/models/test_integration.py |
Excludes MIRepNet from TorchScript tests due to its polymorphic output. |
test/integration_tests/test_pretrained_hub_models.py |
Adds pretrained MIRepNet Hub loading and forward-pass coverage. |
docs/whats_new.rst |
Documents the new model and pretrained weights. |
docs/references.bib |
Adds the MIRepNet publication citation. |
docs/api.rst |
Lists MIRepNet in the model API documentation. |
braindecode/modules/attention.py |
Adds configurable attention-score scaling. |
braindecode/models/util.py |
Registers MIRepNet constructor metadata. |
braindecode/models/summary.csv |
Adds MIRepNet model summary metadata. |
braindecode/models/mirepnet.py |
Implements the MIRepNet architecture and model utilities. |
braindecode/models/__init__.py |
Exports MIRepNet publicly. |
NOTICE.txt |
Records the MIT-licensed MIRepNet implementation. |
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## master #1146 +/- ##
==========================================
+ Coverage 86.57% 87.05% +0.48%
==========================================
Files 142 145 +3
Lines 16303 16655 +352
==========================================
+ Hits 14114 14499 +385
+ Misses 2189 2156 -33 🚀 New features to boost your workflow:
|
bruAristimunha
added a commit
to qinxwew/braindecode
that referenced
this pull request
Sep 29, 2026
Resolve models/__init__.py and docs/whats_new.rst conflicts against master's newly-merged MIRepNet (braindecode#1146) and BaRISTA (braindecode#1173): keep both registry entries, restore alphabetical import/__all__ order, keep both changelog entries.
bruAristimunha
added a commit
to bruAristimunha/braindecode
that referenced
this pull request
Oct 6, 2026
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.
Summary
Scope
This intentionally excludes InterpolatedMIRepNet, channel interpolation, preprocessing, Euclidean alignment, and the masked reconstruction decoder. The raw checkpoint contains 143 keys; all 109 inference keys used by this downstream implementation load, while decoder, mask-token, and forward-unused channel-embedding keys remain excluded.
Testing
python -m pytest -q: 3,391 passed, 232 skippedCloses #1126
Replication results — original-author implementation
Scope: these real-data measurements use the original-author MIRepNet implementation and released checkpoint, not the Braindecode port. They support provenance, not an end-to-end benchmark of this PR. The paper's 82.36 ± 0.10% was not matched.
Each seed mean is an unweighted mean over all nine subjects; SD is across the three seed means. All 27 subject/seed runs per filter used the last epoch of ten, sequential subject RNG (seeded once per full nine-subject run), unstratified 30/70 splits and the same 1,400 exact trial identities from the first test session. Test sets contain 112 trials per subject, except subject 1 with 84; these are not pooled-trial averages. The control changed only the upper band edge and moved the mean by +0.242504 pp; deficits from the paper remain 2.046949 pp and 1.804444 pp, respectively. No seed/filter sweep or best-test-epoch selection.
Protocol caveats: Euclidean alignment is applied separately to native-channel train/test EEG before interpolation to 45 channels. This is transductive: unlabeled test EEG supplies its own covariance reference; test labels are not used for training. The CPU environment and seeds 666/667/668 are reconstructed, while the original prepared arrays, historical seeds and complete environment are unavailable. Neither filter is proven to be the historical setting; this is practical code-faithful reconstruction, not exact historical replication. Original fine-tuning loaded 108/110 downstream tensors with a new two-class head; that is distinct from the 109 inference tensors and port-parity checks described above.
Full protocol, limitations, source/checkpoint hashes and both per-subject tables · All 54 observations with integer correct counts (CSV).
New verification for this documentation-only update: independently recomputed all 54 accuracies, seed means and both SD definitions; verified ten epochs, sequential RNG chains, identical trial/split metadata between filters and all nine recorded terminal job/container receipts (five successful jobs, four diagnosed prerequisite failures). Both complete evaluations exited 0. Repository pre-commit checks for the two added evidence files and
git diff --checkpassed. No model code changed; the earlier unit-test claims in Testing are preserved historical results, not a newly rerun full suite. No CI or merge outcome is claimed.