Spaces:
Running on Zero
Running on Zero
Commit History
-
Refresh greedy-decode log-prob baseline for torch 2.8.0
The reproducibility baseline in tests/expected_results.json was generated against torch 2.4.1+cu124. After bumping to torch 2.8.0 (cu126) for ZeroGPU compatibility, the predicted peptides match but the log-probabilities differ in the low-order floats from underlying kernel changes. Regenerated via: pytest tests/test_reproducibility.py::TestExpectedResults::test_generate_expected_results -s Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>303b44b -
Update tests
a9d662c -
Update InstaNovo to v1.2.2 with reproducibility fixes
- Update instanovo dependency from >=1.1.4 to ==1.2.2 - Add set_seed() function for deterministic behavior across environments - Set random seeds for Python, NumPy, and PyTorch - Enable torch.backends.cudnn.deterministic for consistent CUDA results - Reset seed before each prediction call - Add comprehensive reproducibility tests This addresses the variability issue between local and HF Spaces runs reported in: https://github.com/instadeepai/InstaNovo/issues/92d66a489
-
Adapt to InstaNovo v1.2.2 API
- Switch imports to TransformerDataProcessor / DiffusionDataProcessor, ResidueSet, Metrics, SpectrumDataFrame from their new modules. - Bump TRANSFORMER_MODEL_ID to instanovo-v1.2.0. - Build dataloaders via process_dataset / collate_fn; read batches as dicts (batch["spectra"], batch["precursors"], batch["spectra_mask"]). - Handle the new dict return contract: GreedyDecoder / KnapsackBeamSearchDecoder / DiffusionDecoder all return dicts with predictions / mass_error / prediction_log_probability / prediction_token_log_probabilities; KnapsackBeamSearch now requires beam_size on decode(). - Update tests to use the new data processor API. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>043641a
-
Update InstaNovo to v1.2.2 with reproducibility fixes
- Update instanovo dependency from >=1.1.4 to ==1.2.2 - Add set_seed() function for deterministic behavior across environments - Set random seeds for Python, NumPy, and PyTorch - Enable torch.backends.cudnn.deterministic for consistent CUDA results - Reset seed before each prediction call - Add comprehensive reproducibility tests This addresses the variability issue between local and HF Spaces runs reported in: https://github.com/instadeepai/InstaNovo/issues/922c21889