Release InstaNovo v1.1.3 · instadeepai/InstaNovo

What's Changed

Update codebase to use new diffusion checkpoint, instanovoplus-v1.1.0
Update diffusion predict script with refinement configuration options, and multiple prediction sampling for improved performance
- refine_all: If True, all predictions will be refined
- refine_threshold: Only predictions with a confidence score less than this will be refined
- n_preds: Number of diffusion predictions to sample per spectrum
Include updated model performance benchmarking in README.md
Added option to specify valid_path as a DictConfig, where keys represent the validation group and values represent the validation path.
- Eg. "acpt": "/path/to/acpt/*-valid-*.parquet"
- This will add an "acpt" group in the validation metrics
Added add_source_file_column option to SpectrumDataFrame, which adds the path to the original input file as a column "source_file"
Added add_spectrum_id option to SpectrumDataFrame, which adds unique index values to the input file as a column "spectrum_id"
Updated s3 utils to use a class instead of individual functions
- Added functionality to wrap write methods and optionally specify an s3 path which will get automatically uploaded
- Added functionality to automatically download files to a temporary directory when converting s3 paths
Allow --output-path to remain unspecified in evaluation mode for:
- Transformer model
- Diffusion model when not in refinement mode

Trained a new InstaNovo+ v1.1.0 checkpoint on the Extended Massive-KB dataset - instanovoplus-v1.1.0.ckpt. To be used standalone, or in conjunction with instanovo-v1.1.0.ckpt for best results.

Full Changelog: 1.1.2...1.1.3