Ongoing Research Contributions

[jakehlee]

This pull request introduces a new unmixing training pipeline for the SpecTfEncoder model, adds support for out-of-distribution (OOD) data injection during classifier training, and updates the data configuration to include OOD datasets. The main changes are the addition of two new scripts for unmixing evaluation and training, improvements to OOD handling in classifier training, and updates to the YAML config to support these features.

New unmixing pipeline and evaluation scripts:

• Added models/spectf/report_spectf_unmixing.py for evaluating unmixing models, including metrics computation, scatter plot generation, and saving results.
• Added models/spectf/training_spectf_unmixing.py for unmixing model training, supporting both standard and focal loss, OOD validation, and detailed WandB logging and reporting.

Enhancements to classifier training for OOD support:

• Modified models/spectf/training_spectf.py to inject OOD spectra into the training dataloader (inject_ood=True), and mask out OOD entries (label -1) during loss computation to prevent them from affecting backpropagation. Also logs OOD injection in WandB config. [1] [2] [3]

Data configuration updates:

• Updated models/spectf/v1/data_config.yaml to include paths for ood-train-set and set percent-ood-data in the dataloader section. [1] [2]

These changes collectively enable robust training and evaluation of unmixing models with explicit support for OOD data handling and reporting.