Context
The STAMP integration supports three task types via STAMP_TASK:
classification (currently tested)
survival
regression
The deploy test (run_stamp_deploy_test.sh) currently only validates classification with the VIT model. Survival and regression tasks have code paths in stamp_training.py (e.g., survival-specific label handling at line 345) but are untested in a real multi-node swarm deployment.
Proposed Solution
- Create synthetic datasets for survival and regression tasks (extend
create_synthetic_stamp_dataset.py)
- Add additional model runs to the deploy test (e.g.,
run_single_model "STAMP_classification" "mlp" for other backbones)
- Optionally add a
STAMP_survival job configuration
Related Files
scripts/deploy/run_stamp_deploy_test.sh — currently runs only VIT classification
application/jobs/STAMP_classification/app/custom/stamp_training.py — supports survival/regression
application/jobs/STAMP_classification/app/scripts/create_synthetic_dataset/create_synthetic_stamp_dataset.py — only generates classification data
Context
The STAMP integration supports three task types via
STAMP_TASK:classification(currently tested)survivalregressionThe deploy test (
run_stamp_deploy_test.sh) currently only validatesclassificationwith the VIT model. Survival and regression tasks have code paths instamp_training.py(e.g., survival-specific label handling at line 345) but are untested in a real multi-node swarm deployment.Proposed Solution
create_synthetic_stamp_dataset.py)run_single_model "STAMP_classification" "mlp"for other backbones)STAMP_survivaljob configurationRelated Files
scripts/deploy/run_stamp_deploy_test.sh— currently runs only VIT classificationapplication/jobs/STAMP_classification/app/custom/stamp_training.py— supports survival/regressionapplication/jobs/STAMP_classification/app/scripts/create_synthetic_dataset/create_synthetic_stamp_dataset.py— only generates classification data