rpotter6298
708fbc70ce
- Introduced `bridge_attention_ceiling_check.py` for variance decomposition analysis on bridge attention configurations. - Added `bridge_attention_readout.py` to perform per-tower gate and contribution readouts, including AUC sanity checks. - Created multiple JSON configuration files for backbone replication experiments, including anonymous CV variants and basic backbones. - Implemented sensitivity experiments to evaluate the impact of axial length inclusion and EfficientNetV2-M performance at higher resolutions. - Added a memory probe script to assess GPU memory usage during training with EfficientNetV2-M.
11 lines
612 B
JSON
11 lines
612 B
JSON
[
|
|
{
|
|
"_note": "10-rep checkpointed run of the production bilateral img+cd ensemble at the refugelike R50 backbone (headline configuration). Per-fold tower and stage_models state_dicts saved under each rep's checkpoints/foldN/ directory. Used for F8 explainability: reconstructing per-tower predictions at the nt (eye) and hb (patient) levels so we can compare img-tower-only, cd-tower-only, eye-fusion, and patient-fusion outputs on the headline backbone.",
|
|
"run_name": "experiments/explainability/ensemble_refugelike_ckpt",
|
|
"reps": 10,
|
|
"overrides": {
|
|
"save_checkpoints": true
|
|
}
|
|
}
|
|
]
|