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.
25 lines
1.2 KiB
JSON
25 lines
1.2 KiB
JSON
[
|
|
{
|
|
"_note": "Sensitivity: refuge V2-M ensemble (img + cd, bilateral, ortho-w0.1 inner-Hadamard not applied here — matches the plain V2-M ensemble baseline at 224 from efficientnet/refuge_efficientnetv2_m which gave 0.9132 ± 0.019). This run feeds EfficientNetV2-M at its NATIVE 480x480 input resolution instead of the pipeline default 224. Activation memory roughly 4.6x; bf16 autocast keeps bs=8 fit in 16 GB (probe peak 9.07 GB on 7800 XT). Matched seed/fold_seed (1234/100) inherited from ensemble_fused.json so reps 1-10 here pair with reps 1-10 of the 224 baseline for paired statistics. 10 reps queued — kill early if wall-clock proves prohibitive. save_checkpoints + save_predictions enabled for downstream analysis if the result is promising.",
|
|
"run_name": "experiments/sensitivity/v2m_at_480_amp",
|
|
"reps": 10,
|
|
"overrides": {
|
|
"save_checkpoints": true,
|
|
"save_predictions": true,
|
|
"training": {
|
|
"amp": true,
|
|
"amp_dtype": "bfloat16"
|
|
}
|
|
},
|
|
"tower_overrides": {
|
|
"img": {
|
|
"args": {
|
|
"backbone": "refuge_efficientnet_v2_m",
|
|
"freeze_ratio": 0.0,
|
|
"crop_size": 480
|
|
}
|
|
}
|
|
}
|
|
}
|
|
]
|