naamiinepal / xrayto3D-benchmark

GNU General Public License v3.0
9 stars 3 forks source link

(test) define swinunetr architecture hyperparameter configuration #15

Closed msrepo closed 1 year ago

msrepo commented 1 year ago

SwinUNETR lr 2e-4 batch_size 8 feature_size: {12,24,48} num_heads: default (3,6,12,24), Tiny (2,2,2,2)

msrepo commented 1 year ago

we need a driver python script specific to swinunetr to which configuration can be passed.

sweep_config = {
    "program": "test_sweep.py",
    "method": "grid",
    "early_terminate": {
        "type": "hyperband",
        "min_iter": 3,
    },
    "metric": {"name": "val/dice", "goal": "maximize"},
    "parameters": {
        "feature_size": {"values": [12, 24, 48]},
        "num_heads": {"values": [(3, 6, 12, 24), (2, 2, 2, 2)]},
    },
}

current configuration looks something like this. fill in the required details to call train.py

msrepo commented 1 year ago

had issue in passing list (for num_heads) to the driver program via argparse. image

So, instead of passing list, we pass a string placeholder like this image and expand in the driver program later like this image