Under review at MICCAI 2024
Access to the TMED dataset can be obtained here.
torch 2.1.0 torchinfo 1.8.0 numpy 1.26.0 tqdm 4.66.1 scikit-learn 1.4.1 pickle 0.0.12
To train the baseline:
python train_vc.py --CONFIG avc_tmed2_resnet
To train w/ BackMix
python train_vc.py --CONFIG avc_tmed2_resnet_random_bg
To train w/ BackMix, but only apply aug to a fraction f of training examples (semi-supervised)
python train_vc.py --CONFIG avc_tmed_resnet_random_bg_semi_${f}
To train as above, but with wBackMix (lambda=1, f=0.05)
python train_vc_weighted.py --CONFIG avc_tmed_resnet_random_bg_semi_0_05_weighted_1
All training files require a random seed using the --SEED
arg.
To evaluate on TMED (regardless of config used for training)
python evaluate_vc.py --CONFIG avc_tmed_2 --RUN_ID {specify results folder which contains weights}
Weights are provided in weights/
.