How osupcvlab/mobileunetr works
MobileUNETR is a medical-image segmentation model and experiment framework. The principal workflow is to launch a dataset-specific experiment, load image/mask pairs from CSV-backed ISIC datasets, augment and batch them, build a MobileUNETR variant, then train and validate it with configured losses, optimization, and segmentation metrics. The README also documents offline/online experiment tracking and reports results across several medical datasets. The graph follows the sampled 2016 runner and indexed counterparts; model internals and some trainer behavior are only partially excerpted.
Main entry points: exp_2_dice_b8_a2/run_experiment.py ISIC 2016 runner · exp_2_dice_b8_a2/run_experiment.py ISIC 2017 runner · exp_2_dice_b8_a2/run_experiment.py ISIC 2018 runner · exp_2_dice_b8_a2/run_experiment.py PH2 runner
Components 18
Experiment setup
- ISIC 2016 runner
experiments_medical/isic_2016/exp_2_dice_b8_a2/run_experiment.py - ISIC 2017 runner
experiments_medical/isic_2017/exp_2_dice_b8_a2/run_experiment.py - ISIC 2018 runner
experiments_medical/isic_2018/exp_2_dice_b8_a2/run_experiment.py - PH2 runner
experiments_medical/ph2/exp_2_dice_b8_a2/run_experiment.py
Data preparation
- CSV image/mask data
- Dataset builder
dataloaders/build_dataset.py - ISIC datasets
dataloaders/isic_dataset.py - Augmentation transforms
augmentations/segmentation_augmentations.py - Torchvision transforms
augmentations/segmentation_augmentations_tv.py
Segmentation model
- Architecture selector
architectures/build_architecture.py - MobileUNETR variants
architectures/mobileunetr.py
Training and evaluation
- Segmentation losses
losses/losses.py - Optimizer builder
optimizers/optimizers.py - Learning-rate schedulers
optimizers/schedulers.py - Segmentation trainer
train_scripts/segmentation_trainer.py - EMA updates
train_scripts/ema.py
Other components
- Experiment user
- Weights & Biases
Connections 38
- Experiment user → ISIC 2016 runner: launches
README.md - Experiment user → ISIC 2017 runner: launches
README.md - Experiment user → ISIC 2018 runner: launches
README.md - Experiment user → PH2 runner: launches
README.md - ISIC 2016 runner → Dataset builder: builds loaders
experiments_medical/isic_2016/exp_2_dice_b8_a2/run_experiment.py - ISIC 2017 runner → Dataset builder: builds loaders
experiments_medical/isic_2017/exp_2_dice_b8_a2/run_experiment.py - ISIC 2018 runner → Dataset builder: builds loaders
experiments_medical/isic_2018/exp_2_dice_b8_a2/run_experiment.py - PH2 runner → Dataset builder: builds loaders
experiments_medical/ph2/exp_2_dice_b8_a2/run_experiment.py - Dataset builder → ISIC datasets: selects dataset
dataloaders/build_dataset.py - Dataset builder → Augmentation transforms: uses transforms
dataloaders/build_dataset.py - Dataset builder → Torchvision transforms: uses transforms
dataloaders/build_dataset.py - CSV image/mask data → ISIC datasets: provides image paths
README.md - ISIC 2016 runner → Architecture selector: builds model
experiments_medical/isic_2016/exp_2_dice_b8_a2/run_experiment.py - ISIC 2017 runner → Architecture selector: builds model
experiments_medical/isic_2017/exp_2_dice_b8_a2/run_experiment.py - ISIC 2018 runner → Architecture selector: builds model
experiments_medical/isic_2018/exp_2_dice_b8_a2/run_experiment.py - PH2 runner → Architecture selector: builds model
experiments_medical/ph2/exp_2_dice_b8_a2/run_experiment.py - Architecture selector → MobileUNETR variants: selects variant
architectures/build_architecture.py - ISIC 2016 runner → Segmentation losses: builds criterion
experiments_medical/isic_2016/exp_2_dice_b8_a2/run_experiment.py - ISIC 2017 runner → Segmentation losses: builds criterion
experiments_medical/isic_2017/exp_2_dice_b8_a2/run_experiment.py - ISIC 2018 runner → Segmentation losses: builds criterion
experiments_medical/isic_2018/exp_2_dice_b8_a2/run_experiment.py - PH2 runner → Segmentation losses: builds criterion
experiments_medical/ph2/exp_2_dice_b8_a2/run_experiment.py - ISIC 2016 runner → Optimizer builder: builds optimizer
experiments_medical/isic_2016/exp_2_dice_b8_a2/run_experiment.py - ISIC 2017 runner → Optimizer builder: builds optimizer
experiments_medical/isic_2017/exp_2_dice_b8_a2/run_experiment.py - ISIC 2018 runner → Optimizer builder: builds optimizer
experiments_medical/isic_2018/exp_2_dice_b8_a2/run_experiment.py - PH2 runner → Optimizer builder: builds optimizer
experiments_medical/ph2/exp_2_dice_b8_a2/run_experiment.py - ISIC 2016 runner → Learning-rate schedulers: builds schedulers
experiments_medical/isic_2016/exp_2_dice_b8_a2/run_experiment.py - ISIC 2017 runner → Learning-rate schedulers: builds schedulers
experiments_medical/isic_2017/exp_2_dice_b8_a2/run_experiment.py - ISIC 2018 runner → Learning-rate schedulers: builds schedulers
experiments_medical/isic_2018/exp_2_dice_b8_a2/run_experiment.py - PH2 runner → Learning-rate schedulers: builds schedulers
experiments_medical/ph2/exp_2_dice_b8_a2/run_experiment.py - ISIC 2016 runner → Segmentation trainer: configures trainer
experiments_medical/isic_2016/exp_2_dice_b8_a2/run_experiment.py - ISIC 2017 runner → Segmentation trainer: configures trainer
experiments_medical/isic_2017/exp_2_dice_b8_a2/run_experiment.py - ISIC 2018 runner → Segmentation trainer: configures trainer
experiments_medical/isic_2018/exp_2_dice_b8_a2/run_experiment.py - PH2 runner → Segmentation trainer: configures trainer
experiments_medical/ph2/exp_2_dice_b8_a2/run_experiment.py - Segmentation trainer → EMA updates: updates EMA
train_scripts/segmentation_trainer.py - ISIC 2016 runner → Weights & Biases: initializes tracking
experiments_medical/isic_2016/exp_2_dice_b8_a2/run_experiment.py - ISIC 2017 runner → Weights & Biases: initializes tracking
experiments_medical/isic_2017/exp_2_dice_b8_a2/run_experiment.py - ISIC 2018 runner → Weights & Biases: initializes tracking
experiments_medical/isic_2018/exp_2_dice_b8_a2/run_experiment.py - PH2 runner → Weights & Biases: initializes tracking
experiments_medical/ph2/exp_2_dice_b8_a2/run_experiment.py