UIMNET
This is UIMNET
an Imagenet scale uncertanity quantification benchmark suite.
Usage
Setup and installation
Install environment
conda env create -f conda_env.yaml
Activate environement
. ./script/setup
Install package in development mode
pip install -e .
Run Unit tests
pytests ./tests
Training
Single model
./scripts/run_trainer.py -a ALGORITHM -m MODEL_DIR -a ARCHITECTURE -c CLUSTERING_FILE
python3 -m torch.distributed.launch --nproc_per_node=8 ./scripts/run_trainer.py -a ALGORITHM -m MODEL_DIR -a ARCHITECTURE -c CLUSTERING_FILE -d
Sweep over all models
./run_training.py -s SWEEP_DIR
Calibration
Single model
./scripts/run_calibrator.py -m MODEL_DIR -c CLUSTERING_FILE
python3 -m torch.distributed.launch --nproc_per_node=8 ./scripts/run_calibrator.py -m MODEL_DIR -c CLUSTERING_FILE -d
Sweep over all models
./run_calibration.py -s SWEEP_DIR
In-domain prediction
Single model
./scripts/run_predictor.py -m MODEL_DIR -c CLUSTERING_FILE
python3 -m torch.distributed.launch --nproc_per_node=8 ./scripts/run_predictor.py -m MODEL_DIR -c CLUSTERING_FILE -d
Sweep over all models
./run_prediction.py -s SWEEP_DIR
Ensembling
./scripts/run_ensembles.py -s SWEEP_DIR
Out-of-domain evaluation
Single model
./scripts/run_evaluator.py -m MODEL_DIR -c CLUSTERING_FILE --measure MEASURE
python3 -m torch.distributed.launch --nproc_per_node=8 ./scripts/run_evaluator.py -m MODEL_DIR-c CLUSTERING_FILE -d --measure MEASURE
Sweep over all models
./scripts/run_evaluation.py -s SWEEP_DIR --measures MEASURE1 MEASURE2 ...
License
This source code is released under the MIT license, included Here.