How perslev/u-time works
U-Time is a command-line toolkit for training and evaluating U-Time and U-Sleep models for sleep staging and time-series segmentation. The documented workflow initializes a project, fetches PSG datasets, prepares data splits, trains a model, and predicts or evaluates results. Internally, dataset sequences form batches and apply augmentation; model-building and training utilities support optimization, while evaluation and prediction provide separate result paths. The README describes additional CLI capabilities, but several command implementations are unsampled, so only documented links are shown for them.
Main entry points: ut.py U-Time CLI · utime.py U-Time network · train.py Training workflow · batch_sequence.py BatchSequence
Components 27
Project workflow
- U-Time CLI
utime/bin/ut.py - Project initialization
utime/bin/init.py - Dataset fetching
utime/bin/fetch.py - Hyperparameters
utime/hyperparameters/__init__.py
Data and batching
- Data preparation
utime/bin/preprocess.py - Dataset splits
utime/bin/cv_split.py - Batch sequences
utime/sequences/base_sequence.py - BatchSequence
utime/sequences/batch_sequence.py - Multi-dataset sequence
utime/sequences/multi_sequence.py - Random sampling
utime/sequences/random_batch_sequence.py - Balanced sampling
utime/sequences/balanced_random_batch_sequence.py - Signal augmentation
utime/augmentation/augmenters.py
Models and training
- Model initialization
utime/models/model_init.py - U-Sleep network
utime/models/usleep.py - U-Time network
utime/models/utime.py - Convolution utilities
utime/utils/conv_arithmetics.py - Training engine
utime/train/trainer.py - Training utilities
utime/train/utils.py - Training callbacks
utime/callbacks/callbacks.py - Training workflow
utime/utils/scriptutils/train.py - GPU selection
utime/utils/system.py
Prediction and evaluation
- Prediction workflow
utime/utils/scriptutils/predict.py - Single-study prediction
utime/bin/predict_one.py - Evaluation workflow
utime/bin/evaluate.py - Evaluation metrics
utime/evaluation/metrics.py - Confusion analysis
utime/bin/cm.py
Other components
- Researcher
Connections 27
- Researcher → U-Time CLI: runs commands
- U-Time CLI → Project initialization: initializes project
- U-Time CLI → Dataset fetching: fetches PSG
- U-Time CLI → Dataset splits: splits datasets
- U-Time CLI → Training workflow: trains models
- U-Time CLI → Evaluation workflow: evaluates models
- U-Time CLI → Single-study prediction: predicts studies
- Training workflow → Batch sequences: builds sequences
utime/utils/scriptutils/train.py - BatchSequence → Batch sequences: inherits from
utime/sequences/batch_sequence.py - Batch sequences → Signal augmentation: applies augmentation
utime/sequences/base_sequence.py - Multi-dataset sequence → BatchSequence: uses batches
- Random sampling → BatchSequence: specializes batching
- Balanced sampling → BatchSequence: specializes batching
- Training workflow → Model initialization: initializes model
utime/models/model_init.py - Model initialization → U-Time network: selects model
- Model initialization → U-Sleep network: selects model
- U-Time network → Convolution utilities: uses arithmetic
utime/models/utime.py - U-Sleep network → Convolution utilities: uses arithmetic
utime/models/usleep.py - U-Time network → Training utilities: uses training utilities
utime/models/utime.py - U-Sleep network → Training utilities: uses training utilities
utime/models/usleep.py - Training engine → Training callbacks: configures callbacks
utime/train/trainer.py - Training engine → Training utilities: uses training utilities
utime/train/trainer.py - Training workflow → GPU selection: selects GPUs
- Evaluation workflow → Prediction workflow: predicts sequences
utime/bin/evaluate.py - Evaluation workflow → Evaluation metrics: computes metrics
utime/bin/evaluate.py - Single-study prediction → U-Time network: loads model
utime/bin/predict_one.py - Confusion analysis → U-Time network: uses model
utime/bin/cm.py