Examples#

The benchmarks/ directory of the repository contains eight complete, self-contained training examples – one per challenge / database – written against the modern library API. They also serve as the reference usage for the whole package: each one shows reader and dataset usage, model construction, training with the library trainer, and evaluation with the official challenge metrics where applicable.

Each benchmark is runnable as a module with a --smoke debug mode (a few hundred iterations on a subset, no full training):

python -m benchmarks.train_crnn_cinc2021.train --help
python -m benchmarks.train_crnn_cinc2021.train --smoke

The examples#

Benchmark	Task / Database	Architectures
`train_crnn_cinc2021`	26-class classification, CinC Challenge 2021 (all 6 lead sets)	`ECG_CRNN`
`train_crnn_cinc2020`	27-class classification, CinC Challenge 2020 (strict subset of the 2021 data)	`ECG_CRNN`
`train_unet_ludb`	ECG waveform delineation, LUDB	`ECG_UNET`
`train_multi_cpsc2019`	In-hospital arrhythmia detection, CPSC2019	`ECG_SEQ_LAB_NET`, `ECG_SUBTRACT_UNET`, `ECG_CRNN`
`train_hybrid_cpsc2021`	LOC/RLOC regression, CPSC2021 (3 sub-tasks)	`ECG_UNET_LUDB` variants
`train_hybrid_cpsc2020`	End-to-end + segmentation challenge score, CPSC2020	`ECG_CRNN`, `ECG_SEQ_LAB_NET`
`train_crnn_cinc2023`	24-class classification on ECG + metal-oxide gas sensors, CinC Challenge 2023	`CRNN_CINC2023`
`train_mtl_cinc2022`	Multi-task (classification + regression) on circulatory sounds, CinC Challenge 2022	`CRNN_CINC2022`, `Wav2Vec2` variants

Points to note#

Determinism: every train.py calls seed_everything() and uses worker_init_fn; a given --seed fully determines the run.
Library datasets/readers/metrics: the examples consume torch_ecg.databases readers and datasets instead of local copies, and compute official challenge metrics via the library.
Offline preparation: train_hybrid_cpsc2021 runs the (idempotent) offline preparation via persistence() automatically.
Checkpoints: written via save() with the full model config, loadable with from_checkpoint().

Tests of the examples#

test/test_pipelines/ contains thin pytest wrappers over the benchmarks (fast lanes run the non-downloading parts; the nightly full suite trains briefly on real data), so the examples are continuously verified against the library.