Training Entry Points

All families are trained by one runner, ecgen train <config.yaml> (see Training with ecgen train). The per-model scripts scripts/train/train_{vae_mimic,vqvae_mimic,prior_mimic,pulse2pulse}.py and their script-local DataModules were removed.

Entry point	What it runs
ecgen train <config> [key=value ...] / python -m ecgen train ...	The runner, for any family
scripts/shell/run_train_vae_mimic_config.sh [config] [overrides]	VAE (vae_mimic.yaml)
scripts/shell/run_train_vqvae_mimic.sh [config] [overrides]	VQ-VAE stage 1 (vqvae_mimic.yaml)
scripts/shell/run_train_vqvae_prior_mimic.sh <vqvae_ckpt> [config] [overrides]	PixelCNN prior (prior_mimic.yaml, model.params.vqvae_checkpoint=<vqvae_ckpt>)
scripts/shell/run_train_pulse2pulse_mimic.sh [config] [overrides]	Pulse2Pulse (pulse2pulse_mimic.yaml)
scripts/shell/run_train_pulse2pulse_standalone.sh	Pulse2Pulse, settings from env vars DATA_DIR, EXP_NAME, BATCH_SIZE, MAX_EPOCHS, GPU_ID, NUM_WORKERS
scripts/shell/run_sweep.sh [overrides]	Every configs/experiments/*.yaml except the prior, then the prior on runs/vqvae_mimic/seed_42/checkpoints/last.ckpt if it exists
python -m ecgen.training.train --config <yaml>	Deprecated wrapper that calls the runner
single_file/run_train_vqvae.sh → single_file/train_vqvae_standalone.py --stage {1,2}	Self-contained VQ-VAE (both stages), flags/env vars only, no package imports

All wrappers in scripts/shell/ first cd to the repository root, so they can be run from anywhere, and pass extra arguments through to ecgen train as overrides. Overrides are applied after the YAML, so the command line always wins.

scripts/train/train_template.py now only points at the runner: to add a model, register a family and write a config instead of a new script (see Extending).

Standalone VQ-VAE

single_file/train_vqvae_standalone.py re-implements the VQ-VAE, prior, dataset and training loop in one file and is independent of ecgen.families.vqvae. See Standalone VQ-VAE.