Sampling with ecgen sample

Every checkpoint written by ecgen train records its family and the training data contract (lead names, sampling rate, length, scaling, condition normalizers), so one call samples any model:

import ecgen

x = ecgen.sample("runs/diffusion_mimic_conditional/seed_42/checkpoints/last.ckpt", 16,
                 cond={"measurements": [800, 40, 128, 170, 258, 520, 60, 30, 40]},  # raw units
                 guidance=2.0, steps=50, seed=0,
                 output_dir="samples/", format="csv", plot=2)
# x: [16, 12, 5000] float32

ecgen sample runs/.../last.ckpt -n 16 -o samples/ --format csv \
    --cond measurements=800,40,128,170,258,520,60,30,40 --guidance 2 --steps 50 --seed 0 --plot 2
ecgen sample runs/.../last.ckpt -n 8 --cond report="Sinus tachycardia" --format wfdb

Argument	Meaning
cond	Raw condition values: numeric fields in their original units (NaN = unknown), class indices, multi-hot vectors, strings for text fields. A value with a leading dimension of n (an [n, F] array, n class indices, n strings) gives one condition per sample; anything else is used for all samples. Fields left out use the model's null embedding.
obs, obs_mask	Partially observed ECGs and masks (1 = observed), e.g. only lead I, for lead reconstruction / imputation
guidance	Classifier-free guidance scale (conditional diffusion, flow matching, PixelCNN prior)
steps	Sampling steps (diffusion / flow)
seed	Reproducible sampling
derive_12	Complete the 8 independent leads (I, II, V1–V6) to the standard 12 (Einthoven / Goldberger)
output_dir, format	npy (one array), csv (one file per ECG, lead-name header), wfdb (.hea/.dat), or None; samples.json records leads, fs, units, conditions
plot	Render the first N samples as interactive HTML (ecgen.plot_ecg)
family	Registry name, only for checkpoints that predate family recording

Other keyword arguments go to the model's sample() (e.g. eta, solver, inpaint="repaint", temperature). Units are mV for data trained with scale: none; z-scored models output normalized units (NU).

ecgen.load_model(checkpoint) returns the model itself, and hf://<repo>/<path> checkpoints are downloaded as for ecgen.generate. ecgen.generate (the original Pulse2Pulse .pt loader) is unchanged.