Command line

dietnet --help
dietnet <command> --help
dietnet --version

command	does
create-dataset	PLINK + labels → upstream's HDF5 dataset
partition	dataset → folds (upstream partition_data.py)
generate-embedding	per-fold genotype × class frequencies (upstream generate_embedding.py)
build-embedding	the auxiliary network's input from several feature sources
alphagenome	AlphaGenome variant effects per SNP, for build-embedding
compute-input-stats	per-fold SNP means and SDs (upstream compute_input_features_statistics.py)
train	upstream train.py per (seed, fold) → model packages
match	a cohort on the models' SNPs, ahead of predict
predict	a cohort through every selected model, combined
check	accuracy against labels, or each package's held-out accuracy
published	the 15 published models, as packages
info	versions, and whether the upstream code is intact

Options that take several integers (--seeds, --folds) accept comma-separated values (--seeds 78,79,80) or the flag repeated (--seeds 78 --seeds 79). Paths marked existing must exist.

dietnet create-dataset

Build upstream's HDF5 dataset from PLINK, without the text round-trip (schema). At least one of --labels and --regression-labels is required.

option		default
--plink TEXT	PLINK prefix (.bed/.bim/.fam)	required
--labels PATH	existing TSV: sample id, class label
--regression-labels PATH	existing TSV: sample id, value
--out PATH	output .h5	required

dietnet partition

Split samples into folds (upstream partition_data.py).

option		default
--dataset PATH	existing dataset	required
--out PATH	output .npz (.npz is appended if missing)	required
--nb-folds INTEGER	number of folds	5
--seed INTEGER	random seed	23
--train-valid-ratio FLOAT	train share of train + valid; unbalanced split only	0.75
--balanced / --unbalanced	balance class labels across splits, as in the paper	balanced

dietnet generate-embedding

Per-fold genotype × class frequency embedding (upstream generate_embedding.py). Stops, naming them, if a class has no training samples in some fold.

option		default
--dataset PATH	existing dataset	required
--partition PATH	existing partition	required
--out PATH	output .npz	required
--ncpus INTEGER	processes	1

dietnet build-embedding

SNP features for the auxiliary network from several sources; writes embedding.npz and embedding.npz.json. See SNP features.

option		default
--dataset PATH	existing dataset	required
--partition PATH	existing partition	required
--spec PATH	existing YAML list of feature blocks	required
--out PATH	output .npz	required

dietnet alphagenome

AlphaGenome variant effects for every SNP, through the hosted API, with the key in ALPHAGENOME_API_KEY. Writes a table for build-embedding's precomputed provider (orient: effect). Resumable: rerun the same command after an interruption. Needs the alphagenome extra, and sends every SNP's coordinates and alleles to Google's API; see SNP features.

option		default
--dataset PATH	existing dataset; SNP ids must be chr:pos:REF:ALT on GRCh38	one of these two
--bim PATH	a PLINK .bim instead: only its SNP ids are read (same chr:pos:REF:ALT form), so no genotypes are needed	one of these two
--out PATH	feature table (.tsv)	required
--scorers TEXT	AlphaGenome recommended scorers, comma-separated	ATAC,DNASE,CHIP_TF,CHIP_HISTONE,CAGE,RNA_SEQ
--sequence-length	context around each variant: 16KB, 100KB, 500KB or 1MB	1MB
--chunk INTEGER	SNPs per saved part	200
--max-workers INTEGER	parallel API requests	5

dietnet adjust-phenotype

A trait regressed on covariates and standardised, as regression labels for create-dataset --regression-labels: the residual of trait ~ 1 + covariates divided by its SD. Text covariates (e.g. assessment centre) become indicator columns. Samples missing the trait or a covariate are left out; the summary (samples, covariates' R², coefficients) goes to <out>.json. For height: --trait height --covariates sex,age, since a model of autosomes cannot see sex.

option		default
--phenotypes PATH	existing TSV, header row, sample id first	required
--trait TEXT	the column to predict	required
--covariates TEXT	columns to adjust for, comma-separated	none
--out PATH	labels TSV	required

dietnet pack-genotypes

Write a 2-bit copy of the dataset's genotypes next to it (<dataset>.packed.npy, with a .packed.json recording what it was built from), for train --inputs packed: a quarter of the bytes, read a batch at a time. Rerun it after the dataset changes; a stale copy is refused.

option		default
--dataset PATH	existing dataset	required

dietnet compute-input-stats

Per-fold SNP means and SDs (upstream compute_input_features_statistics.py).

option		default
--dataset PATH	existing dataset	required
--partition PATH	existing partition	required
--out PATH	output .npz	required
--ncpus INTEGER	processes	1

dietnet train

Train each (seed, fold) with upstream train.py; package the results under EXP_DIR/packages/seed_<seed>/fold_<fold>/. See training.

option		default
--dataset PATH	existing dataset	required
--partition PATH	existing partition	required
--embedding PATH	existing auxiliary-network input: required for the Diet Network, refused for the MLP
--input-stats PATH	existing per-fold statistics	required
--config PATH	existing upstream training config (YAML)	required
--exp-dir PATH	runs, logs, configs and packages	required
--seeds TEXT	e.g. 78,79,80; each overrides the config's seed	required
--folds TEXT	e.g. 0,1,2,3,4	required
--task	classification or regression	classification
--normalize / --no-normalize	standardise inputs with the fold's means and SDs	normalize
--model	dietnet, upstream's Diet Network, or mlp, its MLP baseline (no auxiliary network)	dietnet
--inputs	where training reads genotypes: ram copies them all into memory first (upstream), memmap reads each batch from the dataset file, packed from its 2-bit copy (pack-genotypes); same model in every case. auto uses the packed copy if there is one, else memory-maps above 2 GiB	auto
--loader	fast reads each batch in one call with loader workers and fills missing values in one operation; upstream leaves train.py's loader alone. Same model	fast
--workers INTEGER	loader processes for --loader fast	the CPUs available, less one, at most 8
--gpus INTEGER	above 1, train each model data-parallel over this many GPUs on one node (CPU processes where there are none): upstream's model and epoch logic, the config's batch_size split across them; not bit-identical to train.py, see Scaling	1
--engine	upstream: train.py itself, unmodified (bit-identical model); dietnet: upstream's model and epoch logic in dietnet's own loop, packed genotypes decoded on the GPU, without train.py's per-batch bookkeeping (much faster; not bit-identical)	upstream on one GPU, dietnet for --gpus > 1

dietnet match

Match a cohort to the models' SNPs, ahead of predict. CPU and I/O only: run it as its own job, then dietnet predict with the same --work-dir on a GPU, which finds the matched cohort and goes straight to inference. Prints the matching counts, and warns if fewer than 1,000 model SNPs were found.

option		default
--model PATH	existing package, or a directory of packages	required
--plink TEXT	cohort PLINK prefix	required
--work-dir PATH	where the matched cohort HDF5 goes	required
--seeds TEXT	only these seeds	all
--folds TEXT	only these folds	all

dietnet predict

Predict a cohort with every selected model and combine them (output). Warns if fewer than 1,000 model SNPs were found.

option		default
--model PATH	existing package, or a directory of packages	required
--plink TEXT	cohort PLINK prefix	required
--out PATH	predictions TSV	required
--work-dir PATH	where the matched cohort HDF5 goes, or already is	next to --out
--seeds TEXT	only these seeds	all
--folds TEXT	only these folds	all
--batch-size INTEGER	samples read and scored at a time	1024
--device TEXT	cpu, cuda, ...	cuda if available
--scores PATH	also save per-model scores (.npz)

dietnet check

Accuracy of predictions against labels, overall and per class; or, with --packages, each package's accuracy on its own held-out test fold (from the test_predictions.tsv that train writes). Give --packages, or --predictions with --labels.

option		default
--predictions PATH	existing output of dietnet predict
--packages PATH	existing directory of packages from train
--labels PATH	existing TSV: sample id, label
--min-accuracy FLOAT	exit with status 1 below this

dietnet published

Fetch the 15 published models (Zenodo record 21133813) as model packages, in DEST/packages/. The archive is checked against Zenodo's checksum.

option		default
--dest PATH	where the download and packages go	required
--release PATH	existing already unpacked Dietnet_training/ (skips the download)

dietnet info

Versions, and whether the upstream code is exactly the pinned commit. Exits with status 1, naming the files, if not. No options.