CLI reference

Generated from the typer application in src/inmotion/cli/ by walking the click command tree, so it lists exactly what the program accepts.

inmotion --help

analyze

Analyse results.

analyze interference

Cross-route interference from the concurrent_noise_path column.

Options:

• --data (default: data/processed/dataset.csv) - Dataset CSV with concurrent_noise_path.
• --output-dir (default: plots/interference) - Figure output directory.

analyze seeds

Combine and plot results from multi-seed runs.

Options:

• --results-dir (default: results) - Directory of per-seed result CSVs.
• --output-dir (default: combined) - Where to write the combined analysis.
• --seeds (default: 3,5,42) - Comma-separated seeds to combine.

data

Build and clean datasets.

data augment

Augment a dataset with noise injection and intra-class mixup.

Options:

• --data - Dataset CSV to augment.
• --output, -o - Destination CSV.

data clean

Audit a dataset per device and drop rows that poison training.

Options:

• --data - Dataset CSV to audit.
• --output - Write the cleaned CSV here.
• --keep-shifted (default: False) - Keep device-shifted but self-consistent rows.
• --drop-ambiguous (default: False) - Also drop samples the model rejects.
• --ambiguity-thresh (default: 0.9) - Confidence above which a wrong prediction is dropped.

data export

Turn a raw Wavecom capture into a per-route CSV.

Options:

• --input, -i - Raw Wavecom capture (.txt).
• --output, -o - Destination CSV.
• --group - label:mac1,mac2,… (repeatable).
• --group-label - group_name:final_label (repeatable).

data merge

Merge per-route CSVs into one dataset.

Options:

• --input, -i - Directory of per-route CSVs.
• --output, -o - Merged dataset CSV.
• --pattern (default: *.csv) - Glob for input files.

datasets

Dataset versions and provenance.

datasets list

List recorded dataset versions, newest lineage last.

Options:

• --json (default: False) -

datasets manifest

Regenerate the manifest’s measured fields, or print it.

Options:

• --write (default: False) - Rewrite data/DATASETS.json.

datasets resolve

Print the path of a dataset version, warning when it is not trustworthy.

Arguments:

• IDENTIFIER (required) - Dataset id or file name.

datasets show

Show lineage and label provenance for one dataset version.

Arguments:

• IDENTIFIER (required) - Dataset id or file name.

datasets verify

Re-check every dataset against its recorded checksum and row count.

doctor

Report on the environment and the artifact store.

evaluate

Evaluate and combine models.

evaluate checkpoints

Retroactive per-class metrics for saved DL checkpoints.

Options:

• --data (default: /home/andre/Desktop/UA/IT/inMotion/inMotion/data/processed/dataset.csv) - Dataset CSV.
• --seed (default: 42) - Random seed.
• --checkpoints-dir (default: models/dl) - Directory of .pt checkpoints.
• --output-dir (default: results/dl) - Where to write the results.
• --optuna-db - Optuna storage URL, for HPO checkpoints.
• --no-wandb (default: False) - Disable WandB logging.

evaluate ensemble

Greedy ensemble selection over existing checkpoints.

Options:

• --data (default: /home/andre/Desktop/UA/IT/inMotion/inMotion/data/processed/dataset.csv) - Dataset CSV.
• --seed (default: 42) - Random seed.
• --rich (default: False) - Use the 18-channel feature set.

evaluate feature-importance

Permutation feature importance for the DL models.

Options:

• --data (default: /home/andre/Desktop/UA/IT/inMotion/inMotion/data/processed/dataset.csv) - Dataset CSV.
• --seed (default: 42) - Random seed.
• --epochs (default: 100) - Epochs per model.
• --patience (default: 20) - Early-stopping patience.
• --output-dir (default: docs/icaisf/paper/images) - Figure output directory.
• --checkpoints-dir (default: models/dl) - Directory of DL checkpoints.

evaluate mega-ensemble

OOF stacking over every model family, with calibration.

Options:

• --data (default: /home/andre/Desktop/UA/IT/inMotion/inMotion/data/processed/dataset.csv) - Dataset CSV.
• --seed (default: 42) - Random seed.
• --size-frontier (default: False) - Report the ensemble-size frontier.
• --no-tabpfn (default: False) - Exclude TabPFN.
• --no-classical (default: False) - Exclude the classical ML members.
• --output - Write the result table here.

evaluate shap

SHAP channel attribution for the world models.

Options:

• --checkpoint (default: backup/ts_jepa_ft_seed42.pt) - Model checkpoint to explain.
• --data (default: /home/andre/Desktop/UA/IT/inMotion/inMotion/data/processed/dataset.csv) - Dataset CSV.
• --model (default: ts_jepa) - Model family: t_jepa, ts_jepa, lejepa, cf_jepa or sigreg.
• --samples (default: 50) - Number of samples to explain.
• --nsamples (default: 150) - SHAP evaluation budget.
• --seed (default: 42) - Random seed.
• --output-dir (default: plots/shap) - Figure output directory.

model

Inspect, verify and run trained models.

model adopt

Recover a checkpoint’s architecture and write a sidecar.

Use –factory to resolve structural ambiguity such as TS-JEPA versus CF-JEPA at the same width, which are indistinguishable from weights alone.

Arguments:

• CHECKPOINT (required) - Path to a .pt checkpoint.

Options:

• --name - Name for the adopted model.
• --factory - Pin the architecture when ambiguous.
• --write - Write a sidecar-tagged copy to this directory.

model info

Show the architecture and provenance of one model.

Arguments:

• NAME (required) - Registered model name.

model inventory

Adopt every checkpoint under a directory and report what is recoverable.

Arguments:

• DIRECTORY (required) - Directory to scan for .pt files.

Options:

• --json (default: False) -

model list

List every registered model.

Options:

• --json (default: False) - Machine-readable output.

model predict

Run a trained model over a CSV of RSSI windows.

The model’s preprocessing statistics were never persisted by the original training code, so scaling is reconstructed from --reference (the same CSV the model trained on) unless a saved pipeline exists next to the weights.

Arguments:

• NAME (required) - Registered model name.

Options:

• --input, -i - CSV of RSSI rows.
• --reference - CSV used to fit scaling statistics.
• --out, -o -

model verify

Strict-load every model and check it against its recorded parameter count.

Arguments:

• NAME (optional) - Verify one model.

plots

Regenerate figures from saved results.

plots classical

Rebuild the classical-ML figures from saved CSVs.

Options:

• --csv - A single classification_results.csv.
• --results-dir - A directory of result CSVs.
• --plots-dir (default: plots) - Figure output directory.
• --eda (default: False) - Also regenerate the EDA figures.
• --data - Dataset CSV, needed with –eda.

plots dl

Rebuild the DL figures from saved CSVs.

Options:

• --results-csv (default: results/dl/dl_detailed_seed42.csv) - Extended DL results CSV.
• --output-dir (default: plots/dl) - Figure output directory.
• --eda (default: False) - Also regenerate the EDA figures.
• --data - Dataset CSV, needed with –eda.

plots quantization

Rebuild the quantization figures from the statistics CSVs.

Options:

• --data-dir (default: results/quant) - Statistics CSVs and device benchmark JSONs.
• --perf-dir (default: results/perf) - Before/after latency CSVs.
• --out-dir (default: results/quant/figures) - Figure output directory.
• --dpi (default: 600) - Raster DPI (PDF is vector).

quantize

Quantize models and measure the cost.

quantize eval

Measure accuracy under quantization on dataset.csv.

Options:

• --models - Comma-separated; default: all five world models.
• --formats - Comma-separated; default: DEFAULT_PRECISIONS.
• --data - Dataset CSV; default: dataset.csv.
• --seed (default: 42) - Split seed.
• --out (default: results/quant/sweep_dataset.csv) - Where to write the results CSV.
• --producer (default: in-house) - Measure a producer’s schemes instead of the in-house presets: in-house, torch-ao or bitsandbytes. See quantize producers.
• --allow-unrecorded-baseline (default: False) - Accept a model with no recorded fp32 baseline, such as a checkpoint from another training run. A recorded model that has MOVED still fails.

quantize export

Write a portable artifact for a non-Python runtime.

Arguments:

• MODEL (required) - Registered model name.

Options:

• --out, -o - Destination directory.
• --format, -f (default: fp32) - Scheme, or fp32 for none. fp16/bf16/fp8 are cast precisions and cannot be written as artifacts.
• --producer (default: in-house) - Which quantizer: in-house, torch-ao or bitsandbytes. inmotion quantize producers lists them; with a producer, fp32 means that producer’s own default scheme.
• --golden (default: True) - Also write golden vectors.
• --tol (default: 0.0002) - Parity tolerance recorded in golden.json.

quantize formats

List the available quantization schemes.

quantize producers

List the quantizers and what the runtime can read.

quantize show

Quantize one model and report size and error.

Arguments:

• MODEL (required) - Registered model name.

Options:

• --format, -f (default: int8) - Scheme, e.g. q4.

quantize stats

Write the quantization statistics tables the figures read.

Options:

• --models - Comma-separated; default: all five world models.
• --formats - Comma-separated; default: DEFAULT_PRECISIONS.
• --data - Dataset CSV; default: dataset.csv.
• --seed (default: 42) - Split seed.
• --sweep (default: results/quant/sweep_dataset.csv) - Accuracy/size CSV from quantize eval.
• --out-dir (default: results/quant) - Where the tables and raw capture go.
• --latency-iters (default: 200) - Batch-1 torch reference timings; 0 disables.
• --latency-warmup (default: 20) - Warmup iterations per measurement.

train

Train models.

train classification

Classical ML classifiers: EDA, CV, Optuna, importance.

Options:

• --data (default: /home/andre/Desktop/UA/IT/inMotion/inMotion/data/processed/dataset.csv) - Dataset CSV.
• --seed (default: 42) - Random seed.
• --optimize (default: False) - Run the Optuna tuning stage.
• --dashboard (default: False) - Launch optuna-dashboard after tuning.
• --models-dir (default: models) - Model output directory.
• --results-dir (default: results) - Results output directory.
• --plots-dir (default: plots) - Figure output directory.

train dl

Supervised deep learning: baselines, HPO/NAS, MoE, DeepStack.

Options:

• --data (default: /home/andre/Desktop/UA/IT/inMotion/inMotion/data/processed/dataset.csv) - Dataset CSV.
• --seed (default: 42) - Random seed.
• --epochs (default: 150) - Max epochs per model.
• --trials (default: 50) - Optuna trials per study.
• --batch-size (default: 64) - Batch size.
• --num-gpus (default: 0) - GPUs to use; 0 = auto-detect.
• --models-dir - Checkpoint output directory.
• --results-dir - Results CSV directory.
• --plots-dir - Figure output directory.
• --wandb-project - Override the WandB project.
• --no-wandb (default: False) - Disable WandB logging.
• --no-optuna (default: False) - Skip HPO and NAS studies.
• --no-moe (default: False) - Skip the mixture-of-experts phase.
• --no-deepstack (default: False) - Skip the DeepStack phase.
• --no-meta (default: False) - Skip the metadata-fusion phases.
• --no-resume (default: False) - Retrain even when a checkpoint exists.

train exotic

JEPA world models and SIGReg: SSL pretraining and fine-tuning.

Options:

• --model (default: t_jepa) - t_jepa, ts_jepa, lejepa, cf_jepa, sigreg or mamba3_.
• --data (default: /home/andre/Desktop/UA/IT/inMotion/inMotion/data/processed/dataset.csv) - Dataset CSV.
• --seed (default: 42) - Random seed.
• --batch-size (default: 128) - Batch size.
• --gpu (default: 0) - GPU device index.
• --rich-features (default: False) - Use the 18-channel feature set.
• --pretrain-epochs (default: 300) - Stage-1 epochs.
• --finetune-epochs (default: 50) - Stage-2 epochs.
• --epochs (default: 150) - Epochs for directly-supervised models.
• --patience (default: 25) - Early-stopping patience.
• --lr (default: 0.001) - Learning rate for directly-supervised models.
• --pretrain-lr (default: 0.0003) - Stage-1 learning rate.
• --finetune-lr (default: 0.001) - Stage-2 learning rate.
• --probe-cadence (default: 10) - Epochs between linear probes.
• --probe-patience (default: 15) - Probe-based patience.
• --checkpoint - Resume from a pretrained encoder.
• --finetune-only (default: False) - Skip pretraining.
• --d-model (default: 256) - Model width.
• --nhead (default: 8) - Attention heads.
• --num-layers (default: 4) - Transformer layers.
• --dim-ff (default: 512) - Feed-forward width.
• --pred-dim (default: 128) - Predictor width.
• --pred-layers (default: 2) - Predictor depth.
• --d-state (default: 16) - Mamba-3 state width.
• --dropout (default: 0.2) - Dropout for the Mamba-3 hybrids.
• --mimo-rank (default: 4) - Mamba-3 MIMO rank.
• --num-filters (default: 192) - SIGReg convolutional filters.
• --num-blocks (default: 3) - SIGReg residual blocks.
• --latent-dim (default: 128) - SIGReg latent width.
• --sigreg-lambda (default: 0.01) - SIGReg regulariser weight.
• --hpo (default: False) - Run an Optuna study first.
• --hpo-trials (default: 30) - Trials per study.
• --hpo-objective (default: mcc) - mcc, efficiency or pareto.
• --models-dir (default: models/exotic) - Checkpoint output directory.
• --results-dir (default: results/exotic) - Results CSV directory.
• --viz-dir (default: models/exotic/viz) - Figure output directory.
• --no-wandb (default: False) - Disable WandB logging.
• --wandb-project (default: inMotion-exotic-hpo) - WandB project name.

train hpo-paper

Retrain the paper’s HPO-best configurations.

Options:

• --data (default: /home/andre/Desktop/UA/IT/inMotion/inMotion/data/processed/dataset.csv) - Dataset CSV.
• --seed (default: 42) - Random seed.
• --epochs (default: 150) - Epochs per model.
• --patience (default: 20) - Early-stopping patience.
• --models - Comma-separated subset, e.g. gru,tcn.
• --models-dir (default: models/dl) - Checkpoint output directory.
• --results-dir (default: results/dl) - Results output directory.

Pipeline modules

Every training, evaluation and analysis pipeline exposes the same shape: a typed RunConfig dataclass and a run(args) -> int function. The CLI builds the config from typed options, so the pipelines contain no argument parsing and can be driven from Python directly:

from inmotion.pipelines.exotic import RunConfig, run

run(RunConfig(model='ts_jepa', seed=42, pretrain_epochs=500))

The two largest pipelines are packages with one module per responsibility:

pipeline	package	modules
inmotion train dl	inmotion.pipelines.dl
		config, devices, builders, workers, pipeline
inmotion train exotic	inmotion.pipelines.exotic
		config, devices, data, builders, training,
		hpo, hpo_jepa, hpo_sigreg, hpo_mamba3, pipeline

python -m inmotion.pipelines.<name> runs the same CLI for those packages.