flygen-ml

Machine-learning tooling for asking whether a fly's trajectory carries enough information to identify labels such as genotype, antenna condition, cohort, or other manifest metadata.

The current pipeline focuses on between-reward trajectories from paired .data and .trx files. It extracts canonical trajectory segments, builds engineered movement features, exports low-level trajectory tensors, and trains fly-level classifiers with grouped evaluation.

The package is organized as a self-contained pipeline for this task: it reads paired .data and .trx recording files, extracts trajectory segments, builds feature tables, and writes model artifacts without requiring notebook-specific analysis state.

Input Data

The pipeline expects paired .data and .trx files from the lab's trajectory recording workflow. It does not depend on notebook state or external analysis objects, but it is format-tethered to these files: loaders assume the recording metadata, protocol fields, timestamps, and per-fly trajectory arrays used by that workflow.

Each manifest row points to one .data / .trx pair and one experimental fly within that recording. The loader uses those files to recover:

• selected training bounds
• experimental fly identity
• reward geometry and reward-entry / reward-exit events
• per-frame x / y trajectories for the selected fly
• manifest labels such as genotype, cohort, chamber, and training_idx

For the expected input contract, see docs/source-data-contract.md.

Current Scope

The project supports two complementary modeling paths:

• engineered fly-level baselines: per-segment movement summaries aggregated per fly, then classified with NumPy logistic regression / softmax regression
• sequence models: fixed-length, reward-normalized trajectory tensors encoded at the segment level, aggregated at the fly level, and classified with dual genotype/cohort output heads

The first target was genotype classification. The modeling path now also supports cohort for antennae-intact vs antennae-removed classification, and the strongest sequence models predict both axes jointly.

The current strongest model family is GRU-128 over ordered Conv1D segment embeddings, with head-specific fly-level pooling and behavior-derived side features. Across grouped CV seed sweeps, the current results are:

trajectory evidence	model	joint	genotype	genotype bal	cohort	cohort bal
Training 2 only	engineered logreg	0.664	0.738	0.724	0.900	0.897
Training 2 only	GRU-128	0.764	0.788	0.779	0.948	0.949
Training 1 only	engineered logreg	0.659	0.794	0.783	0.857	0.854
Training 1 only	GRU-128	0.809	0.861	0.856	0.929	0.928
Training 1 + Training 2	engineered logreg	0.677	0.774	0.757	0.882	0.881
Training 1 + Training 2	GRU-128	0.803	0.833	0.827	0.954	0.954

Training 1-only is currently strongest for genotype and joint classification. Combined Training 1 + Training 2 evidence is currently strongest for antenna-condition/cohort classification. In all tested regimes, GRU-128 outperforms the engineered logistic-regression baseline.

Installation

Use any Python environment you prefer: Conda, venv, uv, or another local environment manager. From the repository root, install the package in editable mode:

pip install -e .[dev]

This installs the runtime dependencies plus development and test dependencies from pyproject.toml.

For a runtime-only install:

pip install -e .

After installation, CLI commands should work as python -m flygen_ml.... For a temporary shell-only alternative without installing the package:

export PYTHONPATH="$PWD/src"

End-to-End Pipeline

The typical workflow writes intermediate artifacts under artifacts/ and model runs under runs/.

1. Build A Manifest

If a manifest already exists, this step can be skipped. A manifest row identifies one experimental fly in one recording and includes labels such as genotype and cohort.

python -m flygen_ml.cli.build_manifest_from_globs \
  --spec configs/manifest_globs/yang_2025_antennae_kir_v1.csv \
  --output artifacts/manifest.csv \
  --repeat-fly-indices 0,1

For Training 1 runs, use the Training 1 glob spec and a separate manifest path:

python -m flygen_ml.cli.build_manifest_from_globs \
  --spec configs/manifest_globs/yang_2025_antennae_kir_t1.csv \
  --output artifacts/manifest_t1.csv \
  --repeat-fly-indices 0,1

Expected manifest columns include:

• sample_key
• data_path
• trx_path
• genotype
• cohort
• date
• chamber
• training_idx
• fly_idx

2. Extract Between-Reward Segments

python -m flygen_ml.cli.extract_segments \
  --config configs/dataset/v1_binary.yaml \
  --manifest artifacts/manifest.csv \
  --output artifacts/segments_with_cohort.csv

For Training 1:

python -m flygen_ml.cli.extract_segments \
  --config configs/dataset/v1_binary.yaml \
  --manifest artifacts/manifest_t1.csv \
  --output artifacts/segments_t1_with_cohort.csv

The segment table keeps source provenance and frame boundaries so raw trajectory slices can be recovered later. It also carries manifest metadata such as genotype and cohort.

3. Build Fly-Level Features

python -m flygen_ml.cli.build_features \
  --feature-set engineered_v1 \
  --segments artifacts/segments_with_cohort.csv \
  --output artifacts/features_antennae_no_training_end.csv

For Training 1:

python -m flygen_ml.cli.build_features \
  --feature-set engineered_v1 \
  --segments artifacts/segments_t1_with_cohort.csv \
  --output artifacts/features_t1_antennae_no_training_end.csv

By default, feature building omits segments that ended only because the selected training ended. To include those segments:

python -m flygen_ml.cli.build_features \
  --feature-set engineered_v1 \
  --segments artifacts/segments_with_cohort.csv \
  --output artifacts/features_with_training_end.csv \
  --include-training-end-segments

The feature table is one row per fly. Metadata columns such as genotype, cohort, chamber_type, and training_idx are preserved for grouping and labeling, but are excluded from numeric model features.

The combined Training 1 + Training 2 feature table used in the current experiments is one row per fly in artifacts/features_t1_t2_antennae_no_training_end.csv. It should be built by combining the T1 and T2 fly-level feature rows into a single fly-level row, not by treating T1 and T2 as independent labeled examples.

4. Train A Genotype Classifier

Genotype is the default target label.

python -m flygen_ml.cli.train_model \
  --config configs/model/logreg.yaml \
  --features artifacts/features_antennae_no_training_end.csv \
  --output runs/logreg_v1_movement_only

5. Train An Antenna-Condition Classifier

Use --label-key cohort to classify antennae-intact vs antennae-removed.

python -m flygen_ml.cli.train_model \
  --config configs/model/logreg.yaml \
  --features artifacts/features_antennae_no_training_end.csv \
  --output runs/logreg_v1_movement_only_antennae_condition \
  --label-key cohort

The same target can also be set in a config file:

label_key: cohort

target_key is accepted as a backward-compatible alias, but label_key is the preferred name.

6. Evaluate A Saved Run

python -m flygen_ml.cli.evaluate_model \
  --run-dir runs/logreg_v1_movement_only_antennae_condition

This prints a compact text summary. Use --json to print the raw saved metrics payload. The command auto-detects holdout runs and grouped CV runs.

To include a confusion matrix and misclassified validation rows:

python -m flygen_ml.cli.evaluate_model \
  --run-dir runs/logreg_v1_movement_only_antennae_condition \
  --confusion \
  --misclassifications

7. Run Grouped Cross-Validation

python -m flygen_ml.cli.train_model \
  --config configs/model/logreg.yaml \
  --features artifacts/features_antennae_no_training_end.csv \
  --output runs/logreg_v1_movement_only_antennae_condition_cv \
  --label-key cohort \
  --cv-folds 5

The splitter groups by group_key from the model config, currently fly_id, and stratifies using the active label_key.

The same evaluation command works for CV runs:

python -m flygen_ml.cli.evaluate_model \
  --run-dir runs/logreg_v1_movement_only_antennae_condition_cv \
  --confusion \
  --misclassifications

8. Export Trajectory Tensors

The sequence-model path uses the segment table as an index back into the raw .trx trajectories. It exports fixed-length, reward-normalized per-segment tensors while preserving fly labels for fly-level aggregation.

python -m flygen_ml.cli.export_sequence_tensors \
  --segments artifacts/segments_with_cohort.csv \
  --output artifacts/sequences_v1.npz \
  --target-length 128

For Training 1:

python -m flygen_ml.cli.export_sequence_tensors \
  --segments artifacts/segments_t1_with_cohort.csv \
  --output artifacts/sequences_t1_v1.npz \
  --target-length 128

The default channels are x_rel, y_rel, dx_rel, dy_rel, speed_rel, and r_rel, where positions are centered on the reward location and scaled by the reward radius.

The combined Training 1 + Training 2 sequence artifact used in the current experiments is artifacts/sequences_t1_t2_v1.npz. It combines existing T1 and T2 segment tensors while preserving a single fly-level prediction unit.

To export a richer per-timestep motion representation, use --channel-set rich. This keeps the same fly-level training path but adds channels for acceleration, radial/tangential motion, heading, and turning:

python -m flygen_ml.cli.export_sequence_tensors \
  --segments artifacts/segments_with_cohort.csv \
  --output artifacts/sequences_rich_v1.npz \
  --target-length 128 \
  --channel-set rich

9. Train A Fly-Level Sequence Model

The first sequence model is a small NumPy baseline: flatten each segment tensor, encode it with a one-hidden-layer MLP, mean-pool segment embeddings per fly, and predict both genotype and cohort with separate softmax heads. During training, it randomly samples up to train_max_segments_per_fly segments per fly each epoch. During evaluation, eval_max_segments_per_fly: 0 means use all available segments for each fly.

python -m flygen_ml.cli.train_sequence_model \
  --config configs/model/segment_meanpool.yaml \
  --sequences artifacts/sequences_v1.npz \
  --output runs/segment_meanpool_v1_cv \
  --cv-folds 5

Sequence predictions are one row per fly, not one row per segment, and include both genotype and cohort predictions plus joint correctness.

To print a compact summary of a sequence run:

python -m flygen_ml.cli.evaluate_sequence_model \
  --run-dir runs/segment_meanpool_v1_cv

For a stronger segment encoder, install PyTorch in the active environment and train the Conv1D sequence model:

python -m flygen_ml.cli.train_sequence_model \
  --config configs/model/segment_conv1d_meanpool.yaml \
  --sequences artifacts/sequences_v1.npz \
  --output runs/segment_conv1d_meanpool_v1_cv \
  --cv-folds 5

To concatenate engineered fly-level movement summaries onto the learned trajectory embedding before the prediction heads:

python -m flygen_ml.cli.train_sequence_model \
  --config configs/model/segment_conv1d_meanpool_fused.yaml \
  --sequences artifacts/sequences_v1.npz \
  --output runs/segment_conv1d_meanpool_fused_v1_cv \
  --cv-folds 5

Use --seed to override random_seed from the config without creating a seed-specific config file:

python -m flygen_ml.cli.train_sequence_model \
  --config configs/model/segment_conv1d_headpool_fused_wide.yaml \
  --sequences artifacts/sequences_v1.npz \
  --output runs/segment_conv1d_headpool_fused_wide_v1_seed1_cv \
  --cv-folds 5 \
  --seed 1

To test short ordered chains of consecutive between-reward trajectories instead of pooling individual segment embeddings directly. The chain config prints fold/epoch progress every five epochs because these runs are heavier than the single-segment model:

python -m flygen_ml.cli.train_sequence_model \
  --config configs/model/segment_chain_conv1d_headpool_fused_wide.yaml \
  --sequences artifacts/sequences_v1.npz \
  --output runs/segment_chain_conv1d_headpool_fused_wide_v1_cv \
  --cv-folds 5

To run a longer chain experiment with validation monitoring and best-epoch restoration:

python -m flygen_ml.cli.train_sequence_model \
  --config configs/model/segment_chain_conv1d_headpool_fused_wide_long.yaml \
  --sequences artifacts/sequences_v1.npz \
  --output runs/segment_chain_conv1d_headpool_fused_wide_long_v1_cv \
  --cv-folds 5

To run an ordered-segment GRU experiment, which encodes individual trajectories with the Conv1D segment encoder and then passes the ordered segment embeddings through a GRU before fly-level pooling:

python -m flygen_ml.cli.train_sequence_model \
  --config configs/model/segment_gru_conv1d_headpool_fused_wide_long.yaml \
  --sequences artifacts/sequences_v1.npz \
  --output runs/segment_gru_conv1d_headpool_fused_wide_long_v1_cv \
  --cv-folds 5

The current GRU-128 configs use the same ordered-segment architecture with a larger recurrent hidden state. Training 1-only is the strongest current setup for genotype and joint accuracy:

python -m flygen_ml.cli.train_sequence_model \
  --config configs/model/segment_gru128_t1_conv1d_headpool_fused_wide_long.yaml \
  --sequences artifacts/sequences_t1_v1.npz \
  --output runs/segment_gru128_t1_conv1d_headpool_fused_wide_long_v1_cv \
  --cv-folds 5

Combined Training 1 + Training 2 is the strongest current setup for cohort:

python -m flygen_ml.cli.train_sequence_model \
  --config configs/model/segment_gru128_t1_t2_conv1d_headpool_fused_wide_long.yaml \
  --sequences artifacts/sequences_t1_t2_v1.npz \
  --output runs/segment_gru128_t1_t2_conv1d_headpool_fused_wide_long_v1_cv \
  --cv-folds 5

To repeat seed sweeps without creating seed-specific config files:

python -m flygen_ml.cli.train_sequence_model \
  --config configs/model/segment_gru128_t1_conv1d_headpool_fused_wide_long.yaml \
  --sequences artifacts/sequences_t1_v1.npz \
  --output runs/segment_gru128_t1_conv1d_headpool_fused_wide_long_v1_seed1_cv \
  --cv-folds 5 \
  --seed 1

Output Artifacts

A standard holdout run writes:

• run_metadata.json: paths, split sizes, labels, model kind, and active label_key
• model_artifact.json: trained weights, feature names, feature scaling values, labels, and active label_key
• metrics_summary.json: train/validation accuracy, balanced accuracy, per-label recall, support, and evidence-bin summaries
• predictions.csv: per-fly predictions with generic target fields

Prediction rows use:

• label_key
• actual_label
• predicted_label
• predicted_probability

For genotype-mode runs, the prediction output also includes actual_genotype and predicted_genotype as backward-compatible aliases.

Cross-validation writes:

• cv_metrics_summary.json
• cv_predictions.csv
• run_metadata.json

Sequence CV runs can be summarized one at a time:

python -m flygen_ml.cli.evaluate_sequence_model \
  --run-dir runs/segment_gru128_t1_conv1d_headpool_fused_wide_long_v1_cv

To summarize repeated sequence CV runs across seeds:

python -m flygen_ml.cli.summarize_sequence_runs \
  --run-dir runs/segment_gru128_t1_conv1d_headpool_fused_wide_long_v1_cv \
  --run-dir runs/segment_gru128_t1_conv1d_headpool_fused_wide_long_v1_seed1_cv \
  --run-dir runs/segment_gru128_t1_conv1d_headpool_fused_wide_long_v1_seed2_cv \
  --run-dir runs/segment_gru128_t1_conv1d_headpool_fused_wide_long_v1_seed3_cv

Use --output-json to save the across-seed summary as JSON.

Evaluate whether two independently trained fly-level classifiers jointly identify each fly:

python -m flygen_ml.cli.evaluate_joint_predictions \
  --axis-a-run runs/logreg_v1_movement_only_genotype_cv \
  --axis-b-run runs/logreg_v1_movement_only_antennae_condition_cv \
  --axis-a-name genotype \
  --axis-b-name cohort \
  --split valid \
  --join-without-fold \
  --output runs/joint_genotype_cohort_eval/joint_predictions.csv

The joint evaluator auto-detects cv_predictions.csv or predictions.csv inside run directories, or accepts explicit prediction CSVs with --axis-a-predictions and --axis-b-predictions. Rows are joined by fly_id, sample_key, split when present on both axes, and fold when present on both axes. Use --join-without-fold for out-of-fold validation predictions from CV runs whose fold assignments are not aligned across label axes.

Inspection Helpers

Rank extracted segments by an engineered metric:

python -m flygen_ml.cli.inspect_segments \
  --segments artifacts/segments_with_cohort.csv \
  --metric end_radius_px \
  --limit 20 \
  --output artifacts/top_end_radius_segments.csv

Inspect misclassified validation flies and their largest feature contributors:

python -m flygen_ml.cli.inspect_misclassifications \
  --run-dir runs/logreg_v1_movement_only_antennae_condition \
  --output runs/logreg_v1_movement_only_antennae_condition/misclassified_valid_fly_inspection.csv

Use --include-correct to inspect all rows in the selected split.

Export a compact prediction review table joined to fly-level metadata:

python -m flygen_ml.cli.inspect_predictions \
  --run-dir runs/logreg_v1_movement_only_antennae_condition_cv \
  --output runs/logreg_v1_movement_only_antennae_condition_cv/valid_error_review.csv \
  --errors-only

Add --include-features to include all feature columns from the feature table.

Export all between-reward segment rows for selected prediction-review flies:

python -m flygen_ml.cli.export_prediction_segments \
  --prediction-review runs/logreg_v1_movement_only_antennae_condition_cv/valid_error_review.csv \
  --segments artifacts/segments_with_cohort.csv \
  --output runs/logreg_v1_movement_only_antennae_condition_cv/high_confidence_error_segments.csv \
  --errors-only \
  --min-decision-margin 0.30

This creates a plotting-ready segment table with prediction metadata prepended to each segment row.

Segment Occlusion For Sequence Models

For trained PyTorch sequence holdout runs, segment occlusion estimates how much each between-reward trajectory contributes to a fly-level prediction. The workflow removes one segment at a time from each selected fly, reruns the model, and records probability/logit deltas for the genotype and cohort heads.

First train a holdout sequence model so that model_artifact.json is available:

python -m flygen_ml.cli.train_sequence_model \
  --config configs/model/segment_gru128_conv1d_headpool_fused_wide_long.yaml \
  --sequences artifacts/sequences_v1.npz \
  --output runs/segment_gru128_conv1d_headpool_fused_wide_long_v1_holdout_for_occlusion \
  --seed 0

Then run leave-one-segment-out occlusion on the validation flies:

python -m flygen_ml.cli.explain_sequence_occlusion \
  --run-dir runs/segment_gru128_conv1d_headpool_fused_wide_long_v1_holdout_for_occlusion \
  --sequence-path artifacts/sequences_v1.npz \
  --output-csv runs/segment_gru128_conv1d_headpool_fused_wide_long_v1_holdout_for_occlusion/segment_occlusion_valid.csv \
  --split valid

The output contains baseline and occluded probabilities/logits, predicted and actual class deltas, and flags indicating whether removing a segment changed the genotype, cohort, or joint prediction. Deltas are computed as:

baseline value - occluded value

Thus, a positive logit delta means the removed segment supported the tracked class; a negative delta means the segment opposed the tracked class.

To prepare visual-review subsets, filter rows by fly type, output head, sign of the logit delta, and optional correctness of the baseline prediction. For example, to review PFN>Kir antennae-intact flies where the genotype prediction was correct:

python -m flygen_ml.cli.filter_occlusion_segments \
  --occlusion-csv runs/segment_gru128_conv1d_headpool_fused_wide_long_v1_holdout_for_occlusion/segment_occlusion_valid.csv \
  --output-dir runs/segment_gru128_conv1d_headpool_fused_wide_long_v1_holdout_for_occlusion/structured_review/pfn_intact/genotype \
  --deficiency pfn-intact \
  --head genotype \
  --require-correct genotype \
  --max-segments-per-fly 3

Supported fly-type filters are:

• pfn-intact: PFN>Kir + antennae-intact
• control-removed: Control>Kir + antennae-removed
• control-intact: Control>Kir + antennae-intact

Supported correctness filters are none, genotype, cohort, and joint. The filter command writes separate positive and negative logit-delta CSVs. Positive rows support the tracked class; negative rows oppose it.

A typical structured review matrix is:

fly type	head	correctness filter	question
pfn-intact	genotype	genotype	Which segments support or oppose the PFN>Kir genotype call when antennae are intact?
control-removed	cohort	cohort	Which segments support or oppose the antennae-removed call when genotype is control?
control-intact	genotype	genotype	Which segments support or oppose the Control>Kir genotype call in intact controls?
control-intact	cohort	cohort	Which segments support or oppose the antennae-intact call in intact controls?

For quick dependency-light inspection, plot normalized tensor paths directly:

python -m flygen_ml.cli.plot_occlusion_segments \
  --occlusion-csv runs/segment_gru128_conv1d_headpool_fused_wide_long_v1_holdout_for_occlusion/structured_review/pfn_intact/genotype/pfn-intact_genotype_predicted_positive_logit_delta.csv \
  --sequence-path artifacts/sequences_v1.npz \
  --output-dir runs/segment_gru128_conv1d_headpool_fused_wide_long_v1_holdout_for_occlusion/structured_review/pfn_intact/genotype/quick_plots_positive \
  --change-head genotype \
  --include-unchanged

To plot with external trajectory-plotting tools that need original frame metadata, join the filtered occlusion rows back to the segment table:

python -m flygen_ml.cli.export_occlusion_segments \
  --occlusion-csv runs/segment_gru128_conv1d_headpool_fused_wide_long_v1_holdout_for_occlusion/structured_review/pfn_intact/genotype/pfn-intact_genotype_predicted_positive_logit_delta.csv \
  --segments artifacts/segments_with_cohort.csv \
  --output runs/segment_gru128_conv1d_headpool_fused_wide_long_v1_holdout_for_occlusion/structured_review/pfn_intact/genotype/pfn-intact_genotype_predicted_positive_plot_ready.csv

This plot-ready CSV preserves occlusion evidence columns and adds source segment metadata such as data_path, trx_path, experimental_fly_idx, anchor_reward_frame, and end_reward_frame.

Current limitation: occlusion currently explains saved holdout sequence runs. Grouped CV runs do not yet save fold-specific model artifacts, so validation flies from CV runs cannot yet be explained using the exact fold model that predicted them. For stronger systematic analysis, repeat holdout runs across seeds or extend CV training to save one model artifact per fold.

Prediction Error Comparison

Compare the error sets from two saved prediction runs:

python -m flygen_ml.cli.compare_prediction_errors \
  --run-a runs/logreg_v1_movement_only_genotype_cv \
  --run-b runs/segment_gru128_conv1d_headpool_fused_wide_long_v1_cv \
  --axis genotype \
  --run-a-name logreg \
  --run-b-name gru128 \
  --join-without-fold \
  --output runs/error_analysis/logreg_vs_gru128_genotype.csv

Summarize correctness buckets after an error comparison:

python -m flygen_ml.cli.inspect_error_buckets \
  --comparison runs/error_analysis/logreg_vs_gru128_genotype.csv \
  --features artifacts/features_antennae_no_training_end.csv \
  --output runs/error_analysis/logreg_vs_gru128_genotype_summary.json \
  --examples-output runs/error_analysis/logreg_vs_gru128_genotype_examples.csv

Modeling Notes

The baseline model automatically ignores known metadata columns, including fly_id, sample_key, genotype, cohort, chamber, chamber_type, training_idx, date, fly_idx, and prediction-label fields. Additional numeric features can be excluded in configs/model/logreg.yaml:

exclude_feature_names: n_segments,n_segments_with_qc_flags

The current split protects against fly-level leakage. It does not, by itself, guarantee independence across recording date, experimental batch, chamber, or other possible nuisance structure. For stronger claims, inspect those covariates and prefer grouped cross-validation or stricter grouping strategies where appropriate.

Sequence models also protect evaluation at the fly level: per-segment trajectories are evidence for a fly-level label, not independent labeled examples. During training, configs such as train_max_segments_per_fly: 200 randomly sample segment evidence per fly each epoch when a fly has more than the cap. During evaluation, eval_max_segments_per_fly: 0 means use all available segments for each fly.

Current interpretation notes:

• Training 1-only GRU-128 is strongest for genotype and joint accuracy.
• Training 1 + Training 2 GRU-128 is strongest for cohort accuracy.
• Adding more trajectory evidence does not automatically help every label axis; T1 and T2 appear to emphasize different behavioral signals.
• The preferred next inputs remain behavior-derived trajectory context features, rather than shortcut metadata used only to maximize classification.

Development

Run tests from the repository root:

pytest

Useful CLI entry points:

python -m flygen_ml.cli.build_manifest_from_globs --help
python -m flygen_ml.cli.extract_segments --help
python -m flygen_ml.cli.build_features --help
python -m flygen_ml.cli.train_model --help
python -m flygen_ml.cli.evaluate_model --help
python -m flygen_ml.cli.export_sequence_tensors --help
python -m flygen_ml.cli.train_sequence_model --help
python -m flygen_ml.cli.evaluate_sequence_model --help
python -m flygen_ml.cli.summarize_sequence_runs --help
python -m flygen_ml.cli.evaluate_joint_predictions --help
python -m flygen_ml.cli.compare_prediction_errors --help
python -m flygen_ml.cli.inspect_error_buckets --help
python -m flygen_ml.cli.inspect_segments --help
python -m flygen_ml.cli.inspect_misclassifications --help
python -m flygen_ml.cli.inspect_predictions --help
python -m flygen_ml.cli.export_prediction_segments --help

Reference Docs

• docs/source-data-contract.md
• docs/upstream-notes.md
• docs/implementation-spec.md
• docs/experiment-plan.md