Pipeline overview

stride tmaze and stride gait add no computation of their own: every step calls the same stage function the rest of the package uses, so the tables are the same whichever way the stages are started.

flowchart TB
  subgraph prep [Scale and inference]
    A[Distortion check: report only] --> S[Scale px/cm: given value, ChArUco board, or ROI corridor width]
    S --> B[Pose inference: 15 keypoints, every frame]
    S --> C[ROI inference: best frame of the 23-node maze skeleton]
    C --> D[7 region polygons; consensus ROI when a corner is hidden]
  end
  subgraph tmaze [Decisions]
    B --> E[Body-hull coverage per region per frame]
    D --> E
    E --> F[Gate, arm entry, correctness, probes, junction exploration]
    F --> R[Decision rules: decisions_clean.csv]
  end
  subgraph gait [Gait]
    B --> G[Preprocess: fill, smooth, speeds]
    G --> H[Swing detection and stride segmentation]
    H --> I[Measures per stride and per step]
    I --> J[Filters: confidence, edges, turning, forward direction]
  end
  subgraph stats [Optional statistics]
    F --> K[Tag strides with trial phase]
    J --> K
    K --> L[Mixed models M1 and M2]
    L --> M[WT vs Tau classifier]
  end

Steps of stride tmaze

step	module	reads	writes
distortion_check	distortion_check	frames of 6 sampled videos	scale/distortion_check.csv
calibrate	scale_calibration	the ChArUco board in view	scale/px_per_cm.txt, readings, QC images
pose_inference	pose_inference	video, pose model	<video>.slp
roi_inference	roi_inference, roi_selection, slp_to_yaml	video, ROI model	<video>.preds.v2.best1.slp, <video>.rois.yml
roi_consensus	roi_consensus	the ROI files of the folder	<video>.rois.yml for clips without one, QC overlays
decisions	decision_analysis	pose file, ROI file, metadata	decisions.csv, events.csv, metrics.csv
decision_rules	decision_rules	decisions.csv	decisions_clean.csv
gait	gait_extraction, run_gait_extraction	pose file, scale	the gait tables
mixed_models, classifier (off by default)	gait_lmm, classification	the tables above, stats_config	stats/

stride gait runs calibrate, pose_inference, gait_extraction, keypoint_confidence and stride_filtering, with the same modules and no ROI.

The numbered stages of run_all.py

The stage-by-stage batch runner uses the stage numbers of the original study.

stage	module	reads	writes
1 to 2	run_inference	videos, models	distortion report, .slp, ROI region files
3	run_tmaze	pose, ROI, meta_trials_csv	decisions.csv, events.csv, metrics.csv, learning cycles
4 to 6	run_gait_extraction	pose, ROI or px_per_cm_override	gait tables, confidence table, filtered strides
7 to 9	gait_lmm	filtered strides, decisions	lmm_results_all.csv, lmm_fit_log.csv, heatmaps
10	classification	per-mouse features	feature ranking, AUC table, ROC curves

The two inference steps need the sleap-nn runtime and, in practice, a GPU; everything after reads files and runs on any machine.