Upload, auto-label, and pipeline raw Halo activity-log and GPS-log data into an Edge Impulse project.
Find it in Studio under Dashboard → Keys.
AdvancedUpload a local activity log file directly, or run the automated pipeline to process all sessions from the bucket.
Parse a raw Halo activity log (and optional user_reported.txt) from your computer and upload the resulting samples straight into this project.
Runs an Edge Impulse Pipeline that (1) executes the "Activity Log Processor" transformation block directly on the mounted bucket — converting raw activity session folders (*.log + *_user_reported.txt) into per-timestep labeled CSVs matching this project's CSV Wizard config — then (2) syncs those CSVs into the target project. No file browsing or manual download required.
| Step | Duration | Status |
|---|---|---|
| No run history yet. | ||
Runs an Edge Impulse Pipeline that (1) executes the "GPS Log Processor" transformation block directly on the mounted bucket — converting raw ai-logs/*.log + app-logs-zip/*.zip into labeled feature CSVs — then (2) syncs those CSVs into the target project. No file browsing or manual download required.
| Step | Duration | Status |
|---|---|---|
| No run history yet. | ||
Retrain and build a combined multi-impulse (activity + GPS) firmware deployment, then patch it directly into your local halo-fw checkout.
Score GPS session zips (collar_collected / user_reported file pairs) against the Assessment Framework release bar. Put the zips for one test in Session A and, optionally, a second test in Session B — zips are pooled only within their own group. The top summary compares A vs B (accuracy with 95% CI, streaks, latency, scored time); full details for each group sit in collapsible sections under the Session A / Session B tabs.
Score activity session zips (collar_collected / user_reported file pairs) against a per-group release bar. Every tagged second (after a 10s grace period) is compared with what the collar was outputting, grouped as Active (Walking, RunningPlaying), Idle (Lying, Sitting, Standing) and Carry (HumanCarry) and OnShelf. Put the zips for one test in Session A and, optionally, a second test in Session B. Activity zips are cumulative, so sessions repeated across zips are counted once.
Is there enough training data? Inspect the dataset, check the training curve, and run learning-curve / hold-out experiments against this project's impulse (Raw data → Spectral features → activity classifier). Includes a per-class check for the 8-class model..
Scores the current model the way the Labeling Capacity Planner does: the dog is the unit. Accuracy = % of test windows correct per dog, averaged across dogs. The measured between-dog SD replaces the planner's placeholder, giving how many scored dogs the validation set needs, per stratum, and how many label-minutes per round.
Plots the loss / accuracy per epoch from the last training of the selected impulse. A validation loss that turns upward, or a large train-vs-validation gap, means the model is overfitting — usually a sign of too little or too homogeneous training data.
Keeps the test set fixed and retrains on growing subsets of the training data (nested; subsample whole dogs to see whether more dogs help). If test accuracy is still climbing at 100%, more data will help; if it has flattened, you have enough of this kind of data.
Removes one dog (or other metadata group) from the training set, retrains, and measures accuracy on that group's test samples. A big drop means the model needs data from more distinct groups, not just more data.
Is there enough training data? Inspect the dataset, check the training curve, and run learning-curve / hold-out experiments against this project's impulse (Raw data → Classifier).
Scores the current model the way the Labeling Capacity Planner does: the dog is the unit. Accuracy = % of test windows correct per dog, averaged across dogs. The measured between-dog SD replaces the planner's placeholder, giving how many scored dogs the validation set needs, per stratum, and how many label-minutes per round.
Plots the loss / accuracy per epoch from the last training of the selected impulse. A validation loss that turns upward, or a large train-vs-validation gap, means the model is overfitting — usually a sign of too little or too homogeneous training data.
Keeps the test set fixed and retrains on growing subsets of the training data (nested; subsample whole dogs to see whether more dogs help). If test accuracy is still climbing at 100%, more data will help; if it has flattened, you have enough of this kind of data.
Removes one dog (or other metadata group) from the training set, retrains, and measures accuracy on that group's test samples. A big drop means the model needs data from more distinct groups, not just more data.