Skip to content
GitHub · 2.7.0

Concepts

File names

Generated take names and model autosaves share one pattern. They identify the participant and contents in Finder and Explorer without opening the files.

Text
data/<ID>/recordings/ <ID>_<YYYYMMDD-HHMM>_<what>.mg.zip
data/<ID>/predictions/ <ID>_<YYYYMMDD-HHMM>_<what>.mg.zip
data/<ID>/models/ <ID>_<YYYYMMDD-HHMM>_<model>.joblib

The folder says what kind of file it is, so the name never repeats it. The participant ID stays in the name, so a file copied out of data/ still says whose it is.

The parts

Part Example Rule
<ID> P-023 ASCII letters, digits, - and _; other characters become dashes, repeated dashes collapse, and the result is trimmed to 40 characters with edge dashes/underscores removed. Left out when no usable ID remains.
<YYYYMMDD-HHMM> 20261105-1533 Local time the take started, to the minute. meta.json’s created keeps it to the second.
<what> F, R+F, follow-F What happened in the take, see below.
<model> spar-mlp The model’s type, lowercase.

_ separates the parts, - joins the words inside one. A name taken already, by another take of the same minute with the same <what>, gets -2, -3, … at the end. Nothing is ever overwritten.

StatusBar shows the participant folder using the same safe ID, before the kind subfolder: P 001 becomes data/P-001. If no usable ID remains, it shows the data root.

<what>

Take <what> Example
Calibration, one gesture its code P-023_20261105-1533_F.mg.zip
Calibration, several gestures their codes joined by + P-023_20261103-1035_Re+T+I+M+Ri+L.mg.zip
More than 40 characters of codes the first three and a count …_R+F+P+6-more.mg.zip
Free prediction free, or free- and the labelled gesture …_free-F.mg.zip
Guided run guided …_guided.mg.zip
Follow run follow- and the gesture it followed …_follow-F.mg.zip
Named on Stop the codes, then the name …_R+F_practice.mg.zip

A take in recordings/ names the classes with recorded time, in the app’s class order. An unnamed calibration take with no recorded class time omits the _<what> suffix. A take in predictions/ is a free run unless its Session.name says otherwise. FollowRun names its take follow-<code>, and an app sets guided for a guided run.

Gesture codes

A code is the first letter of the gesture, capitalized. A gesture of several words takes the first letter of each, so Three-Finger-Pinch becomes Tfp. A number stays whole, so Pose 10 becomes P10. If codes match when ignoring case, both take more letters of their first word until they differ. Rest and Ring become Re and Ri.

Python
from myogestic.session import class_codes
print(class_codes(["Rest", "Thumb", "Index", "Middle", "Ring", "Little"]))
# {'Rest': 'Re', 'Thumb': 'T', 'Index': 'I', 'Middle': 'M', 'Ring': 'Ri', 'Little': 'L'}
print(class_codes(["Rest", "Fist", "Pinch", "Three-Finger-Pinch"]))
# {'Rest': 'R', 'Fist': 'F', 'Pinch': 'P', 'Three-Finger-Pinch': 'Tfp'}

Codes are worked out over the app’s whole class list, so a gesture keeps its code in every take of that app. Adding Ring to an app with Rest and Rest moves from R to Re. Once an app has data, pin its codes next to its class list:

Python
from myogestic import App
app = App("Bracelet")
app.ctx.class_names = ["Rest", "Fist", "Pinch", "Ring"]
app.ctx.class_codes = {"Rest": "R"} # the rest are worked out around it: Ring → Ri

Every take saves the whole map as meta.json’s class_codes, so any file decodes later, whatever the app’s codes are by then.

Models

Pipeline.autosave_dir saves each trained model as <ID>_<YYYYMMDD-HHMM>_<model>.joblib, where <model> is pipeline.model_name ("spar-mlp") or, unset, the trained object’s class (MLPDirectV3 → mlp-direct-v3). An app that saves models itself gets the same name from model_path(root, participant, model).

Older files

open_session_store accepts any .zip with meta.json and its declared stream arrays, including P-023_2026-11-05_15-33-30_Fist.session.zip. Match takes with *.zip and models with *.joblib. Never match a fixed naming suffix. Files on disk keep their names.