Requests
Signals & montages
Send the trials you recorded. Wizard converts them for each method it tests; this page says what each kind of method does with your channels and samples.
What Wizard accepts
| Channels | any 10-10 or 10-20 layout, one name per row, in any order; methods requiring two channels are excluded for a single-channel recording |
| Sampling rate | any; each method resamples to its own rate |
| Trial length | any; see Windows below |
| Units | uV, mV or V, converted to microvolts |
A method that cannot read your data (your layout, a rate too low for its band, a trial much shorter than its window) is not tried, and the request is not refused for it; the create is refused only when no compatible method can read it.
What each kind of method reads
| Method | Channels it reads |
|---|---|
| Methods using the supplied layout | The channels you sent, as they are: any layout of at least two channels. Predictions must send the same names. |
| Position-aware processing | Any layout whose every channel has a 10-10, 10-20 or 10-05 name: the position is read from the name. A channel without such a name (EXG1) rules the method out, so drop such channels before sending. |
| Methods with a fixed 19-site layout | The 19 sites of the 10-20 layout, matched by name in any order; a site you did not send is filled from the nearest electrode within 60 mm, or left silent. |
| Methods with a named vocabulary | Any layout that shares at least one channel with the method's own list of names. |
Naming channels
Use 10-10 names, the current names for the 10-20 sites (Fp1, Cz, P8, …); capitalization is ignored. The legacy names T3, T4, T5 and T6 are read as T7, T8, P7 and P8. Extra channels such as EOG, EMG or triggers are dropped by the methods that read named sites.
Names are matched exactly; prefixes and suffixes are not stripped: a file label such as EEG FP1-REF matches nothing. Clean the labels first (the helper on Dataset format does). A bipolar channel such as Fp1-F7 is not an electrode against a common reference: re-reference before sending.
Windows
A pretrained encoder reads fixed-length windows. For each pretrained model, a trial at least two of the encoder’s windows long is cut into back-to-back windows and their features are averaged; stride_seconds sets the step. A shorter trial is kept whole as one window, center-cropped or zero-padded to the window length. Classical methods read their own fixed epoch, or the whole trial for sleep (the methods).
Before a network encoder, each window has its DC offset removed (each channel’s mean over the window) and is resampled to the encoder’s rate (polyphase). No band-pass or notch filter is applied beyond the resampler’s own anti-aliasing low-pass; if your analysis needs a filter, apply it yourself and report it. Resampling cannot restore content above half your original rate: a 128 Hz recording carries nothing above 64 Hz.
Units and scale
Set units to what your amplifier exports: uV, mV or V. All three are converted to microvolts. unitless, normalized and unknown are refused with 422 UNITS_REQUIRED: the encoders need a physical unit.