Jupyter integration
Execute Aner notebooks with a registered native kernel, persistent sessions, development observations, and explicit reset and error contracts.
Aner executes .ipynb notebooks through a registered Aner kernel and the native interpreter. This reference describes a configured notebook environment. Kernel registration and interpreter installation are separate deployment prerequisites.
Requirements
- A notebook application with a registered Aner kernel.
- An updated native
anerexecutable supporting persistent sessions and development observations. - An accessible working directory containing any input files referenced by the notebook.
The kernel configuration identifies the native executable. Execution uses the configured path; if no path is configured, the integration checks ANER_EXECUTABLE and then the aner command on PATH. This reference does not imply that the kernel is included in every binary package.
Execution
Select Aner from the notebook application's kernel selector. Open the development notebook or the neural research notebook. Code cells contain Aner source; Markdown cells contain authored document text.
Imports, functions, classes, and typed variables persist between successful cells. Restart or reset the kernel before redefining a retained function or class. A static error preserves earlier state. A runtime error or transport failure clears native state; rerun the required setup cells.
Runtime modes
Place control directives in separate cells:
| Directive | Behavior |
|---|---|
%aner dev | Enable bounded development observations; reset session state |
%aner dev values | Enable observations with bounded individual values; reset state |
%aner prod | Disable automatic recording; reset session state |
%aner timeout 600 | Set a 600 second limit for subsequent cells; retain session variables |
%aner reset | Clear native session state |
%aner variables | Display the most recent development observation without executing user code |
Production is the default. Mode changes and reset clear variables. These directives configure observation policy, not neural training/evaluation behavior.
Development observations
Completed cells expose read only variable tables, object references, Tensor placement, training loss curves, operation topology, and measured training progress when available. Wide tables and diagrams scroll within their panels. Each fit has a separate loss series; phase labels distinguish losses before an update from final trained loss. Mini batch observations additionally chart row weighted online loss, held out validation loss, and validation accuracy when supplied. These measurements appear after the completed cell.
The transport buffers output until a cell completes. Streaming progress, source breakpoints, live variable editing, and pause/step execution are not implemented. The standalone training report provides detailed dependencies, neurons, and available gradients; see development mode and visual training reports.
Limits and failure behavior
Each source request is limited to 1 MiB. The transport timeout defaults to 120 seconds and includes startup, request writing, and execution. Before a longer cell, run %aner timeout 600 in a separate cell. The argument must be an integer from 1 through 86,400 seconds. Changing the limit preserves current variables and affects subsequent cells; a running request keeps its original deadline. Interrupting a running request terminates its owned native session and clears variables. Observations have independent bounded payloads and report omitted values or timing events.
Notebook documents retain code, Markdown, metadata, and saved output. Saved output is historical document content; it is not a fitted model checkpoint. Distributed examples contain no saved execution output. .anernb is the VS Code notebook suffix; .ipynb is the notebook format used by this integration.