Experiments
The Experiments page is where runs logged with the neural_studio SDK
appear: every ns.init(...) in a notebook, a script on your laptop or a
cluster creates a run in the workspace, and run.log(...) streams its
metrics here while it trains. How to log is on the SDK page;
this page is about what you can do with the runs once they exist.
The runs list
The sidebar on the left lists the workspace's runs. It can be collapsed (Cmd/Ctrl+.) or expanded into a full table.
- Search matches names and tags. The status filter offers All statuses, Running, Completed, Failed, Stopped and Disconnected.
- Filter runs builds conditions on config keys and summary values (fields are sampled from the newest 1,000 runs). With no filter, every run in the workspace is listed.
- Group runs by a config key; grouped runs take their group's colour on every chart, and a band shows the spread within a group.
- Metric columns picks which config keys and summary metrics the table shows.
- Check several runs for the bulk bar: add or remove tags, archive or unarchive, delete. Only a run's creator or the workspace owner can delete it.
Each run's menu offers Open, Edit (name, tags and notes), Archive (hidden from the default list, with its page, shares and links kept; Unarchive brings it back), Copy run ID and Delete.
Which runs are checked decides which runs the dashboard shows.
The dashboard
The dashboard is yours: every member keeps their own, and it saves itself as you change it (undo with Cmd/Ctrl+Z). It is made of sections of panels.
Add panels… offers:
- a line panel per logged metric, over the visible runs;
- a pattern panel: a regular expression such as
loss$or^val/, filled in as matching metrics are logged; - a bar panel: one bar per run from a metric's Last, Lowest or Highest value;
- a run comparer: the visible runs' config and summary side by side;
- a text panel in Markdown;
- a scatter of two fields across runs, parallel coordinates and parameter importance for hyperparameter sweeps;
- a media panel for one media key (images, a table or a histogram).
The toolbar sets the X axis (Step, Time, or Epoch from each run's logged epoch metric) and Smoothing (No smoothing, EMA, Running average or Gaussian) for every chart at once. A panel's menu has Panel settings…, Rename…, Duplicate, Move to section, Full screen and Download PNG. Charts update live while a run is logging.
Saved views
Saved views keeps named snapshots of the whole setup: filters, grouping, columns and panels. Save view stores the current one for you; an editor can share a view with the whole workspace. Opening a shared view and then Make this your own workspace copies it into your dashboard.
A run's page
Opening a run shows its name, status, tags and notes, a Share button, and these tabs:
| Tab | What it shows |
|---|---|
| Metrics | Every logged metric as a chart, with the same axis and smoothing controls. |
| Media | Images, tables and histograms logged with run.log, one key at a time, with a step slider and Follow the latest step while the run is live. A step shows its first 64 items. |
| Layers | Per-layer gradient and weight statistics from run.watch(model). |
| System | CPU, RAM and per-GPU utilisation, memory, temperature and power over the run. |
| Logs | Everything the process printed, captured by the SDK. |
| Artifacts | Files attached with run.log_artifact, and attention maps from run.log_attention, by name and step. |
| Config | The run's configuration, including updates made during the run, and the environment and git snapshot the SDK captured. |
A run's status is Running while the SDK heartbeats, Completed or
Failed when run.finish says so (an uncaught exception finishes the run
as failed), Stopped when the code says finish(status="stopped"), and
Disconnected when the process went away without finishing.
Sharing a run
Share on the run page gives a named person read access to that one run, or creates a public link anyone can open without an account. Recipients see the same tabs; they can never edit the run or see the rest of the workspace. See Sharing.
Related
- Python SDK:
ns.init,run.log, media, artifacts and offline runs. - Limits: media and artifact caps.
- Sharing: grants and public links.