Metrics
A metric is a series of values over steps, such as a loss per epoch. Each run keeps its own series.
Add a metric
import hectiqlab as hl
hl.add_metric("loss", step=epoch, value=loss)| Name | Type | Description |
|---|---|---|
name | str | The series. |
step | float | The step. |
value | float | The value. |
add_metric returns at once. Samples are batched and sent in the background, so it's safe in a tight loop. It needs an open run.
The API keeps at most 500 samples per series, thinned.
Read metrics
loss = hl.metrics.get("loss", 42).unwrap()
every = hl.metrics.list(42).unwrap()Each series carries its samples, ordered by step. hl.metrics.list(runs=[...]) reads the series of several runs, by id, in one request.
Describe or pin a series
A description and a pin apply to the series on every run of the project. A pinned series shows first in the web app.
hl.metrics.update("loss", description="Cross-entropy on the training set", pinned=True)