Querying model metadata
For revisiting past experiment runs, the Verta API offers the ability to find runs based on their properties and logged values.
To start, you can obtain a collection of experiment runs under a project or an experiment using their expt_runs
attribute:
These are ExperimentRuns
objects, which have a find()
method for filtering their contents:
Multiple queries can be specified together, which returns the intersection of their results.
Query syntax
Queries are Python strings that follow this syntax:
For example, lets say I have several runs, and I've been logging each multi-layer perceptron's hidden layer size as a hyperparameter with the key "hidden_size".
To find all runs with a hidden size greater than 512, the query would look like this:
If the value is a string, it should be surrounded by quotes. As examples:
Fields
The fields currently supported by the client are:
id
project_id
experiment_id
name
date_created
attributes
hyperparameters
metrics
date_created
is represented as a Unix timestamp, in milliseconds.
Some fields inherently do not have keys, such as id
and name
, in which case their queries are even more straightforward:
For example:
Operators
The operators currently supported by the client are:
==
!=
>
>=
<
<=
For string values, it is recommended to only use ==
and !=
.
Last updated