Training experiment

As explained in Designing Deep Learning Systems: A guide for software engineers

A grouping construct that organizes multiple related machine learning training runs.

A training experiment groups related machine learning training runs around a shared development goal. One experiment can contain multiple runs, enabling execution history to be organized and compared in a metadata-store interface. section 8.3.2

Use the experiment as the organizing layer around each run’s operational evidence:

• Associate every training run with its experiment so related executions can be grouped in the UI. section 8.3.2
• Retain parameters such as dataset ID, dataset version, and hyperparameters alongside metrics such as F2 score. section 8.3.2
• Link models, input data, metrics, and artifact URLs through the shared training_run_id. section 8.3.2

Intent classification development

Define intent classification model development as one training experiment, then associate each intent classification training execution with it. section 8.3.2

Figure 8.5 An entity relationship diagram of the data schema of the metadata storage

Inspect the entity relationship diagram to see how experiments, training runs, parameters, metrics, models, and artifacts connect. section 8.3.2

How Designing Deep Learning Systems: A guide for software engineers covers it

How deep it goes

Defines experiment-to-run grouping and connects runs to metadata and artifacts, without detailing experiment lifecycle operations.

Examples

Uses an intent classification development project containing multiple intent classification training executions.

What you can do after

Evaluate related training executions together and trace their inputs, metrics, models, and artifact URLs.

How it connects

Connects training runs with parameters, metrics, models, datasets, metadata stores, and model artifacts.

liveBookmanning.com

Read Designing Deep Learning Systems: A guide for software engineers on liveBook Get the book at manning.com

Related concepts

• runuses · A run is a single execution of a process, test, workflow, or operation, often associated with its inputs, configuration, outputs, and results.

This explainer was generated by Manning's knowledge graph from the text of Designing Deep Learning Systems: A guide for software engineers. The section links point to the parts of the book it draws on.