liveBook · Manning

Metadata

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

Data that describes an artifact, experiment, execution, or relationship between machine learning artifacts.

Metadata is structured reference data about other data or objects. In deep learning systems, it describes training runs, workflows, models, datasets, and other artifacts. Unlike service logs and metrics used for operational monitoring, it records machine learning activity so teams can compare experiments, reproduce models, and investigate degradation. section 8.2 section 8.0

A metadata and artifact store separates descriptive records from raw files. Datasets, code, processed data, and model files remain in file or cloud storage; the metadata store retains their URLs, identifiers, versions, execution IDs, facts, and lineage. Searching metadata can then reveal a model's related training inputs and outputs. section 8.1 section 8.2.1

Useful metadata categories include:

  • Training-run metadata: dataset identity and version, hyperparameters, allocated and consumed hardware, code version and environment configuration, training metrics, and evaluation metrics. section 8.2.1
  • General-artifact metadata: file location, version, description, and audit history. section 8.2.1
  • Model-file and orchestration-workflow metadata: records that identify artifacts and connect them to the executions and workflows that produced or use them. section 8.2.1

Metadata-driven model serving

A prediction request supplies a model ID. The frontend looks up metadata containing the model version, algorithm type, and file URL; it selects a compatible predictor backend and downloads the model file to shared storage when needed. Metadata therefore connects a logical model request to its deployable artifact and serving implementation. section 7.1.2 section 7.1.3

For a complete metadata and artifact store, MLflow adds automatic logging, experiment visualization and comparison, a model registry, artifact management, reproducibility support, and deployment-oriented MLOps features. ML Metadata (MLMD) is a lightweight alternative for ingesting and searching run-based metadata within an existing artifact registry and metric-visualization system. section 8.4.3

Figure 8.1 Artifacts are associated with their metadata in the metadata store.

This layout shows raw artifacts in file storage associated with searchable metadata in a separate metadata store. section 8.1

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

How deep it goes
Defines metadata, details training-run and artifact fields, and explains a separate artifact-store architecture without prescribing one schema.
Examples
Uses food nutrition labels, cloud file URLs, an intent-classification serving request, and MLflow-versus-MLMD deployment choices.
What it assumes
Introduces metadata and artifacts from zero while assuming familiarity with deep learning projects, training, datasets, and models.
What you can do after
Operate/configure metadata storage to retrieve artifacts, compare training experiments, reproduce models, and trace model-performance problems.
How it connects
Connects artifacts, model training runs, workflows, datasets, model serving, lineage, hyperparameters, and training metrics.
Cover of Designing Deep Learning Systems: A guide for software engineers
liveBookmanning.com
Read Designing Deep Learning Systems: A guide for software engineers on liveBook Get the book at manning.com

Also explained in

Compare how 21 books explain metadata →

Related concepts

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.