DOCSEARCHbyAlgolia
Submit Docs

Neptune Documentation

website·Indexed 40 days ago

https://docs.neptune.ai

Documentation for Neptune, a platform for tracking model-training experiments, logging metadata, and visualizing results. Includes guides on using the neptune-scale and neptune-query Python packages to log configs, histograms, and files, as well as managing runs, creating dashboards, and querying metadata via the Neptune Query API.

Tokens
238.8K
Snippets
979
Records
1.9K
Agent score
65%

What's inside Neptune

  1. Overview of Neptune experiment tracking

    Neptune is a scalable experiment tracker designed specifically for monitoring and debugging foundation model training. It allows developers to log metadata, visualize results, and collaborate with teams through a combination of Python APIs and a web-based application.

  2. Overview of Neptune experiment tracking

    Neptune is a scalable experiment tracker designed specifically for monitoring and debugging foundation model training. It allows developers to log metadata, visualize metrics, and collaborate on model training results via a web application.

  3. Overview of Neptune Dashboards

    A dashboard in Neptune is a customizable display template used to analyze runs within a single project. It allows users to visualize metadata at a glance, such as overlaid metrics and loss for training/validation on the same chart, model parameters, and scatter plots showing relationships between configs and metrics.

    Dashboard vs. Report:

    • Dashboards are ideal for analyzing ongoing experiments within one project. The configuration applies to whichever runs are currently visible to the viewer.
    • Reports are ideal for analyzing final results, allowing comparisons across multiple projects, pre-selecting visible runs for viewers, and adding comments or external resources.

  4. Overview of Neptune integration for PyTorch Lightning

    The Neptune integration for PyTorch Lightning allows developers to monitor training in real-time, log hyperparameters, model checkpoint paths, and metrics, and visualize metadata within the Neptune app. This integration is built on top of pytorch-lightning==2.5.1rc2 and is not part of the official PyTorch Lightning release.

  5. Overview of neptune-query Filter methods

    The neptune-query package provides a set of Filter methods to specify criteria for experiments or attributes. These filters can be negated or joined with logical operators. When comparing series attributes, the last logged value is used.
    MethodDescriptionExample
    name()Run/experiment name matches regex or listFilter.name(["name1"])
    eq()Attribute value equalsFilter.eq("lr", 0.001)
    ne()Attribute value doesn't equalFilter.ne("sys/owner", "bot")
    gt()Attribute value is greater thanFilter.gt("acc", 0.9)
    ge()Attribute value is greater than or equal toFilter.ge("acc", 0.93)
    lt()Attribute value is less thanFilter.lt("loss", 0.1)
    le()Attribute value is less than or equal toFilter.le("loss", 0.11)
    matches()String attribute matches regexFilter.matches("opt", r"^Ada")
    contains_all()StringSet/String contains all specified substringsFilter.contains_all("tags", ["v1", "v2"])
    contains_none()StringSet/String contains none of the specified substringsFilter.contains_none("tok", "bpe")
    exists()Attribute existsFilter.exists("metric7")

  6. Overview of Neptune custom views

    Custom views allow you to save specific runs table configurations—including applied query filters, column configurations, and grouping—to organize and analyze runs without clutter. Custom views are available across the All runs, Charts, Dashboards, and Side-by-side tabs.

  7. Overview of Neptune attribute types and logging methods

    An attribute is the location of a piece of metadata in a Neptune run object. The resulting attribute type is determined by the metadata type and the logging method used. The type determines which operations are available for that attribute.
    Metadata typeExampleLogging methodResulting attribute type
    Single valueParameters, final scores, text, timelog_configs()Float, Integer, Boolean, String, or Datetime
    Series of valuesMetrics, loss, accuracy, text logslog_metrics()FloatSeries
    Series of stringsText logs, noteslog_string_series()StringSeries
    Single fileImage, plot file, data sampleassign_files()File
    Series of filesImage series, predictions, video/audio, JSON, CSVlog_files()FileSeries
    Series of histogramsActivations, distributionslog_histograms()HistogramSeries
    TagsText tags for runs or groupsadd_tags()StringSet

  8. Overview of Neptune experiment tracking

    Neptune is a scalable experiment tracker designed for monitoring and debugging foundation model training. It allows developers to log metadata, visualize results via a web app, and collaborate with teams through shared views, reports, and role-based access control.

  9. Overview of Neptune integration for PyTorch Lightning

    The Neptune integration for PyTorch Lightning allows developers to monitor training in real-time, log hyperparameters, track model checkpoint paths, and record metrics. These logged metadata are then visualized within the Neptune app. This integration is built on top of pytorch-lightning==2.5.1rc2 and is not part of the official PyTorch Lightning release.

  10. Overview of Neptune attribute types and logging methods

    In Neptune, an attribute is the location of metadata within a run object. The resulting attribute type is determined by the metadata type and the specific logging method used. The attribute type dictates what operations are available for that metadata.
    Metadata typeExampleLogging methodResulting attribute type
    Single valueParameters, final scores, text, timelog_configs()Float, Integer, Boolean, String, or Datetime
    Series of valuesMetrics, loss, accuracy, text logslog_metrics()FloatSeries
    Series of stringsText logs, noteslog_string_series()StringSeries
    Single fileImage, plot file, data sampleassign_files()File
    Series of filesImage series, predictions, video/audio, JSON, CSVlog_files()FileSeries
    Series of histogramsActivations, distributionslog_histograms()HistogramSeries
    TagsText tags for runs or groupsadd_tags()StringSet

  11. Understand Neptune self-hosted architecture requirements

    Neptune is deployed on a Kubernetes cluster. The architecture is decomposable, allowing components to run within a single cluster or integrate with external managed services (e.g., Kafka). A functional deployment requires the following database instances, which can either be provided by the user or managed as part of the deployment: ClickHouse, MySQL, and Redis.

  12. Organize Neptune metadata using namespaces and attributes

    In a Neptune Run, metadata is organized into a folder-like structure using attributes and namespaces. An attribute is a specific location for storing a piece of metadata, and attributes can be grouped under namespaces, which can be nested to create a hierarchy.