Neptune Query API

The neptune_query package is a read-only API for fetching metadata.

With the Query API, you can:

• List experiments, runs, and attributes of a project.
• Fetch experiment or run metadata as a data frame.
• Define filters to fetch experiments, runs, and attributes that meet certain criteria.

Installation

pip install "neptune-query<2.0.0"

Set your Neptune API token and project name as environment variables:

export NEPTUNE_API_TOKEN="ApiTokenFromYourNeptuneProfile"

export NEPTUNE_PROJECT="workspace-name/project-name"

Note: You can also pass the project path to the project argument of any querying function.

Usage

import neptune_query as nq

Available functions:

• download_files() – download files from the specified experiments.
• fetch_experiments_table() – runs as rows and attributes as columns.
• (experimental) fetch_experiments_table_global() – like fetch_experiments_table(), but searches across all projects that the user has access to.
• (experimental) fetch_metric_buckets() – get summary values split by X-axis buckets.
• fetch_metrics() – series of float or int values, with steps as rows.
• (runs API) fetch_runs_table() – like fetch_experiments_table(), but for individual runs.
• fetch_series() – for series of strings or histograms.
• list_attributes() – all logged attributes of the target project's experiment runs.
• list_experiments() – names of experiments in the target project.
• set_api_token() – set the Neptune API token to use for the session.

For details, see the API reference.

Runs API

You can target individual runs by ID instead of experiment runs by name.

To use the corresponding methods for runs, import the runs module:

import neptune_query.runs as nq_runs

nq_runs.fetch_metrics(...)

You can use these methods to target individual runs by ID instead of experiment runs by name.

Documentation

For how-tos and the complete API reference, see the docs directory.

Examples

The following are some examples of how to use the Query API. For all functions and options, see the API reference.

Example 1: Fetch metric values

To fetch values at each step, use fetch_metrics().

• To filter experiments to return, use the experiments parameter.
• To specify attributes to include as columns, use the attributes parameter.
• To limit the returned values, use the available parameters.

nq.fetch_metrics(
    experiments=["exp_dczjz"],
    attributes=r"metrics/val_.+_estimated$",
    tail_limit=10,
)

                  metrics/val_accuracy_estimated  metrics/val_loss_estimated
experiment  step
exp_dczjz    1.0                        0.432187                    0.823375
             2.0                        0.649685                    0.971732
             3.0                        0.760142                    0.154741
             4.0                        0.719508                    0.504652

Example 2: Fetch metadata as one row per run

To fetch experiment metadata from your project, use the fetch_experiments_table() function. The output mimics the runs table in the web app.

• To filter experiments to return, use the experiments parameter.
• To specify attributes to include as columns, use the attributes parameter.

nq.fetch_experiments_table(
    experiments=r"^exp_",
    attributes=["metrics/train_accuracy", "metrics/train_loss", "learning_rate"],
)

            metrics/train_accuracy   metrics/train_loss   learning_rate
experiment
exp_ergwq                 0.278149             0.336344            0.01
exp_qgguv                 0.160260             0.790268            0.02
exp_hstrj                 0.365521             0.459901            0.01

For series attributes, the value of the last logged step is returned.

Example 3: Define filters

List my experiments that have a "dataset_version" attribute and "val/loss" less than 0.1:

from neptune_query.filters import Filter


owned_by_me = Filter.eq("sys/owner", "sigurd")
dataset_check = Filter.exists("dataset_version")
loss_filter = Filter.lt("val/loss", 0.1)

interesting = owned_by_me & dataset_check & loss_filter
nq.list_experiments(experiments=interesting)

['exp_ergwq', 'exp_qgguv', 'exp_hstrj']

Then fetch configs from the experiments, including also the interesting metric:

nq.fetch_experiments_table(
    experiments=interesting,
    attributes=r"config/ | val/loss",
)

            config/optimizer  config/batch_size  config/learning_rate   val/loss
experiment
exp_ergwq               Adam                 32                 0.001     0.0901
exp_qgguv           Adadelta                 32                 0.002     0.0876
exp_hstrj           Adadelta                 64                 0.001     0.0891

Example 4: Exclude archived runs

To exclude archived experiments or runs from the results, create a filter on the sys/archived attribute:

import neptune_query as nq
from neptune_query.filters import Filter


exclude_archived = Filter.eq("sys/archived", False)
nq.fetch_experiments_table(experiments=exclude_archived)

To use this filter in combination with other criteria, use the & operator to join multiple filters:

name_matches_regex = Filter.name(r"^exp_")
exclude_archived = Filter.eq("sys/archived", False)

nq.fetch_experiments_table(experiments=name_matches_regex & exclude_archived)

Example 5: Fetch runs belonging to specific experiment

Each run's experiment information is stored in the sys/experiment namespace.

To query runs belonging to a specific experiment, use the runs API and construct a filter on the sys/experiment/name attribute:

import neptune_query.runs as nq_runs
from neptune_query.filters import Filter


experiment_name_filter = Filter.eq("sys/experiment/name", "kittiwake-week-1")
nq_runs.list_runs(runs=experiment_name_filter)

---

License

This project is licensed under the Apache License Version 2.0. For details, see Apache License Version 2.0.