Neptune example#

Neptune is the MLOps stack component for experiment tracking. It offers a single place to log, compare, store, and collaborate on experiments and models. This plugin enables seamless use of Neptune within Flyte by configuring links between the two platforms. In this example, we learn how to train scale up training multiple XGBoost models and use Neptune for tracking.

from typing import List, Tuple

import numpy as np
from flytekit import (
    ImageSpec,
    Resources,
    Secret,
    current_context,
    dynamic,
    task,
    workflow,
)
from flytekitplugins.neptune import neptune_init_run

First, we specify the Neptune project that was created on Neptune’s platform. Please update NEPTUNE_PROJECT to the value associated with your account.

NEPTUNE_PROJECT = "username/project"

Neptune requires an API key to authenticate with their service. In the above example, the secret is created using Flyte’s Secrets manager.

api_key = Secret(key="neptune-api-token", group="neptune-api-group")

Next, we use ImageSpec to construct a container with the dependencies for our XGBoost training task. Please set the REGISTRY to a registry that your cluster can access;

REGISTRY = "localhost:30000"

image = ImageSpec(
    name="flytekit-xgboost",
    packages=[
        "neptune",
        "neptune-xgboost",
        "flytekitplugins-neptune",
        "scikit-learn==1.5.1",
        "numpy==1.26.1",
        "matplotlib==3.9.2",
    ],
    builder="default",
    registry=REGISTRY,
)

First, we use a task to download the dataset and cache the data in Flyte:

@task(
    container_image=image,
    cache=True,
    cache_version="v2",
    requests=Resources(cpu="2", mem="2Gi"),
)
def get_dataset() -> Tuple[np.ndarray, np.ndarray]:
    from sklearn.datasets import fetch_california_housing

    X, y = fetch_california_housing(return_X_y=True, as_frame=False)
    return X, y

Next, we use the neptune_init_run decorator to configure Flyte to train an XGBoost model. The decorator requires an api_key secret to authenticate with Neptune and the task definition needs to request the same api_key secret. In the training function, the Neptune run object is accessible through current_context().neptune_run, which is frequently used in Neptune’s integrations. In this example, we pass the Run object into Neptune’s XGBoost callback.

@task(
    container_image=image,
    secret_requests=[api_key],
    requests=Resources(cpu="2", mem="4Gi"),
)
@neptune_init_run(project=NEPTUNE_PROJECT, secret=api_key)
def train_model(max_depth: int, X: np.ndarray, y: np.ndarray):
    import xgboost as xgb
    from neptune.integrations.xgboost import NeptuneCallback
    from sklearn.model_selection import train_test_split

    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=123)
    dtrain = xgb.DMatrix(X_train, label=y_train)
    dval = xgb.DMatrix(X_test, label=y_test)

    ctx = current_context()
    run = ctx.neptune_run
    neptune_callback = NeptuneCallback(run=run)

    model_params = {
        "tree_method": "hist",
        "eta": 0.7,
        "gamma": 0.001,
        "max_depth": max_depth,
        "objective": "reg:squarederror",
        "eval_metric": ["mae", "rmse"],
    }
    evals = [(dtrain, "train"), (dval, "valid")]

    # Train the model and log metadata to the run in Neptune
    xgb.train(
        params=model_params,
        dtrain=dtrain,
        num_boost_round=57,
        evals=evals,
        callbacks=[
            neptune_callback,
            xgb.callback.LearningRateScheduler(lambda epoch: 0.99**epoch),
            xgb.callback.EarlyStopping(rounds=30),
        ],
    )

With Flyte’s dynamic workflows, we can scale up multiple training jobs with different max_depths:

@dynamic(container_image=image)
def train_multiple_models(max_depths: List[int], X: np.ndarray, y: np.ndarray):
    for max_depth in max_depths:
        train_model(max_depth=max_depth, X=X, y=y)


@workflow
def train_wf(max_depths: List[int] = [2, 4, 10]):
    X, y = get_dataset()
    train_multiple_models(max_depths=max_depths, X=X, y=y)

To run this workflow on a remote Flyte cluster run:

union run --remote neptune_example.py train_wf

To enable dynamic log links, add plugin to Flyte’s configuration file:

plugins:
  logs:
    dynamic-log-links:
      - neptune-run-id:
          displayName: Neptune
          templateUris: "{{ .taskConfig.host }}/{{ .taskConfig.project }}?query=(%60flyte%2Fexecution_id%60%3Astring%20%3D%20%22{{ .executionName }}-{{ .nodeId }}-{{ .taskRetryAttempt }}%22)&lbViewUnpacked=true"