Neptune + XGBoost integration

Experiment tracking, model registry, data versioning, and live model monitoring for XGBoost trained models.

What will you get with this integration?

• Log, display, organize, and compare ML experiments in a single place
• Version, store, manage, and query trained models, and model building metadata
• Record and monitor model training, evaluation, or production runs live

What will be logged to Neptune?

• metrics,
• parameters,
• learning rate,
• pickled model,
• visualizations (feature importance chart and tree visualizations),
• hardware consumption (CPU, GPU, Memory),
• stdout and stderr logs,
• training code and Git commit information,
• other metadata

Example dashboard with train-valid metrics and selected parameters

Resources

• Documentation
• Code example on GitHub
• Example of a run logged in the Neptune app
• Run example in Google Colab

Example

On the command line:

pip install xgboost>=1.3.0 neptune-xgboost

In Python:

import neptune
import xgboost as xgb
from neptune.integrations.xgboost import NeptuneCallback

# Start a run
run = neptune.init_run(
    project="common/xgboost-integration",
    api_token=neptune.ANONYMOUS_API_TOKEN,
)

# Create a NeptuneCallback instance
neptune_callback = NeptuneCallback(run=run, log_tree=[0, 1, 2, 3])

# Prepare datasets
...
data_train = xgb.DMatrix(X_train, label=y_train)

# Define model parameters
model_params = {
    "eta": 0.7,
    "gamma": 0.001,
    "max_depth": 9,
    ...
}

# Train the model and log metadata to the run in Neptune
xgb.train(
    params=model_params,
    dtrain=data_train,
    callbacks=[neptune_callback],
)

Support

If you got stuck or simply want to talk to us, here are your options:

• Check our FAQ page
• You can submit bug reports, feature requests, or contributions directly to the repository.
• Chat! When in the Neptune application click on the blue message icon in the bottom-right corner and send a message. A real person will talk to you ASAP (typically very ASAP),
• You can just shoot us an email at [email protected]