EXGBoost

Elixir bindings to the XGBoost C API using Native Implemented Functions (NIFs).

EXGBoost provides an implementation of XGBoost that works with Nx tensors.

Xtreme Gradient Boosting (XGBoost) is an optimized distributed gradient boosting library designed to be highly efficient, flexible and portable. It implements machine learning algorithms under the Gradient Boosting framework. XGBoost provides a parallel tree boosting (also known as GBDT, GBM) that solve many data science problems in a fast and accurate way. The same code runs on major distributed environment (Hadoop, SGE, MPI) and can solve problems beyond billions of examples.

Installation

def deps do
[
  {:exgboost, "~> 0.5"}
]
end

API Data Structures

EXGBoost's top-level EXGBoost API works directly and only with Nx tensors. However, under the hood, it leverages the structs defined in the EXGBoost.Booster and EXGBoost.DMatrix modules. These structs are wrappers around the structs defined in the XGBoost library. The two main structs used are DMatrix to represent the data matrix that will be used to train the model, and Booster which represents the model.

The top-level EXGBoost API does not expose the structs directly. Instead, the structs are exposed through the EXGBoost.Booster and EXGBoost.DMatrix modules. Power users might wish to use these modules directly. For example, if you wish to use the Booster struct directly then you can use the EXGBoost.Booster.booster/2 function to create a Booster struct from a DMatrix and a keyword list of options. See the EXGBoost.Booster and EXGBoost.DMatrix modules source for more implementation details.

Basic Usage

key = Nx.Random.key(42)
{x, key} = Nx.Random.normal(key, 0, 1, shape: {10, 5})
{y, key} = Nx.Random.normal(key, 0, 1, shape: {10})
model = EXGBoost.train(x, y)
EXGBoost.predict(model, x)

Training

EXGBoost is designed to feel familiar to the users of the Python XGBoost library. EXGBoost.train/2 is the primary entry point for training a model. It accepts a Nx tensor for the features and a Nx tensor for the labels. EXGBoost.train/2 returns a trainedBooster struct that can be used for prediction. EXGBoost.train/2 also accepts a keyword list of options that can be used to configure the training process. See the XGBoost documentation for the full list of options.

EXGBoost.train/2 uses the EXGBoost.Training.train/1 function to perform the actual training. EXGBoost.Training.train/1 and can be used directly if you wish to work directly with the DMatrix and Booster structs.

One of the main features of EXGBoost.train/2 is the ability for the end user to provide a custom training function that will be used to train the model. This is done by passing a function to the :obj option. The function must accept a DMatrix and a Booster and return a Booster. The function will be called at each iteration of the training process. This allows the user to implement custom training logic. For example, the user could implement a custom loss function or a custom metric function. See the XGBoost documentation for more information on custom loss functions and custom metric functions.

Another feature of EXGBoost.train/2 is the ability to provide a validation set for early stopping. This is done by passing a list of 3-tuples to the :evals option. Each 3-tuple should contain a Nx tensor for the features, a Nx tensor for the labels, and a string label for the validation set name. The validation set will be used to calculate the validation error at each iteration of the training process. If the validation error does not improve for :early_stopping_rounds iterations then the training process will stop. See the XGBoost documentation for a more detailed explanation of early stopping.

Early stopping is achieved through the use of callbacks. EXGBoost.train/2 accepts a list of callbacks that will be called at each iteration of the training process. The callbacks can be used to implement custom logic. For example, the user could implement a callback that will print the validation error at each iteration of the training process or to provide a custom setup function for training. See the EXGBoost.Training.Callback module for more information on callbacks.

Please notes that callbacks are called in the order that they are provided. If you provide multiple callbacks that modify the same parameter then the last callback will trump the previous callbacks. For example, if you provide a callback that sets the :early_stopping_rounds parameter to 10 and then provide a callback that sets the :early_stopping_rounds parameter to 20 then the :early_stopping_rounds parameter will be set to 20.

You are also able to pass parameters to be applied to the Booster model using the :params option. These parameters will be applied to the Booster model before training begins. This allows you to set parameters that are not available as options to EXGBoost.train/2. See the XGBoost documentation for a full list of parameters.

EXGBoost.train(X,
              y,
              obj: &EXGBoost.Training.train/1,
              evals: [{X_test, y_test, "test"}],
              learning_rates: fn i -> i/10 end,
              num_boost_round: 10,
              early_stopping_rounds: 3,
              max_depth: 3,
              eval_metric: [:rmse,:logloss]
              )

Prediction

EXGBoost.predict/2 is the primary entry point for making predictions with a trained model. It accepts a Booster struct (which is the output of EXGBoost.train/2). EXGBoost.predict/2 returns a Nx tensor containing the predictions. EXGBoost.predict/2 also accepts a keyword list of options that can be used to configure the prediction process.

preds = EXGBoost.train(X, y) |> EXGBoost.predict(X)

Serialization

A Booster can be serialized to a file using EXGBoost.write_* and loaded from a file using EXGBoost.read_*. The file format can be specified using the :format option which can be either :json or :ubj. The default is :json. If the file already exists, it will NOT be overwritten by default. Boosters can either be serialized to a file or to a binary string. Boosters can be serialized in three different ways: configuration only, configuration and model, or model only. dump functions will serialize the Booster to a binary string. Functions named with weights will serialize the model's trained parameters only. This is best used when the model is already trained and only inferences/predictions are going to be performed. Functions named with config will serialize the configuration only. Functions that specify model will serialize both the model parameters and the configuration.

Output Formats

• read/write - File.
• load/dump - Binary buffer.

Output Contents

• config - Save the configuration only.
• weights - Save the model parameters only. Use this when you want to save the model to a format that can be ingested by other XGBoost APIs.
• model - Save both the model parameters and the configuration.

Plotting

EXGBoost.plot_tree/2 is the primary entry point for plotting a tree from a trained model. It accepts an EXGBoost.Booster struct (which is the output of EXGBoost.train/2). EXGBoost.plot_tree/2 returns a VegaLite spec that can be rendered in a notebook or saved to a file. EXGBoost.plot_tree/2 also accepts a keyword list of options that can be used to configure the plotting process.

See EXGBoost.Plotting for more detail on plotting.

You can see available styles by running EXGBoost.Plotting.get_styles() or refer to the EXGBoost.Plotting.Styles documentation for a gallery of the styles.

Kino & Livebook Integration

EXGBoost integrates with Kino and Livebook to provide a rich interactive experience for data scientists.

EXGBoost implements the Kino.Render protocol for EXGBoost.Booster structs. This allows you to render a Booster in a Livebook notebook. Under the hood, EXGBoost uses Vega-Lite and Kino Vega-Lite to render the Booster.

See the Plotting in EXGBoost Notebook for an example of how to use EXGBoost with Kino and Livebook.

Examples

See the example Notebooks in the left sidebar (under the Pages tab) for more examples and tutorials on how to use EXGBoost.

Requirements

Precompiled Distribution

We currenly offer the following precompiled packages for EXGBoost:

%{
  "exgboost-nif-2.16-aarch64-apple-darwin-0.5.0.tar.gz" => "sha256:c659d086d07e9c209bdffbbf982951c6109b2097c4d3008ef9af59c3050663d2",
  "exgboost-nif-2.16-x86_64-apple-darwin-0.5.0.tar.gz" => "sha256:05256238700456c57e279558765b54b5b5ed4147878c6861cd4c937472abbe52",
  "exgboost-nif-2.16-x86_64-linux-gnu-0.5.0.tar.gz" => "sha256:ad3ba6aba8c3c2821dce4afc05b66a5e529764e0cea092c5a90e826446653d99",
  "exgboost-nif-2.17-aarch64-apple-darwin-0.5.0.tar.gz" => "sha256:745e7e970316b569a10d76ceb711b9189360b3bf9ab5ee6133747f4355f45483",
  "exgboost-nif-2.17-x86_64-apple-darwin-0.5.0.tar.gz" => "sha256:73948d6f2ef298e3ca3dceeca5d8a36a2d88d842827e1168c64589e4931af8d7",
  "exgboost-nif-2.17-x86_64-linux-gnu-0.5.0.tar.gz" => "sha256:a0b5ff0b074a9726c69d632b2dc0214fc7b66dccb4f5879e01255eeb7b9d4282",
}

The correct package will be downloaded and installed (if supported) when you install the dependency through Mix (as shown above), otherwise you will need to compile manually.

NOTE If MacOS, you still need to install libomp even to use the precompiled libraries:

brew install libomp

Dev Requirements

If you are contributing to the library and need to compile locally or choose to not use the precompiled libraries, you will need the following:

• Make
• CMake
• If MacOS: brew install libomp

When you run mix compile, the xgboost shared library will be compiled, so the first time you compile your project will take longer than subsequent compilations.

You also need to set CC_PRECOMPILER_PRECOMPILE_ONLY_LOCAL=true before the first local compilation, otherwise you will get an error related to a missing checksum file.

Known Limitations

• The XGBoost C API uses C function pointers to implement streaming data types. The Python ctypes library is able to pass function pointers to the C API which are then executed by XGBoost. Erlang/Elixir NIFs do not have this capability, and as such, streaming data types are not supported in EXGBoost.
• Currently, EXGBoost only works with tensors from the Nx.Binarybackend. If you are using any other backend you will need to perform an Nx.backend_transfer or Nx.backend_copy before training an EXGBoost.Booster. This is because Nx tensors are JSON-encoded and serialized before being sent to XGBoost and the binary backend is required for proper JSON-encoding of the underlying tensor.

Roadmap

• CUDA support
• Collective API?

License

Licensed under an Apache-2 license.