Just give me the dataz!

Welcome! Leapy is a library for real-time, sub-millisecond inference; it provides customizable machine learning pipeline export for fast model serving. These pipelines are targeted for using Dask's scalable machine learning, which follows the Scikit-Learn API. However, you can use this framework directly with Scikit-Learn as well.

pipe = Pipeline([
        ('fp', FeaturePipeline([('ohe',
                                 OneHotEncoder(sparse=False),
                                 [0, 1])])),
        ('clf', LogisticRegression())
])

pipe.fit(X, y)

pipeline_runtime = pipe.to_runtime()               # ⚡⚡⚡ 
init('./model_repo', pipeline_runtime, df.head())  # Ready to deploy

And serve this super fast pipeline:

$ leap serve --repo ./model_repo
$ curl localhost:8080:/health
{
  "status": "healthy"
}

(See below for benchmarks and a more detailed usage example.)

Dask is a Python distributed computing environment in Python with a burgeoning machine learning component, compatible with Scikit-Learn's API. Using Leapy's framework, we can serve these pipelines in real-time!

This means:

• No JVM: No reliance on JVM from using Spark.
• Python: All Python development and custom transformers -- no Scala & Java needed!
• Scale: Scikit-Learn logic and pipelines scaled to clusters.
• Fast: You're in control of how fast your transformers are.
• Simple: Easily build and deploy models with Docker.
• Reliable: Encourages a test-driven approach.

• Simple -- Super simple example of creating, testing, and using custom transformers
• XGBoost -- Advanced example of using XGBoost with Dask, saving, and serving the model.

A simple example of what we're going for -- computing a one-hot-encoding, with ~200K labels, of a single data point (dask array x_da and numpy array x = x_da.compute()):

Where ohe_dml (from dask_ml) and ohe (from leapy) are essentially the same; ohe_sk is from scikit-learn and ohe_runtime is from ohe.to_runtime(). And, running compute() on Dask transforms above bumps the time up to about 1 second.

With the time we save using ohe_runtime, we can squeeze in many more transformations and an estimator to still come in under 1ms.

Start with a dataset in dask arrays, X, y, and dataframe ddf: pipeline:

import numpy as np
import pandas as pd
import dask.array as da
import dask.dataframe as dd

X_np = np.array([[1, 'a'], [2, 'b']], dtype=np.object)
df_pd = pd.DataFrame(X_np, columns=['test_int', 'test_str'])
y_np = np.array([0, 1])

X = da.from_array(X_np, chunks=X_np.shape)
y = da.from_array(y_np, chunks=y_np.shape)
ddf = dd.from_pandas(df_pd, npartitions=1)

Create our pipeline:

from sklearn.pipeline import Pipeline
from dask_ml.linear_model import LogisticRegression
from leapy.dask.transformers import OneHotEncoder
from leapy.dask.pipeline import FeaturePipeline
from leapy.serve import init

pipe = Pipeline([
        ('fp', FeaturePipeline([
            # One-Hot-Encode 'test_str' feature, drop 'test_int'
            ('ohe', OneHotEncoder(sparse=False), [1])])),
        ('clf', LogisticRegression())
])

pipe.fit(X, y)

Then we export to a runtime pipeline and get ready for model serving:

pipe_runtime = pipe.to_runtime()
init('./model_repo', pipe_runtime, ddf.head())

Finally we serve the model:

$ leap serve --repo ./model_repo
$ curl localhost:8080/predict \
    -X POST \
    -H "Content-Type: application/json" \
    --data '{"test_int": 1, "test_str": "b"}'
{
  "prediction": 1.0
}

For more on model serving see leapy/serve/README.md.

Leapy is inspired by MLeap.