import FeatureHighlights from "@site/src/components/FeatureHighlights"; import TilesGrid from "@site/src/components/TilesGrid"; import TileCard from "@site/src/components/TileCard"; import { Zap, GitBranch, Package, BookOpen } from "lucide-react";

MLflow spaCy Integration

Introduction

spaCy is an industrial-strength natural language processing library designed for production use. It provides pre-trained models and efficient processing pipelines for tasks like named entity recognition, part-of-speech tagging, and text classification.

MLflow's spaCy integration provides model logging, versioning, and deployment capabilities for NLP workflows.

Why MLflow + spaCy?

<FeatureHighlights features={[ { icon: Package, title: "Model Packaging", description: "Log spaCy models with all pipeline components and dependencies automatically captured.", }, { icon: GitBranch, title: "Experiment Tracking", description: "Track NLP metrics, model performance, and training configurations across experiments.", }, { icon: Zap, title: "Easy Deployment", description: "Deploy spaCy models as REST APIs or batch inference pipelines with MLflow serving.", }, { icon: BookOpen, title: "Version Control", description: "Manage different model versions and pipeline configurations with MLflow's model registry.", }, ]} />

Model Logging

Log spaCy models to MLflow:

import mlflow
import spacy

# Load or train your spaCy model
nlp = spacy.load("en_core_web_sm")

# Log to MLflow
model_info = mlflow.spacy.log_model(nlp, name="spacy_model")

Custom Training with MLflow

Track custom spaCy training with MLflow:

import mlflow
import spacy
from spacy.training import Example

# Load base model
nlp = spacy.blank("en")
nlp.add_pipe("ner")

# Sample training data
TRAIN_DATA = [
    ("Apple is a tech company", {"entities": [(0, 5, "ORG")]}),
    ("Google acquired YouTube", {"entities": [(0, 6, "ORG"), (16, 23, "PRODUCT")]}),
]

# Convert to Examples
examples = [Example.from_dict(nlp.make_doc(text), annotations) for text, annotations in TRAIN_DATA]

# Initialize and train
optimizer = nlp.initialize()

with mlflow.start_run():
    mlflow.log_params({
        "model": "blank_en",
        "pipeline": "ner",
    })

    for epoch in range(10):
        losses = {}
        for example in examples:
            nlp.update([example], sgd=optimizer, losses=losses)

        mlflow.log_metric("loss", losses["ner"], step=epoch)

    # Log the trained model
    mlflow.spacy.log_model(nlp, name="custom_ner_model")

Model Loading

Load spaCy models from MLflow:

import mlflow

# Load as spaCy model
nlp = mlflow.spacy.load_model("models:/<model_id>")
doc = nlp("Apple is looking at buying a startup")

# Load as PyFunc for deployment
predictor = mlflow.pyfunc.load_model("models:/<model_id>")
predictions = predictor.predict(["Text to process"])

Learn More