skorch

Why we included this project

skorch is for teams that know scikit-learn well and want to train neural networks without leaving that ecosystem. It turns PyTorch modules into sklearn-style estimators like NeuralNetClassifier and NeuralNetRegressor, so a network you define in PyTorch can be trained and tuned with the Pipeline, GridSearchCV, and scoring machinery you already use. The package also bundles training conveniences like learning-rate schedulers, early stopping, checkpointing, and parameter freezing, and it can wrap GPyTorch models for Gaussian processes or Hugging Face transformers. That means you can experiment with deep models while keeping the evaluation and deployment patterns you are used to. It is a building block rather than a turnkey app, so you still write your own PyTorch module definitions.