Tutorial 4

Deep Learning Workflows with Skorch and RAPIDS

This notebook focuses on model development workflows that combine Skorch with GPU-enabled tooling, giving attendees a reusable pattern for deep learning experimentation in notebook form.

Skorch is a high-level library for building and training neural networks in PyTorch. It provides a scikit-learn-like interface for PyTorch, making it easier to integrate with other libraries in the Python ecosystem. Skorch allows users to leverage the power of PyTorch while benefiting from the simplicity and ease of use of scikit-learn.

The aim is to keep 99% of the flexibility of PyTorch while being able to leverage most features of scikit-learn.

Distributed Hyperparameter Optimization with Ray Tune and RAPIDS

What is RayTune?

RayTune is a scalable Hyperparameter optimization library. It allows distributed HPO, provides various search algorithms to allow different optimization techniques to be explored with ease. The library also provides scheduling algorithms that allows a smarter way to schedule the different parameter sweep instead of the basic First In-First Out method which is followed by other libraries (Scikit-Learn, Dask-ml) that support HPO. The different scheduling algorithms can make the HPO process resource efficient and help arrive at the best parameters much faster.

High-Throughput HPO using Ray Tune and Fractional GPU Allocation with RAPIDS & PyTorch

This notebook demonstrates how to use Ray Tune for high-throughput hyperparameter optimization while leveraging fractional GPU allocation with RAPIDS and PyTorch.

Zero-Copy Deep Tabular Learning with RAPIDS, Optuna, and PyTorch Lightning

Normally, moving data from a GPU DataFrame into a PyTorch tensor means: copy GPU -> host RAM (as pandas/NumPy), then copy host RAM -> GPU again for torch. For a dataset with millions of rows, that round-trip dominates preprocessing time and doubles peak memory usage.

DLPack is an open, framework-agnostic tensor memory standard. cudf can export a DataFrame’s underlying GPU buffer as a DLPack capsule with to_dlpack(), and PyTorch can import that exact same GPU memory with torch.utils.dlpack.from_dlpack() — no host copy, no new GPU allocation, just a reinterpretation of the existing buffer as a torch.Tensor.

What you will do

open a ready-to-run Colab notebook
work through a GPU-based deep learning training workflow
experiment with tuning and evaluation in an interactive environment
adapt the notebook for follow-up research projects