Skip to content

Overview

DeepGate's Python SDK extends PyTorch to simplify quantizing and exporting models, and provides building blocks for developing more efficient machine learning models for edge deployment on microcontrollers. You author an ordinary PyTorch nn.Module, and the SDK handles the rest of the road to the edge.

The Quickstart gets you running end to end in a few minutes.

The path

Everything the SDK does maps to the steps you'll take in order:

  1. Build – author in PyTorch and use DeepGate layers where needed to push the boundaries of efficiency for your edge AI application.
  2. Pre-process – use image and audio pre-processing that's built into your exported model.
  3. Train / quantize – train and quantize in PyTorch with built-in QAT and PTQ for smaller, faster edge AI models.
  4. Export – directly export quantized PyTorch models using DeepGate's schema.
  5. Compile & benchmark – compile for your target hardware and benchmark your model on the DeepGate platform to unlock the best performance.

Why it exists

PyTorch is the most popular framework for building AI models, but it's not designed for edge deployment. The DeepGate SDK extends PyTorch with everything you need to build, quantize, and export efficient edge AI models in a single workflow – without separate conversion or quantization tools – while giving you access to DeepGate's optimized layers.

TFLite support

Need TFLite? The DeepGate developer platform also supports the TFLite specification. See the Developer Notes for the conventional PyTorch to TFLite workflow.