Quickstart: Run a Model Zoo Model Locally¶

This guide walks you through setting up the DeepGate SDK locally, downloading a Model Zoo model, and exporting the schema JSON that Bitweaver needs to compile and benchmark your model. Once you have the schema, continue to Compile & Benchmark to upload it to Bitweaver and run it on real hardware.

Don't have a Bitweaver account yet? Get in touch.

1. Install the DeepGate SDK¶

pip install deepgate

Requires Python 3.10+ and PyTorch 2.6+. The package is deepgate on PyPI; the import name is dg.

2. Pick a model from the Zoo¶

The Model Zoo is available in the Bitweaver web app under Model Zoo. Pre-trained models live there with the Model-ID you can pass to dg.from_pretrained.

This guide uses wakeword, a 12-class keyword-spotting model trained on Speech Commands v2 (10 keywords + silence + unknown). The Zoo will be continuously populated with more models. The steps below apply to any of them.

3. Download the model¶

import dg

model = dg.from_pretrained("wakeword")

from_pretrained resolves the model ID and downloads model.py, model.pt, and config.json into a cache at ~/.cache/dg/models/wakeword/ (not your current directory). It then instantiates the model and loads the weights. Subsequent calls hit the cache without re-downloading.

4. Export the schema JSON¶

import torch

# Run one forward pass so the model records its input shape (required
# before save_pretrained can write schema.json).
model(torch.zeros(1, 1, 49, 10))

model.save_pretrained("./wakeword")

save_pretrained writes a fresh ./wakeword/ directory in your current working directory (separate from the cache used in step 3) containing four files:

File	What it is
`model.pt`	Trained weights
`model.py`	Self-contained model architecture
`config.json`	Constructor kwargs
`schema.json`	I/O contract and layer layout, the file Bitweaver needs

Input shape per model

The forward pass needs the model's expected input shape. For wakeword it's (1, 1, 49, 10) (batch × channels × MFCC frames × MFCC coefficients). Each Zoo model will document its expected input shape on Bitweaver's Model Zoo.

5. Continue to Compile & Benchmark¶

You now have ./wakeword/schema.json. Continue to:

Compile & Benchmark: upload the schema, pick target boards, and review on-MCU performance.
End-to-End Tutorial: Build and train your own model with dg from scratch, then deploy through Bitweaver.