Skip to content

image-classification

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

image-classification

README.md

ICP image classification

This example demonstrates running an ONNX machine-learning model inside an ICP canister. The canister accepts an image from the user and runs image classification inference using the Tract ONNX inference engine with the MobileNet v2-7 model.

The example uses the WASI polyfill to run Tract (which relies on POSIX file I/O) inside the deterministic ICP runtime, and Wasm SIMD instructions for faster inference.

The example consists of two canisters:

  • backend — embeds the Tract ONNX inference engine with the MobileNet v2-7 model. It provides two classification endpoints:
    • classify(image) — update call, runs under replicated execution on all nodes (secure, slower).
    • classify_query(image) — query call, runs on a single node (fast, not replicated).
  • frontend — serves the web UI (HTML/JS/CSS) from which users upload images and view results.

Build and deploy from the command line

Prerequisites

  • Node.js v18+
  • icp-cli: npm install -g @icp-sdk/icp-cli @icp-sdk/ic-wasm
  • Rust v1.85+ with wasm32-wasip1 target: rustup target add wasm32-wasip1
  • wasi2ic: cargo install wasi2ic

wasm-opt is installed automatically on first deploy if not already present.

Install

git clone https://github.com/dfinity/examples
cd examples/rust/image-classification

Download the MobileNet v2-7 model:

./download_model.sh

Deploy and test

icp network start -d
icp deploy
bash test.sh
icp network stop

If the deployment is successful, the CLI will print the frontend URL. Open that URL in a browser to interact with the canister.

For frontend development with hot reload:

npm run dev --prefix frontend

Updating the Candid interface

Only needed if you change the backend endpoints. Requires candid-extractor (cargo install candid-extractor) and ic_cdk::export_candid!() at the end of backend/src/lib.rs (already present):

icp build backend && candid-extractor target/wasm32-wasip1/release/backend.wasm > backend/backend.did

Security considerations and best practices

Refer to the ICP security best practices for guidance on securing your canister.