ML Inference (TinyGo)

This example demonstrates a simple ML inference workload compiled to WASM. A linear regression model is trained in Python using scikit-learn, and the coefficients are extracted into a pure Go implementation that runs entirely inside the WASM module — no Python runtime needed at inference time.

The model approximates: y = 2*x0 + 3*x1

Before building, a Python training script generates the model and the Go model code is extracted:

cd examples/ml-wasm
python3 train_model.py
go run model_gen.go > /dev/null  # validates coefficients
cd ../..

This creates mymodel.pkl and embeds the coefficients in model_gen.go. The weights are:

• intercept = 0.0
• weight[0] = 2.0 (for input x0)
• weight[1] = 3.0 (for input x1)

Source Code

The source code is available in the examples/ml-wasm directory.

Loading...

Training Script

Loading...

Generated Model

Loading...

Create Task

Build the WASM binary:

cd propeller
GOOS=wasip2 GOARCH=wasm go build -o build/ml-wasm.wasm examples/ml-wasm/main.go examples/ml-wasm/model_gen.go

Now we can create a task:

curl -X POST "http://localhost:7070/tasks" \
-H "Content-Type: application/json" \
-d '{"name": "predict", "inputs": ["200", "300"]}'

Inputs are scaled by 100 (e.g. 200 = 2.0) — the module divides by 100, runs inference, and multiplies the result by 100.

Upload Wasm

curl -X PUT "http://localhost:7070/tasks/<task-id>/upload" \
-F "file=@$(pwd)/build/ml-wasm.wasm"

Start Task

curl -X POST "http://localhost:7070/tasks/<task-id>/start"

The result for inputs [2.0, 3.0] should be approximately 1300 (i.e. (2*2.0 + 3*3.0) * 100 = 1300).

Invoking Using WasmTime

wasmtime --invoke predict ./build/ml-wasm.wasm 200 300