Model-Transpiler

Model-Transpiler converts trained machine-learning models into standalone C or Verilog programs. It lets you take classic scikit-learn estimators or lightweight PyTorch MLPs and run them on hardware targets or bare-metal environments without a Python runtime.

The CLI introspects a serialized model, gathers the metadata it needs (weights, pre-processing parameters, topology, etc.), and stitches that information into language-specific templates located in vendors/. The generated sources can be built with gcc or iverilog and are validated by an integration test-suite under tests/.

Supported inputs

Format	Details
.joblib	Any scikit-learn LinearRegression, LogisticRegression (binary or multi-class), DecisionTreeClassifier, or a Pipeline that optionally includes a MinMaxScaler before one of those estimators.
.pt / .pth	PyTorch state_dict files describing an nn.Sequential multilayer perceptron composed of Linear layers with optional ReLU activations.

The metadata retrievers live in src/metadata_retriever/. Add a new retriever and update src/utils.get_metadata_retriever_class to support additional model types.

Repository layout

• main.py – CLI that loads a serialized model, chooses the target language(s), and writes sources to outputs/.
• src/ – Core transpiler logic, including metadata extraction, dependency resolution, and template loading.
• vendors/ – Vendor-provided snippets split into c/ and verilog/ directories with regular_functions/ and main_functions/.
• models/ – Sample models you can transpile immediately.
• tests/test_pipeline.py – End-to-end tests that transpile, compile, and execute every sample model.
• benchmark.sh / time_inference.py – Utilities to compare Python inference time against generated binaries.

Transpiling models

python main.py --path models/linear.joblib --c --v

• --path points to any supported .joblib, .pt, or .pth file.
• Pass --c and/or --v to restrict languages; omit both to emit both.
• Outputs land in outputs/<model_name>.c or .v. Existing files are overwritten.

You can now compile and run the generated programs:

gcc outputs/linear.c -o outputs/linear_c.out
outputs/linear_c.out 1 2 3

iverilog -o outputs/linear_v.out outputs/linear.v
vvp outputs/linear_v.out 1 2 3

Using your own templates

main.py loads .config to find the vendor directory:

[FunctionLoader]
VENDOR_DIRECTORY = ./vendors

Point it to another folder to swap in different runtime implementations. Each regular function can declare dependencies in src/dependencies.py; the transpiler resolves and inlines those snippets automatically for the chosen language.

Testing

The integration tests ensure end-to-end parity between Python predictions and the generated binaries (including preprocessing paths).

pytest tests/test_pipeline.py

Tests expect gcc and iverilog on PATH. They create temporary directories, so no artifacts remain in outputs/.

Benchmarking

benchmark.sh transpiles every model in models/, builds the outputs, and compares the runtime of:

1. Native Python via time_inference.py
2. Compiled C binary
3. Compiled Verilog simulation

The script requires hyperfine, gcc, and iverilog. Results are written to stdout alongside the generated outputs/<name>_{c,v}.out binaries.