onnx-codegen

Generate self-contained C inference code from ONNX models, targeting bare-metal MCU deployment.

Features

• Pure C output with no external runtime dependencies
• Static-shape ONNX graphs (recommended for MCU targets)
• Quantised inference: 8w8a, 8w16a, 16w16a weight/activation bit-widths with mixed act storage + int32 accumulator/logit runtime types when needed
• Fused Conv1D + requantisation kernels (pulp-nn style)
• Liveness-based scratch buffer allocation to minimise SRAM usage
• Per-layer #if guards for selective compilation

Supported ONNX Ops

Identity, Relu, LeakyRelu, Sigmoid, Tanh, Clip, Add, Mul, Div, Floor, MatMul, Gemm, Conv (1D NCW, 2D NCHW, groups), MaxPool, AveragePool, GlobalAveragePool, Flatten, Reshape, Transpose, Squeeze, Unsqueeze, Softmax, BatchNormalization, Concat, Pad, Slice.

Installation

pip install .

Or in development mode:

pip install -e .

Usage

onnx-codegen --onnx model.onnx --out-dir output/ --prefix my_model --quant 8w8a --rogue 0

Or, to enable Rogue kernels:

onnx-codegen --onnx model.onnx --out-dir output/ --prefix my_model --quant 8w8a --rogue 1

The CLI prints a memory breakdown after generation, including weights/constants, scratch buffers, input/output tensors, and total estimated model memory.

To generate the C and immediately compare it against the ONNX reference with a temporary main.c harness:

onnx-codegen --onnx model.onnx --out-dir output/ --prefix my_model --compare

If the generated C comes from an integer-form model but the reference should be the original float-parameter ONNX, pass it explicitly:

onnx-codegen --onnx model_int.onnx --compare --compare-onnx model_float.onnx

Options:

• --onnx — Path to .onnx file or .zip containing one
• --out-dir — Output directory (default: .)
• --prefix — C symbol prefix (default: sanitised model filename)
• --quant NwMa — Force N-bit weights and requantized activation storage (e.g. 8w8a); accumulators/residual sums/logits may still use int32
• --skip-shape-inference — Skip ONNX shape inference pass
• --custom-kernels-header — Extra header to include in generated .c
• --compare — Compile the generated C and compare it against the ONNX model
• --compare-onnx — Alternate ONNX file used only as the comparison reference
• --compare-random-cases — Additional random inputs for --compare (default: 2)
• --compare-seed — RNG seed for --compare (default: 0)
• --cc — C compiler used by --compare (default: $CC or gcc)

Output Files

File	Description
<prefix>_model.h	Public API: <prefix>_init(), <prefix>_run()
<prefix>_model.c	Inference implementation with all kernels inlined
<prefix>_weights.h	Static constant weight arrays
<prefix>_layer_cfg.h	Per-layer enable/disable switches

License

Apache-2.0