ONNX Import

Overview

nnc import converts ONNX models to NNL format with extracted weights.

nnc import model.onnx -o model.nnl

Supported ONNX Operators

ONNX Op	NNL Layer
Gemm / MatMul	Dense
Conv	Conv2D
MaxPool	MaxPool2D
AveragePool	AvgPool2D
Flatten	Flatten
BatchNormalization	BatchNorm
Dropout	Dropout
Add	Add
Concat	Concat
Relu	ReLU
Sigmoid	Sigmoid
Softmax	Softmax

Weight Handling

Weights are extracted from ONNX initializers and saved as individual .npy files.
Gemm nodes with transB=1 have their weights automatically transposed to the NNL [in, out] layout.
The batch dimension is stripped from input shapes.

Unsupported Operators

Operators without a mapping are emitted as comments in the generated .nnl file:

// UNSUPPORTED: Reshape(reshape_0)

These require manual resolution — replace the comment with an equivalent NNL layer or restructure the model before export.

Round-Trip Workflow

Train your model in PyTorch, TensorFlow, or another framework.
Export to ONNX (e.g., torch.onnx.export(model, dummy, "model.onnx")).
Import into NNL: nnc import model.onnx -o model.nnl
Compile and run: nnc compile model.nnl -o model && ./model
Test outputs against the original framework to verify correctness.

Limitations

Only float32 weights are supported.
External data is not supported — weights must be embedded in the .onnx file.
Dynamic shapes are not supported; all dimensions must be fixed at export time.