Brevitas Export

FINN expects an ONNX model as input. This can be a model trained with Brevitas. Brevitas is a PyTorch library for quantization-aware training and the FINN Docker image comes with several example Brevitas networks. Brevitas provides an export of a quantized network in QONNX representation, which is the format that can be ingested by FINN. In a QONNX graph, all quantization is represented using Quant, BinaryQuant or Trunc nodes. QONNX must be converted into FINN-ONNX by finn.transformation.qonnx.convert_qonnx_to_finn. FINN-ONNX is the intermediate representation (IR) FINN uses internally. In this IR, quantized weights are indicated through tensors with additional attributes to mark low-precision datatypes and quantized activations are expressed as MultiThreshold nodes.

To work with either type of ONNX model, it is loaded into a ModelWrapper provided by FINN.

At this stage we can already use the functional verification flow to simulate the model using Python. For more details please have look at Functional Verification.

The model can now be further processed in FINN, the next flow step is Network Preparation.