ShapeShifter Tutorials#
ShapeShifter is Quark’s pass-based graph-transformation framework for ONNX (and PyTorch) models. ONNX passes run in two stages: preprocessing passes prepare the float model before quantization (constant folding, operator fusion, BatchNorm folding, opset conversion, …), and postprocessing passes adapt the quantized Q/DQ model for a deployment target (Q/DQ scale alignment, bias-scale correction, XINT8/NPU simulation, …). The tutorials below show how to drive these passes.
ShapeShifter on ResNet50 (CLI and ShapeShifterYaml)
Run ShapeShifter ONNX passes on ResNet50 two ways: standalone with the
quark-cli shapeshifter command, and integrated into a quantization script
via the ShapeShifterYaml option (preprocessing + postprocessing).