Tutorial 3: Compass IR Overview

Compass IR defines the operations inside the NN model. It contains info includes:

• model name

• layer number

• input tensor, output tensor

• each layer info:

  • layer id

  • layer name

  • layer type

  • bottom and top tensor(shape,layout, dtype, …)

  • quantization info (scale, shift, …)

Parsing the NN model from ONNX/Pytorch/TensorlowLite, we can obtain:

• A txt file, so call the IR definition for model (compass_ir.txt)

• A .bin binary file for weight data (weight.bin)

Compass IR is compatible with Netron, you can inspect the structure of our Compass IR by using Netron.