aidge-quantization 0.4.0

Quantization algorithms to compress aidge networks.

Aidge Quantization Module

You can find in this folder the library that implements the quantization algorithms. For the moment only Post Training Quantization (PTQ) is available. Its implementation does support multiple branch architectures.

[TOC]

Installation

Dependencies

• GCC
• Make/Ninja
• CMake
• Python (optional, if you have no intend to use this library in python with pybind)

Aidge dependencies

• aidge_core The requirements for installing the library are the followings:

  • GCC, Make and CMake for the compilation pipeline
  • The AIDGE modules aidge_core, aidge_onnx and aidge_backend_cpu
  • Python (> 3.7) if you intend to use the pybind wrapper

Pip installation

pip install . -v

TIPS : Use environment variables to change compilation options :

• AIDGE_INSTALL : to set the installation folder. Defaults to /usr/local/lib. :warning: This path must be identical to aidge_core install path.
• AIDGE_PYTHON_BUILD_TYPE : to set the compilation mode to Debug or Release
• AIDGE_BUILD_GEN : to set the build backend with

User guide

In order to perform a quantization, you will need an AIDGE model (that can be loaded from an ONNX). Then, you will have to provide a calibration dataset consisting of AIDGE tensors (that can be loaded from some numpy arrays). And finally, you will have to specify the quantization number of bits.

Performing the PTQ on your model will then be a one liner:

aidge_quantization.quantize_network(aidge_model, nb_of_bits, calibration_set)

Technical insights

The PTQ algorithm consists of 3 main steps:

- Normalization of the parameters, so that each node set of weights fits in the [-1:1] range.
- Normalization of the activations, so that each node output value fits in the [-1:1] range.
- Quantization of the scaling nodes previously inserted

To achieve those steps, one must propagate the scaling factors inside the network. One should also balance the different branches when they are merging. A particular care is needed for the biases rescaling at each step.

Doing quantization step by step

It is possible to perform the PTQ step by step, thanks to the exposed functions of the API. In that case, here is the standard pipeline:

- Prepare the network for the PTQ (remove the flatten nodes, fuse the BatchNorms ...)
- Insert the scaling nodes that will allow the model calibration
- Perform the Cross Layer Equalization if possible
- Perform the parameter normalization
- Compute the node output ranges over an input calibration dataset
- Adjust the output ranges using a specified error metric (MSE, KL, ...)
- Perform the activation normalization
- Quantize the normalized network
- Convert the scaling factors to bit-shifting operations if needed

Further work

• add Quantization Aware Training (QAT)