(C) Copyright EdgeCortix Inc. 2020

Eager mode quantization in PyTorch

To run the PyTorch eager mode quantization tutorial, cleaned up and modified to use the Mobilenet v3 model:

python tutorial_eager.py

The directory models contains eager-mode, quantization-ready implementations of Mobilenet v2 (Sandler et al. CVPR 2018) and v3. (More quantized models with application to non classification tasks to be added ..)

The script auto_quant.py is an experiment on Torch automatic quantization support. It can be run but currently convolution is performed on fp32. Keep an eye on this thread to track the progress of the development.

TVM QNN Support

The directory tvm_qnn_evaluation contains an evaluation script for TVM QNN implementation in the PR. Unless you are interested in TVM, you can ignore this directory. See the README there for details.

Requirements

wget (pip install wget)
packaging (pip install packaging)
numpy
Pillow >= 0.7
PyTorch >= 1.4.0 (> 1.4.0 to trace quantized mobilenet v3 for JIT)
torchvision >= 0.5.0
TVM, optional (the latest one built from source, required for QNN evaluation)