Exploring the Potential of Flexible 8-bit Format: Design and Algorithm

Requirements

To install requirements:

pip install -r requirements.txt

MMlab Network Config Prepare

For convenience, we have introduced the mmlab framework into the code. If you have problems with the installation of the openmmlab framework, you can refer to the link https://mmcv.readthedocs.io/en/latest/get_started/installation.html

And before you use it, you should prepare dataset like ImageNet, COCO2017, and CityScapes. Organized as below

dataset_root
└── imagenet
    ├── meta
    ├── train
    └── val 

Take resnet50 as an example:

cd ${code_root}
# dataset_root should end with '/'
sh scripts/mmlab/download_network_config/cls/res50.sh ${dataset_root}

Bert Config Prepare

Download bert-base-uncased models from https://huggingface.co/bert-base-uncased

Running Quantization Experiments

# mmlab model
sh script/mix_all_8bit_script/cls/res50.sh

# bert model
sh script/glue/run.sh