Conga2022

Code for using Qtorch+ and reproduce Conga2022 results

Update 2023:

For object dectection model, the torchbench and sotabench are not maintained. Follow the below steps to run the scripts:

Download coco dataseet and setup like :

1. mkdir -p ./.data/vision/coco
2. wget http://images.cocodataset.org/zips/val2017.zip
3. wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
4. unzip both zipped files and put them into ./.data/vision/coco

   Install the modified version of torchbench

5. uninstall existing torchbench if exist: pip uninstall torchbench
6. cd .. ; git clone https://github.com/minhhn2910/torchbench
7. cd torchbench ; pip install -e ./
8. Now you can go back to conga2022 and run the script: cd conga2022 ; python torchbench_coco.py

Part of the raw data is logged in this google docs:

(this includes scale weight and scale act parameters for scaling posit, raw data logged long time a go so please be patient when reading it :D ) https://docs.google.com/document/d/19CYdPaVoKkQGT27jPXwxmeItIZP8jLHvOOHL7lAa-Ms/edit?usp=sharing