QCore

QCore: Data-Efficient, On-Device Continual Calibration for Quantized Models (under review)

How to run the model:

• Train the full-precision model and generate QCore. Run main.py by specifying the datasets for training and streaming. For example: python main.py --bits 4 --core_size 20 --data_source activities --dataset 1 --stream_dataset 2
• Baselines can be found in the /baselines/ folder. To run them, execute main.py while specifying the appropriate model and the original and target domains. For instance, after accessing the folder using cd baselines, a possible command would be: python ./utils/main.py --bits 4 --dataset har --lr 0.01 --buffer_size 20 --data_source activities --model er --data_in 1 --stream_dataset 2
• The data is managed in the dataloader.py file.
• Time-series data can be found in the /data/ folder. The results will be inserted into a local database.
• Detailed information about all the parameters can be found in each execution file.

Citation

If you use the code, please cite the following paper:

  
@article{pvldb/Ca24,
  author    = {David Campos and Bin Yang and Tung Kieu and Miao Zhang and Chenjuan Guo and Christian S. Jensen},
  title     = {{QCore: Data-Efficient, On-Device Continual Calibration for Quantized Models}},
  journal   = {{PVLDB}},
  volume    = {17},
  year      = {2024}
}