HyQ: Hardware-Friendly Post-Training Quantization for CNN-Transformer Hybrid Networks

Authors: Nam Joon Kim, Jongho Lee, Hyun Kim

IJCAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that the proposed quantization method with INT8 precision demonstrated a 0.39% accuracy drop compared with the FP32 baseline on Mobile Vi T-s with the Image Net-1k dataset.
Researcher Affiliation Collaboration Nam Joon Kim1 , Jongho Lee1,2 and Hyun Kim1 1Department of Electrical and Information Engineering and Research Center for Electrical and Information Technology, Seoul National University of Science and Technology 2Squeeze Bits Inc.
Pseudocode Yes Algorithm 1 Integer-only Linear Exponential
Open Source Code Yes The code is available at https://github.com/IDSL-Seoul Tech/Hy Q.
Open Datasets Yes The proposed Hy Q framework was validated using the Image Net-1k benchmark dataset [Deng et al., 2009].
Dataset Splits Yes The proposed Hy Q framework was validated using the Image Net-1k benchmark dataset [Deng et al., 2009]. To optimize the QADS parameters, we used 100 unlabeled images from the Image Net-1k training set.
Hardware Specification Yes Notably, applying QADS to Mobile Vi T-xxs took only 2.5 minutes on a single NVIDIA Ge Force RTX 3090 GPU. [...] We compared the synthesis results of two approaches [...] designed in RTL using the XC7Z0101 chipset in the Zynq-7000 board at 125 MHz.
Software Dependencies No We used the Py Torch framework [Paszke et al., 2019] for all experiments and quantized the pre-trained models provided by the Py Torch Image Models library [Wightman, 2019].
Experiment Setup Yes We performed 1,000 iterations with a batch size of 100 using the Adam optimizer [Kingma and Ba, 2014].