Linear Symmetric Quantization of Neural Networks for Low-precision Integer Hardware

Authors: Xiandong Zhao, Ying Wang, Xuyi Cai, Cheng Liu, Lei Zhang

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, three sets of experiments on Cifar10, Image Net and Pascal VOC datasets are presented.
Researcher Affiliation Academia Institute of Computing Technology, Chinese Academy of Sciences1 University of Chinese Academy of Sciences2 State Key Laboratory of Computer Architecture3 {zhaoxiandong,wangying2009,caixuyi18s,liucheng,zlei}@ict.ac.cn
Pseudocode Yes Algorithm 1 LLSQ
Open Source Code No The paper mentions using "Py Torch Model Zoo" and "open-source Py Torch implementation of YOLOv2" as baselines, but does not explicitly state that their own LLSQ code is open-source or provide a link to it.
Open Datasets Yes In this section, three sets of experiments on Cifar10, Image Net and Pascal VOC datasets are presented.
Dataset Splits No The paper mentions "val accuracy" in Figure 5, implying a validation set was used, but it does not provide specific details on the dataset splits (e.g., percentages, sample counts, or a detailed splitting methodology) for reproducibility.
Hardware Specification Yes Finally, we deploy the quantized network onto our specialized integer-only neural network accelerator... Our accelerators adopt the typical 2D systolic array architecture... The MAC unit in each PE consists of a 4-bit multiplier and a 16-bit accumulator... Finally, we implement the 8/4/2-bit integer neural network processors with Synopsys Design Compiler (DC) under the 40nm technology, clocked at 800MHz.
Software Dependencies No The paper mentions "Py Torch (Paszke et al., 2017)" for implementation and "Synopsys Design Compiler (DC)" for hardware implementation, but does not provide specific version numbers for these software components.
Experiment Setup Yes we train the reference network for 400 epochs using an initial learning rate of 2e-2. And for the training of the quantized network, we use a warmup learning rate scheduler in the first ten epochs with an initial learning rate of 2e-3. In all quantization experiments, the total training epochs are 100.