Towards Efficient and Accurate Winograd Convolution via Full Quantization

Authors: Tianqi Chen, Weixiang Xu, Weihan Chen, Peisong Wang, Jian Cheng

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate the effectiveness of our method, e.g., with 8-bit quantization and a tile size of 6, our method outperforms the previous Winograd PTQ method by 8.27% and 5.38% in terms of the top-1 accuracy on Res Net-18 and Res Net-34, respectively.
Researcher Affiliation Collaboration Tianqi Chen1,2, Weixiang Xu1, Weihan Chen1, Peisong Wang1,2,3 , Jian Cheng1,2,3,4 1Institute of Automation, Chinese Academy of Sciences 2School of Artificial Intelligence, University of Chinese Academy of Sciences 3AIRIA 4Maicro.ai {chentianqi2023, xuweixiang2018,chenweihan2018}@ia.ac.cn, {peisong.wang,jcheng}@nlpr.ia.ac.cn
Pseudocode No The paper does not contain a block or figure explicitly labeled 'Pseudocode' or 'Algorithm'.
Open Source Code No The paper does not provide any explicit statement about open-source code availability or a link to a code repository for the described methodology.
Open Datasets Yes Experiments demonstrate the effectiveness of our method, e.g., with 8-bit quantization and a tile size of 6, our method outperforms the previous Winograd PTQ method by 8.27% and 5.38% in terms of the top-1 accuracy on Res Net-18 and Res Net-34, respectively.
Dataset Splits No The paper mentions 'Image Net validation accuracy' in Table 6, implying the use of a validation set, and states 'we use 1024 unlabeled images' for calibration, but it does not provide specific details on dataset split percentages or counts for training, validation, or test sets.
Hardware Specification Yes All our experiments are conducted on NVIDIA Ge Force RTX 3090 24GB GPU servers and last for several hours.
Software Dependencies No The paper mentions software components like 'Adam [39] optimizer' and refers to existing quantization methods (e.g., LSQ [14], Adaround [18]) but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes Following BRECQ [19], we use 1024 unlabeled images and Adam [39] optimizer with 20k iterations and a batch size of 32. Based on it, we further verify our method on the strong baseline. The learning rates of A, B, and G are set to (1e-4, 1e-4, 5e-4) by default...