[2412.19867] Data-Free Group-Wise Fully Quantized Winograd Convolution via Learnable Scales

View the PDF file for the paper entitled “Winograd in full quantitative

PDF HTML (experimental) view

a summary:Despite the revolutionary breakthroughs of the text spreading models to the image on a large scale for the complex vision and the estuary tasks, its calculations and very high storage limits the ability to use them. The estimation of the proliferation forms in modern works has been explored to reduce the costs of calculating and using the frequency range of memory. To increase the improvement time, rapid wrapping algorithms such as Winograd can be used for spraining layers, which are a large part of the calculations in spread models. However, the great quality loss in the full quantitative Winograd using the pantying pantying methods of beloved, along with the complexity and cost of the matrices in the transfer of Winograd for such large models to restore quality, which makes them inappropriate for large -scale basis models. Motivated by the presence of a large group of values ​​in it, we verify the impact of the exact measurement of the granular group in the prevalence models. While the collective quantitative measurement can significantly deal with quantitative wrapping, it is struggling to deal with a large imbalance in a large part of the Winograd field account. To reduce the differences in the range in the WinogRAD field, we only suggest enjoying the scale parameters for Winograd transfer matrices without using any field training data. Since our way does not depend on any training data, the performance of the circulation of quantitative prevalences is safely guaranteed. For the task of generating the text to an image, the 8 -bit spread model with Winograd provides semi -losing quality (FID and Clip) compared to the full accuracy model. To classify images, our way out of the modern Winograd PTQ method by 1.62 % and 2.56 % in imagenet Top-1 resolution is on Resnet18 and Resnet-34, respectively, with Winograd F (6, 3).