Patch2Vec [WNNA-2022 site] [Poster]
This is an official C# implementation for "Patch2Vec: a simple and efficient convolution algorithm for mobile neural networks" WNNA-2022 poster.
The meaning of most fast convolution algorithms, such as im2col or im2row, involves bringing the convolution to matrix multiplication, which allows optimizing memory access operations by using the processor cache. However, such methods either require a buffer for srcC * kernelY * kernelH * dstH * dstW elements, which is extremely irrational. The proposed patch2vec method unwraps each patch of the input image on the fly, and then applies all convolution filters to it. This implementation is not inferior in efficiency to classical algorithms like im2col, and in practice even surpasses them. The buffer for this algorithm will have the size of srcC * kernelY * kernelX, which is much smaller than in the case of similar methods. Moreover, patch2vec does not impose restrictions on the convolution parameters, unlike, for example, the Shmuel Vinograd method. The proposed algorithm is difficult to fit into classical machine learning frameworks due to the fact that they are focused on using GEMM as the core. Pure C#-based implementations make it easy to do this. For more detailed description, please see Details.md
- Lavin, Scott Gray: Fast Algorithms for Convolutional Neural Networks, arXiv preprint arXiv:1509.09308, 2015.
- Anton V. Trusov, Elena E. Limonova, Dmitry P. Nikolaev and Vladimir V. Arlazarov: p-im2col: Simple Yet Efficient Convolution Algorithm With Flexibly Controlled Memory Overhead, IEEE Access PP(99):1-1 (2021)