Optimizing Conv in Caffe

机器学习

Solution: Converting convolution to GEMM. Done in 2 steps:

• do an im2col to convert the image to a matrix,
• call GEMM to do the actual computation.

Convert input data $C \times H \times W$ matrix to feature matrix $(H \times W) \times (C \times K \times K)$. Do GEMM with $C_{\mathrm{out}} \times (C \times K \times K)$ filter matrix. Output size: $C_{\mathrm{out}} \times H \times W$.