• Goals of this lab
• Introduction
  • process
• Intended Learning Outcomes
• Background
  • where to devise
  • Your Assignment
    • Im2Col
    • Process
    • Result

Goals of this lab

• im2col - 90%
• Results of labels - 10%

Introduction

im2col (image to column) is a technique commonly utilized in image processing and computer vision tasks, notably in the convolution operations of Convolutional Neural Networks (CNNs). The primary aim is to transform convolution operations into matrix multiplications, enabling the use of efficient matrix operation to enhance computational efficiency.

• pros

  • significantly enhances the computational efficiency of convolution operations by utilizing efficient matrix multiplication libraries.

• cons

  • Due to the unfolding of the input image into a matrix, introducing data duplication (especially when convolution windows overlap), the im2col method can consume substantial memory.

process

reference-1
reference-2

Intended Learning Outcomes

• Understanding how to transform the original convolution into 2-D matrix operations.

Background

Before we implement systolic array on convolution layer,we need to do transform the convolution operations matrix multiplication,because convolution operations cannot be applicable to matrix multiplication, your job in this lab is to do im2col in convolution layer for the next lab.

where to devise

 /tensorflow/lite/kernels/internal/reference/integer_ops/conv.h

Your Assignment

Im2Col

• Objective :

  • Let the convolution layer be transformed into 2D matrix operations.

• Input :

  • Model convolution layer

• Output :

  • 2-D Matrix

Process

First, we perform Im2col operation, followed by conventional matrix multiplication to compute ACC. This is due to the unavailability of Systolic Array in CFU-playground, which is the Lab5 assignment.

Result

Show the results of the labels of KWS model after quantized ,and see if they match those of lab2.

Figure 1: label8