Department of Electrical Engineering
Faculty Member:_________________ Date: _________________
Course/Section:____________________ Semester: _______________
CS-477 Computer Vision
Lab 12: Implementation of a Simple
Convolutional Neural Network
Name Reg. No PLO4-CLO4 PLO5-CLO5 PLO8-CLO6 PLO9-CLO7
Investigation Modern Tool Ethics Individual
Usage and Team
Work
5 marks 5 marks 5 marks 5 marks
Lab 9: Implementation of a Simple Convolutional
Neural Network
Objectives
Understand PyTorch’s Tensor library and neural networks at a high
level.
Introduction to CNN
Training a CNN classifier
Lab Report Instructions
All questions should be answered precisely to get maximum credit. Lab report must ensure following
items:
✔ Lab objectives
✔ Python codes
✔ Results (graphs/tables) duly commented and discussed
✔ Conclusion
Introduction to CNN
Convolutional Neural Network (CNN): A CNN is a type of deep neural network
designed to recognize and process visual data with a grid-like structure, such as
images. CNNs are particularly effective in image recognition, object detection, and
other computer vision tasks.
Grid-like Topology: Images can be thought of as a grid of pixels, where each pixel
represents the smallest unit of information. The arrangement of pixels creates a
grid-like structure, and CNNs leverage this spatial organization for more effective
feature extraction.
Digital Image: In the context of CNNs, digital images are represented as a grid of
pixels. Each pixel's position in the grid corresponds to a specific location in the
image, and the pixel value represents the color and intensity at that point. The
combination of pixel values across the grid forms the complete visual representation
of the image.
Binary Representation: While you mentioned a binary representation, it's important
to note that digital images typically use a range of values to represent colors. In the
RGB (Red, Green, Blue) color space, for example, each pixel is often represented by
three values corresponding to the intensity of each color channel. The values are
usually integers ranging from 0 to 255, or they could be normalized to the range [0,
1].
Pixel Values: Pixel values indicate the brightness and color information of the image.
In grayscale images, each pixel has a single value representing intensity. In color
images, each pixel has multiple values corresponding to the intensities of different
color channels.
CNNs use convolutional layers to automatically and adaptively learn spatial
hierarchies of features from input images. These layers contain filters that are
convolved with the input image to extract features such as edges, textures, and more
complex patterns. Pooling layers are often used to reduce the spatial dimensions of the
data, and fully connected layers integrate the learned features for final classification
or regression tasks.
In summary, CNNs are a powerful class of neural networks designed for processing
grid-like data, making them especially effective for tasks involving images and spatial
relationships within those images.
Figure 1: Source: https://pippin.gimp.org/image_processing/images/sample_grid_a_square.png
Convolutional Neural Network (CNN) typically consists of three fundamental layers:
a convolutional layer, a pooling layer, and a fully connected layer.
Figure 2: Source https://www.mathworks.com/videos/introduction-to-deep-learning-what-are-
convolutional-neural-networks--1489512765771.html
Convolution Layer
At the core of the CNN architecture lies the convolutional layer, bearing the primary
computational load of the network. This layer executes a dot product operation
between two matrices – one matrix comprises learnable parameters known as a
kernel, and the other matrix represents a confined section of the receptive field. While
the kernel is spatially smaller than the image, it possesses greater depth. For instance,
in an image with three (RGB) channels, the kernel's height and width are spatially
small, yet its depth extends across all three channels.
Illustration of Convolution Operation
In the forward pass, the kernel traverses the height and width of the image, generating
the image representation of the receptive region. This process yields a two-
dimensional representation known as an activation map, showcasing the kernel's
response at each spatial position within the image. The extent of movement of the
kernel is termed as the "stride."
For an input of size W x W x D, with Dout being the number of kernels, a spatial size
of F, a stride of S, and a specified amount of padding P, the size of the output volume
is determined by the following formula:
Pooling Layer
Following the convolutional layer, the pooling layer plays a crucial role in the
Convolutional Neural Network (CNN) architecture. Its function involves replacing
specific locations in the network's output by computing a summary statistic of the
neighboring outputs. This strategic substitution aids in diminishing the spatial size of
the representation, leading to a reduction in computational demands and the number
of weights. Notably, the pooling operation is applied independently to each slice of
the representation.
Various pooling functions exist, each offering distinct ways of summarizing
information within a neighborhood. Options include the average of the rectangular
neighborhood, the L2 norm of the rectangular neighborhood, and a weighted average
based on the distance from the central pixel. However, among these, max pooling
stands out as the most widely employed method. Max pooling entails selecting the
maximum output value from the neighborhood, providing a robust approach to
retaining essential features while reducing the dimensionality of the representation.
If we have an activation map of size W x W x D, a pooling kernel of spatial size F,
and stride S, then the size of output volume can be determined by the following
formula:
This will yield an output volume of size Wout x Wout x D.In all cases, pooling
provides some translation invariance which means that an object would be
recognizable regardless of where it appears on the frame.
Fully Connected Layer
Neurons in this layer have full connectivity with all neurons in the preceding and
succeeding layer as seen in regular FCNN. This is why it can be computed as usual by
a matrix multiplication followed by a bias effect. The FC layer helps to map the
representation between the input and the output.
Non-Linearity Layers
Given that convolution is inherently a linear operation and images exhibit significant
non-linearity, non-linearity layers are frequently inserted immediately after the
convolutional layer to impart non-linearity to the activation map.
Several types of non-linear operations are commonly employed, with notable
examples including:
Sigmoid:
The sigmoid non-linearity is expressed mathematically as σ(κ) = 1/(1+e¯κ). It takes a
real-valued number and compresses it into a range between 0 and 1. However, a
drawback of the sigmoid function is its tendency to generate gradients close to zero,
particularly at its tails. This property can hinder the backpropagation process,
potentially causing the gradient to become too small and impede effective weight
updates. Additionally, when the input data is consistently positive, the sigmoid may
produce outputs that are either entirely positive or entirely negative, leading to a zig-
zag dynamic in gradient updates for weights.
Tanh:
Tanh transforms a real-valued number to the range [-1, 1]. Similar to sigmoid, tanh
activations can saturate, but unlike sigmoid, its output is zero-centered.
ReLU (Rectified Linear Unit):
ReLU has gained widespread popularity in recent years. It computes the function ƒ(κ)
= max(0, κ), effectively thresholding the activation at zero. Compared to sigmoid and
tanh, ReLU offers faster convergence, accelerating the learning process by a factor of
six. Despite its advantages, ReLU does have a potential drawback during training. If a
large gradient flows through it, the neuron may undergo an update in such a way that
further updates become unlikely. This issue can be mitigated by carefully selecting an
appropriate learning rate
Watch following video to understand how CNN works
https://www.youtube.com/watch?v=HGwBXDKFk9I
Task 1: Convolution on Images ______________________________
Download a color image for this task. Write a function in python that takes as input
arguments: an image, a square filter (3x3, 5x5 etc.), padding size and number of
strides. The output of the function must be an image undergoing the convolution
operation on the input image.Implement convolution and showcase the result by
trying different filters, padding values and number of strides. Provide the code and at
least 4 screenshots for the final outputs.
### TASK 1 EXPLANATION STARTS HERE ###
### TASK 1 EXPLANATION ENDS HERE ###
### TASK 1 CODES START HERE ###
### TASK 1 CODES END HERE ###
### TASK 1 SCREENSHOTS START HERE ###
### TASK 1 SCREENSHOTS END HERE ###
Task 2: Simple CNN ____________________________________________
Build a simple convolutional neural network in PyTorch and train it to recognize
handwritten digits using the MNIST dataset (Training a classifier on the MNIST
dataset can be regarded as the hello world of image recognition).
### TASK 2 EXPLANATION STARTS HERE ###
### TASK 2 EXPLANATION ENDS HERE ###
### TASK 2 CODES START HERE ###
### TASK 2 CODES END HERE ###
### TASK 2 SCREENSHOTS START HERE ###
### TASK 2 SCREENSHOTS END HERE ###
Task 3: CNN _____________________________________________________
Build a simple convolutional neural network in PyTorch and train it to recognize
following fashion object using the fashion MNIST dataset which contains 10 classes
(T-shirt, Trouser, Pullover, Dress, Coat, Sandal, Shift, Sneaker, Bag, Ankle boot).
### TASK 3 EXPLANATION STARTS HERE ###
### TASK 3 EXPLANATION ENDS HERE ###
### TASK 3 CODES START HERE ###
### TASK 3 CODES END HERE ###
### TASK 3 SCREENSHOTS START HERE ###
### TASK 3 SCREENSHOTS END HERE ###
Helpful links
https://towardsdatascience.com/convolutional-neural-networks-explained-9cc5188c4939
https://www.youtube.com/watch?v=HGwBXDKFk9I
https://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html
https://colab.research.google.com/github/pytorch/tutorials/blob/gh-pages/_downloads/
4e865243430a47a00d551ca0579a6f6c/cifar10_tutorial.ipynb#scrollTo=PP9km88QkiZp
