Deeper Networks for Image Classification

• Code by Kaviraj Gosaye QMUL
• Student ID: 220575371

Description 📜

This repository is an analysis of Convolutional Neural Networks (CNNs) on two popular datasets to measure performance of deep networks for image classification tasks. All four models are trained on MNIST and further analysis for CIFAR10 is done using only ResNet50 and MobileNetV2.

Models 🚀

• InceptionV3
• VGG16
• ResNet50
• MobileNetV2

Datasets 📚

• MNIST
• CIFAR10

Tech and Version 🛠️

• Python 3.12.1
• PyTorch 2.2.0
• Matplotlib 3.8.2
• Torchinfo 1.8.0

Results 📈

MNIST

Model	Size (MB)	Trainable parameters	Train Loss (4 s.f.)	Test Loss (4 s.f.)	Train Accuracy (4 s.f.)	Test Accuracy (4 s.f.)	Precision	Recall
VGG-16	549.95	134,308,810	0.01964	0.03202	99.47	98.93	0.99	0.99
Inception V3	92.32	21,805,226	0.02872	0.02745	99.20	98.53	0.99	0.99
ResNet50	105.99	23,519,690	0.01447	0.03467	99.58	98.61	0.99	0.99
MobileNetV2	9.11	1,814,346	0.04369	0.09918	98.61	94.06	0.94	0.94

CIFAR10

Model	Size (MB)	Trainable parameters	Train Loss (4 s.f.)	Test Loss (4 s.f.)	Train Accuracy (4 s.f.)	Test Accuracy (4 s.f.)	Precision	Recall
ResNet50	108.61	23,520,842	0.02102	1.184	99.27	73.58	0.74	0.74
MobileNetV2	73.96	1,814,922	0.04439	1.318	98.42	72.30	0.72	0.72