Learn practical skills, build real-world projects, and advance your career

Exploring SVHN with CNN

SVHN(Street View House Number) Dataset is a real world dataset which is obtained from the Google Street View Images. This Dataset contains over 600,000 digit images.
This Dataset is most likely big brother of MNIST Dataset which is "hello world " dataset for deep Learning.

This dataset full of some harder to recognise images of digits than MNIST. This dataset have two type of data,
first full images and another cropped images.

We are going to explore this dataset and build an deep Neural Network using CNN(Convolution Neural Network).

Dataset Overview

according to source ,

SVHN from stanford

http://ufldl.stanford.edu/housenumbers/

  1. 10 classes, 1 for each digit. Digit '1' has label 1, '9' has label 9 and '0' has label 10.

  2. 73257 digits for training, 26032 digits for testing, and 531131 additional, somewhat less difficult samples, to use as extra training data

**This dataset also available in Pytorch Torchvision datasets. In this project we use from torchvision dataset **

https://pytorch.org/docs/stable/torchvision/datasets.html

Before going to start let's install jovian so we can save our work

This project have following steps :

  • importing required libraries
  • Downloading an image dataset from torchvision dataset
  • exploring images of dataset
  • Understanding convolution and pooling layers
  • Creating a convolutional neural network (CNN) using PyTorch
  • Training a CNN from scratch and monitoring performance
  • Underfitting, overfitting and how to overcome them

Importing necessary libraries