Image Classification Convolutional Neural Network

Classifying images from the CIFAR-10 dataset using a convolutional neural network

This is Project 2 of Udacity's Deep Learning Nanodegree Foundation program. The aim of this project is go through the entire process of building a simple convolutional neural network (CNN) in TensorFlow.

Files

Jupyter Notebook or HTML

Steps

The steps included:

• Normalizing data and one-hot encoding labels
• Building TensorFlow Placeholders for the input, labels and dropout.
• Building a CNN function which does the matrix multiplication with normalized weights and bias, applies a rectified linear unit (ReLU) activation function and finally applies max pooling.
• Creating a flatten function that converts the incoming 4-D tensor into a 2-D flattened matrix.
• Creating a fully connected layer function in TensorFlow
• Building the ouput layer
• Defining the convolutional neural network architecture. This CNN has 4 convolutional layers, followed by 2 fully connected layers and finally the output layer.
• Setting the number of epochs, batch size and dropout (keep probability) for training on a single batch, and then on all 5 batches.

CNN Summary

Layer	Shape	Neurons
Input	32x32x3	3072
CNN Layer 1	16x16x48	12288
CNN Layer 2	8x8x192	12288
CNN Layer 3	4x4x384	6144
CNN Layer 4	2x2x512	2048
Fully Conn Layer 1	1x2048	512
Fully Conn Layer 2	1x512	128
Output	1x10	10

Dropout layer is applied before Fully Conn Layer 1, after flattening CNN Layer 4.

Training hyperparameters:

• Epochs: 600
• Batch size: 256
• Keep Probability: 0.5

Results

Test Accuracy: 57.29%