Developing an Image Classifier with Deep Learning

This project demonstrates how to make use of pre-trained deep learning models, and then enhancing them with custom feed-forward networks. This is implemented in Python3 using PyTorch, and supports GPU-based training. The project is trained and tested on an image dataset of flowers composed of 102 variants, but the project supports the ability to train a classifier on any image dataset.

Currently Supported Models:

A range of pre-trained models are supported, though you can easily add more as required. Currently supported models include:

• AlexNet
• ResNet18
• SqueezeNet 1.0
• SqueezeNet 1.1

Project structure

1. Interactive Jupyter Notebook

This highlights the data processing and model training components, with detailed interactive cells to explain the process. The SqueezeNet 1.1 model is used, and the number of output features are configured to match the number of classes in the input dataset.

2. Console Application

This builds off the components demonstrated in the notebook and provides a highly customisable application allowing for modification of:

• Input dataset (directory structure must match)
• Input class mappings
• Model architecture
• Number of hidden units in model feed-forward network
• Learning rate
• Training epochs
• Loading pre-trained model checkpoint
• Number of predicted class results
• GPU-based training / prediction

The ResNet18 model implements a custom feed-forward network using linear, relu, and dropout layers that can be modified as required.

Package Requirements

Versions are currently agnostic. Install a combination is simply compatible:

• Python3
• PyTorch (with CUDA if GPU is available)
• NumPy
• Pandas
• Matplotlib
• Seaborn

Console Application Arguments:

The console application can be used to either train a model, or analyse a specified input.

Train:

Basic usage:

python train.py [input_data_directory]

Optional arguments:

Argument	Default	Description
--save_dir		checkpoint save directory
--arch	squeezenet1_1	model architecture
--learning_rate	0.001	optimizer learn rate
--hidden_units	256	number of features in feed-forward layer
--epochs	10	number of training iterations
--gpu		use GPU for training model

Example argument usage:

python train.py flowers --arch resnet18 --learning_rate 0.001 --epochs 10 --gpu

Predict:

Basic usage:

python predict.py [path_to_image] [saved_checkpoint]

Optional arguments:

Argument	Default	Description
--top_k	5	number of top predictions to return
--category_names	cat_to_name.json	json file of class mappings
--gpu		use GPU for training model

Example argument usage:

python predict.py flowers/test/1/image_06743.jpg checkpoint_squeezenet1_1.pth --gpu --top_k 10 --category_names cat_to_name.json