Deep Learning Image Classifier for Flower Species

Project Overview

This project implements a command line application for training and predicting flower species using deep learning. The model trains on a dataset of images and uses transfer learning from either ResNet18 or VGG13 architectures to achieve high accuracy in flower classification.

Getting Started

Prerequisites

• Python 3
• PyTorch
• torchvision
• PIL
• numpy
• matplotlib

Installation

1. Clone this repository
2. Install required packages:

   pip install torch torchvision PIL numpy matplotlib

Usage

Training the Model

Train a new network on a data set with train.py:

python train.py data_directory --save_dir save_directory --arch "vgg13" --learning_rate 0.01  --epochs 20 --gpu

Options:

• data_directory: Path to training data (required)
• --save_dir: Directory to save checkpoints (default: current directory)
• --arch: Model architecture - "vgg13" or "resnet18" (default: "resnet18")
• --learning_rate: Set learning rate (default: 0.003)
• --hidden_units: Hidden units for classifier (default: [1024, 512, 256])
• --epochs: Number of training epochs (default: 3)
• --gpu: Use GPU for training if available (default: False)

Making Predictions

Predict flower names from images using predict.py:

python predict.py /path/to/image checkpoint --top_k 3 --category_names cat_to_name.json --gpu

Options:

• Image path (required)
• Checkpoint path (required)
• --top_k: Return top K predictions (default: 1)
• --category_names: Use category names JSON file (default: cat_to_name.json)
• --gpu: Use GPU for inference if available (default: False)

Project Structure

├── train.py           # Script for training the network
├── predict.py         # Script for making predictions
├── cat_to_name.json   # Mapping of categories to flower names  
└── README.md

Model Architecture

The project offers two pre-trained model architectures:

1. ResNet18 (default)

   • Pretrained on ImageNet
   • Custom classifier added with configurable hidden units
   • Dropout added for regularization

2. VGG13

   • Pretrained on ImageNet
   • Modified classifier with configurable hidden units
   • Dropout layers to prevent overfitting

Data Processing

• Images are loaded using torchvision's ImageFolder
• Training transformations include:
  • Random rotation
  • Random resizing & cropping
  • Random horizontal flips
  • Normalization
• Validation/Testing transformations:
  • Resizing
  • Center crop
  • Normalization

Training Process

1. Loads pretrained model and freezes feature parameters
2. Adds new classifier for flower categories
3. Trains using:
   • Adam optimizer
   • NLLLoss criterion
   • Learning rate scheduler
4. Validates accuracy during training
5. Saves checkpoint with model & optimizer state

Checkpointing

Saved checkpoints include:

• Model state dict
• Optimizer state dict
• Class to index mapping
• Epoch completed
• Architecture used
• Hidden layer units
• Learning rate

Acknowledgments

• Project completed as part of the Udacity AI Programming with Python Nanodegree
• Architecture implementations based on torchvision models