Flower Species Image Classifier

Overview

This project is a deep learning-based image classification application developed in Python. It provides a robust command-line interface (CLI) for training a custom neural network classifier on top of pre-trained architectures (like ResNet18 and VGG13) and using it to predict flower species from raw images with high accuracy.

Key Features

• Transfer Learning: Leverages PyTorch’s pre-trained ImageNet models (resnet18 and vgg13) as powerful feature extractors for computer vision tasks.
• Custom Classifier Architecture: Modifies the base models by seamlessly swapping native fully connected layers with a dynamically configurable classifier to map specifically to 102 flower categories.
• CLI Utilities:
  • train.py: Allows users to define custom learning rates, hidden layer sizing, training epochs, model architecture, and GPU acceleration. It saves the training progress, weights, and configurations to a reusable checkpoint.
  • predict.py: Reconstructs the saved model using the dynamic checkpoint metadata to predict the Top-K most probable flower species for any given image input.
• Data Augmentation: Implements torchvision.transforms to augment training data—applying random rotations, resized crops, and horizontal flips—to significantly reduce model overfitting.

Technical Stack

• Language: Python 3
• Deep Learning Framework: PyTorch
• Data Processing & Visualization: NumPy, PIL (Pillow), Matplotlib, torchvision
• Core Concepts: Transfer Learning, Deep Neural Networks, Optimizer Tuning (Adam), Loss Functions (NLLLoss), GPU Acceleration (CUDA)

How It Works

1. Data Preprocessing: Images are loaded, resized, center-cropped to 224x224 pixels, converted to PyTorch tensors, and normalized against ImageNet standard metrics.
2. Training Phase: The model freezes the heavy convolutional base representations (preventing them from losing generalized visual knowledge) and only calculates gradients/backpropagates through the new custom classifier block.
3. Inference Phase: Input images are fed through the evaluation-ready model. Exponentiated LogSoftmax outputs are used to extract the highest probabilities and map them to human-readable flower names via a JSON dictionary.

Result

A highly accurate and configurable inference model that intelligently categorizes botanical species, showcasing strong foundational ML engineering practices and PyTorch framework knowledge.