Flower Image Classifier

Overview

A deep learning image classifier that identifies 102 species of flowers, built as the capstone project of the Udacity AI Programming with Python Nanodegree. The model uses VGG16 pretrained on ImageNet as a feature extractor, with a custom classifier trained on the Oxford 102 Flower Dataset.

Approach

Instead of training a neural network from scratch, I used transfer learning: the convolutional layers of VGG16 are frozen and act as a fixed feature extractor, while only the final classifier layers are trained on the flower dataset. This makes training fast and achieves strong accuracy even on a relatively small dataset.

• Backbone: VGG16 pretrained on ImageNet, all convolutional weights frozen
• Custom classifier: Linear(25088→512) → ReLU → Dropout(0.2) → Linear(512→102) → LogSoftmax
• Data augmentation: random resized crop and horizontal flip during training to reduce overfitting
• Normalization: ImageNet mean and std ([0.485, 0.456, 0.406] / [0.229, 0.224, 0.225])
• Loss / Optimizer: NLLLoss with Adam (lr=0.001), trained for 5 epochs on GPU

Command-Line Interface

Beyond the notebook, the project ships as two standalone CLI scripts:

• train.py flowers/ --arch vgg16 --epochs 5 --gpu trains on any ImageFolder dataset and saves a checkpoint
• predict.py image.jpg checkpoint.pth --top_k 5 --category_names cat_to_name.json loads a checkpoint and prints ranked predictions

What I Learned

• The full PyTorch training loop: ImageFolder → DataLoader → forward pass → loss → backward → optimizer step
• Why transfer learning works: the pretrained network already detects edges, textures, and shapes fine-tuning just maps those features to new classes
• How to save and restore model checkpoints including the class-to-index mapping
• Converting a Jupyter notebook into reusable CLI scripts with argparse
• The difference between training accuracy and validation accuracy, and why validation is the number that matters

Notebook

Full implementation below, from data loading through inference: