Flower Image Classifier (Udacity AIPND)

Command-line (CLI) app to train an image classifier on the Udacity Flowers dataset and predict a flower name from a new image using transfer learning.

TL;DR (try this first):

• No-training quickstart (recommended): download a pretrained checkpoint from GitHub Releases and run predict.py.
• Train your own checkpoint (CPU-only can be slow): python train.py flowers --epochs 2
• Predict a label from an image: python predict.py flowers/test/3/image_06634.jpg save_directory/checkpoint.pth --top_k 5
• Uses a pretrained backbone (default: ResNet-50) + a new classifier head.
• Checkpoint integrity: compare file SHA-256 with the value in the Release notes.

---

Demo (what you should see)

After training, run:

python predict.py flowers/test/3/image_06634.jpg save_directory/checkpoint.pth --top_k 5

Example output (format may vary slightly):

Path to image: flowers/test/3/image_06634.jpg
Path to checkpoint: save_directory/checkpoint.pth
Number of top K classes: 5
Path to category names file: cat_to_name.json
GPU: False

Prediction (name): cape flower
Probability: 0.10262521356344223

Top classes (names): ['cape flower', 'cyclamen', 'lotus lotus', 'magnolia', 'columbine']
Top probabilities: [0.10262521356344223, 0.07787298411130905, 0.05228663235902786, 0.048569660633802414, 0.0458136685192585]

Smoke test

python scripts/smoke_test.py

---

What’s in this repo

• train.py — trains a classifier and saves a checkpoint
• predict.py — loads a checkpoint and predicts top-k classes for an input image
• helper.py — training / preprocessing / checkpoint helpers
• get_input_args.py — CLI argument definitions
• cat_to_name.json — mapping from class id → flower name
• assets/ — screenshots / example images (optional)
• notebooks/ — project notebook (reference / exploration)
• scripts/ — script (smoke test, others.)

Not included: the dataset folder flowers/ (it is ignored by git).

---

Approach

This project uses transfer learning:

1. Load a pretrained convolutional neural network (CNN) backbone (default: ResNet-50).
2. Replace the final classification layer with a new fully-connected classifier head for 102 flower classes.
3. Freeze (or mostly freeze) backbone parameters and train the classifier head on the flowers dataset.
4. Save a checkpoint so you can run fast predictions later without retraining.

---

Quickstart — clone → predict (no training)

1) Create an environment + install

Option A (recommended): pip + virtual environment

python -m venv .venv

# Windows (Git Bash)
source .venv/Scripts/activate

# macOS / Linux
# source .venv/bin/activate

pip install --upgrade pip
pip install -r requirements.txt

Option B: Conda

conda create -n flower_image_classifier python=3.11 -y
conda activate flower_image_classifier
pip install --upgrade pip
pip install -r requirements.txt

2) Dataset layout (expected)

Place the Udacity Flowers dataset in the repo root:

flowers/
  train/
  valid/
  test/

3) Download the pretrained checkpoint (recommended)

Download checkpoint.pth from the repo’s GitHub Release (e.g. v1.0.0) and save it to:

save_directory/checkpoint.pth

Now you can run predictions without training:

python predict.py flowers/test/3/image_06634.jpg save_directory/checkpoint.pth --top_k 5

If you don’t have the dataset locally (flowers/), use any local image path instead of flowers/test/....

4) Train (optional — creates your own checkpoint)

Minimal training run (2 epochs):

python train.py flowers --epochs 2

Defaults:

• checkpoint folder: save_directory/
• checkpoint file: save_directory/checkpoint.pth
• architecture: resnet50

A “more realistic” training example:

python train.py flowers   --arch resnet50   --learning_rate 0.003   --hidden_units 512 256   --dropout 0.2   --epochs 3

5) Predict (top-5)

python predict.py flowers/test/3/image_06634.jpg save_directory/checkpoint.pth --top_k 5

---

GPU usage (optional)

If you have a CUDA-capable GPU and a compatible PyTorch install, add --gpu:

python train.py flowers --epochs 3 --gpu
python predict.py flowers/test/3/image_06634.jpg save_directory/checkpoint.pth --top_k 5 --gpu

If no GPU is available, the code runs on CPU.

---

Results

Setting	Value
Backbone	ResNet-50
Epochs	5
Learning rate	0.0005
Hidden units	512
Dropout	0.2
Validation accuracy	0.920
Test accuracy	0.878

---

CLI reference

Training help

python train.py -h

Common arguments:

• data_dir (positional): dataset folder (e.g. flowers)
• --save_dir: folder where checkpoints are saved (default: save_directory/)
• --arch: pretrained architecture (default: resnet50)
• --learning_rate
• --hidden_units (space-separated list, e.g. --hidden_units 512 256)
• --dropout
• --epochs
• --gpu

Prediction help

python predict.py -h

Common arguments:

• path_to_image (positional)
• path_to_checkpoint (positional)
• --top_k
• --category_names (default: cat_to_name.json)
• --gpu

---

License

This project is licensed under the MIT License — see the LICENSE file for details.