Flower Recognition Project

Fine-tuning an ImageNet-pretrained ResNet-18 on the Oxford 102 Category Flower Dataset.

Setup

This project uses Python 3.10+. There are two ways to install the dependencies.

Option 1: uv (recommended)

If you have uv installed, it reads pyproject.toml and creates the environment automatically:

uv sync

Then run any command in this README with uv run in front, e.g. uv run python code/prepare_data.py.

Option 2: pip

pip install torch torchvision numpy matplotlib pillow scipy pandas

A GPU is recommended but not required. The code runs on CPU as well, just slower.

Preparing the data

The dataset is not downloaded automatically, you must place it in the data/ folder yourself. Download the Oxford 102 Flowers dataset and arrange it like this:

data/
├── jpg/                # all 8189 flower images (image_00001.jpg ... image_08189.jpg)
├── setid.mat           # the official train/val/test split
└── imagelabels.mat     # the per-image class labels

jpg/, setid.mat and imagelabels.mat all come from the official dataset download. Once they are in place, build the CSV split manifests:

python code/prepare_data.py

This creates data/train.csv, data/val.csv and data/test.csv. The training code reads images from data/jpg/ and the split from those CSV files, so this step must be done before any training.

How to run

There are two ways to run everything.

Option A: Jupyter notebook (easiest)

Open run_all.ipynb and run the cells in order. The notebook prepares the data, runs every experiment, evaluates the checkpoints and builds the figures, so you can watch the whole pipeline step by step. The data must still be placed in data/ first, as described above.

Option B: Command line

Run the scripts directly. The dataset must already be prepared (see above).

Baseline — pretrained ResNet-18

python code/train.py --model resnet18 --pretrained \
    --epochs 25 --lr-head 1e-3 --lr-backbone 1e-4 \
    --run-name resnet18_baseline

Hyperparameter sweep — new-layer learning rate

python code/train.py --model resnet18 --pretrained --lr-head 1e-2 --lr-backbone 1e-4 --epochs 25 --run-name sweep_lr1e-2
python code/train.py --model resnet18 --pretrained --lr-head 1e-3 --lr-backbone 1e-4 --epochs 25 --run-name sweep_lr1e-3
python code/train.py --model resnet18 --pretrained --lr-head 1e-4 --lr-backbone 1e-4 --epochs 25 --run-name sweep_lr1e-4

From scratch — no pretraining

python code/train.py --model resnet18 --no-pretrained \
    --epochs 60 --lr-head 1e-2 \
    --run-name resnet18_scratch

Attention — ResNet-18 with SE blocks

python code/train.py --model resnet18-se --pretrained --epochs 25 --run-name resnet18_se

Evaluate a saved checkpoint

python code/evaluate.py --checkpoint checkpoints/resnet18_baseline.pth --model resnet18

Build all figures for the report

python code/plot_results.py

Each training run writes a per-epoch log to logs/<run-name>.csv and saves the best model to checkpoints/<run-name>.pth.