Setup

1. conda create -n how-much-data-experiments python=3
2. pip install tensorflow numpy pandas (tested with TensorFlow 1.12 & 1.8.0)
3. curl -LO http://download.tensorflow.org/example_images/flower_photos.tgz
4. tar xzf flower_photos.tgz
5. pip install tensorflow-hub (tested with version 0.1.1)
6. pip install six scipy Pillow matplotlib scikit-image opencv-python imageio Shapely
7. pip install imgaug
8. pip install vega
9. pip install jupyter
10. pip install tqdm

Running one-off experiments

Open "Experiment.ipynb" in a notebook environment. Run through the cells in order; modifying the parameters at the top as you wish. There's a little busywork you'll need to do if you change foldernames.

Run this notebook on Google Colaboratory.

Running everything

./downsample_data.py  # Runs fast
./augment_data.py     # (Optional) Takes an hour or so 
./train_everything.sh
./infer_folders.sh

Reviewing

Run the Analysis notebook in ./notebooks.