Experiments for the masters thesis of Martin Thoma

Every experiment is a YAML file. See the experiments folder for some examples.

• matplotlib
• Tensorflow 1.0
• Keras 2.0 (adjusted preprocessing/image.py, see misc directory)
• seaborn

If you get TypeError: __init__() got an unexpected keyword argument 'hsv_augmentation' you didn't adjust the image.py. Just copy the one in the misc folder to python -c "import keras.preprocessing.image as k;print(k.__file__)"

• ./run_training.py -f experiments/cifar100_baseline.yaml: Train a model. Downloads everything by its own
• ./analyze_training.py -d artifacts/cifar100_baseline: Show some training statistics
• ./inference_timing.py -f experiments/cifar100_baseline.yaml: Run inference on a given trained model and measure the time
• ./eval_ensemble.py -f ensemble/cifar100_baseline.yaml: Evaluate an ensemble
• ./visualize.py --cm artifacts/cifar100_root/cm-test.json: Confusion matrix optimization
• ./create_cm.py --indices cm.indices.pickle -f experiments/cifar100_root-g5.yaml

Measuring inference time needs about 2 minutes:

$ ./inference_timing.py -f experiments/cifar10_baseline.yaml

Measuring training time takes about 70 minutes:

$ ./run_training.py -f experiments/cifar10_baseline.yaml