nolearn-utils

Iterators and handlers for nolearn.lasagne to allow efficient real-time image augmentation and training progress monitoring

Real-time image augmentation

• ShuffleBatchIteratorMixin to shuffle training samples
• ReadImageBatchIteratorMixin to transform image file path into image as color or as gray, and with specified image size
• RandomFlipBatchIteratorMixin to randomly (uniform) flip the image horizontally or verticaly
• AffineTransformBatchIteratorMixin to apply affine transformation (scale, rotate, translate) to randomly selected images from the given transformation options - BufferedBatchIteratorMixin to perform transformation in another thread automatically and put the result in a buffer (default size = 5)
• LCNBatchIteratorMixin to perform local contrast normalization to images
• MeanSubtractBatchIteratorMixin to subtract samples from the pre-calculated mean

Example of using iterators as below:

train_iterator_mixins = [
    ShuffleBatchIteratorMixin,
    ReadImageBatchIteratorMixin,
    RandomFlipBatchIteratorMixin,
    AffineTransformBatchIteratorMixin,
    BufferedBatchIteratorMixin,
]
TrainIterator = make_iterator('TrainIterator', train_iterator_mixins)

train_iterator_kwargs = {
    'buffer_size': 5,
    'batch_size': batch_size,
    'read_image_size': (image_size, image_size),
    'read_image_as_gray': False,
    'read_image_prefix_path': './data/train/',
    'flip_horizontal_p': 0.5,
    'flip_vertical_p': 0,
    'affine_p': 0.5,
    'affine_scale_choices': np.linspace(0.75, 1.25, 5),
    'affine_translation_choices': np.arange(-3, 4, 1),
    'affine_rotation_choices': np.arange(-45, 50, 5)
}
train_iterator = TrainIterator(**train_iterator_kwargs)

The BaseBatchIterator is also modified from nolearn.lasagne to provide a progress bar for training process for each iteration

Handlers

• EarlyStopping stops training when loss stop improving
• StepDecay to gradually reduce a parameter (e.g. learning rate) over time
• SaveTrainingHistory to save training history (e.g. training loss)
• PlotTrainingHistory to plot out training loss and validation accuracy over time after each iteration with matplotlib

Examples

Example code requires scikit-learn

MNIST

example/mnist/train.py should produce a model of about 99.5% accuracy in less than 50 epoch.

MNIST data can be downloaded from Kaggle.

CIFAR10

CIFAR10 images can be downloaded from Kaggle. Place the downloaded data as follows:

examples/cifar10
├── data
│   ├── train
│   |   ├── 1.png
│   |   ├── 2.png
│   |   ├── 3.png
│   |   ├── ...
│   └── trainLabels.csv
└── train.py

example/cifat10/train.py should produce a model at about 85% accuracy at 100 epoch. Images are read from disk and augmented at training time (from another thread)

TODO

• Embarrassingly parallelize transform

License

MIT & BSD