Computer Vision

Bayesian CNNs (BCNNs) are created and then tested upon a range of computer vision tasks:

vision.action_args(args)[source]

Make GPU specific changes based upon the system’s setup and the user’s arguments. :param args: Argparser containing desired arguments. :return: Set of kwargs.

vision.build_parser()[source]

Setup parser.

vision.fgsm_test(model, adversary, args, test_loader)[source]

Evaluate a standard neural network’s performance when the images being evaluated have adversarial attacks inflicted upon them.

Parameters:
  • model (Torch Model) – A trained CNN
  • adversary – An adversarial object for which attacks can be crafted
  • args – Arguments object
  • test_loader – Testing dataset
vision.load_data(args, kwargs)[source]

Load in the MNIST dataset are setup batching.

Parameters:
  • args – Argparser object
  • kwargs – GPU specific kwargs
Returns:

Train and Test datasets
vision.mcdropout_test(model, args, test_loader, stochastic_passes=100)[source]

Carry out basic tests on the BCNN.

Parameters:
  • model (Torch Model) – A trained BCNN
  • args – Arguments object
  • test_loader – Testing dataset
  • stochastic_passes (int) – Number of stochastic passes to maker per image
vision.test(model, args, test_loader)[source]

Test a CNN performance

Parameters:
  • model (Torch Model) – A trained BCNN
  • args – Arguments object
  • test_loader – Testing dataset
vision.train(model, opt, epoch, args, train_loader)[source]

Train a model.

Parameters:
  • model (Torch Model) – The model to be trained
  • opt – The optimiser to be used during training
  • epoch (int) – Number of epochs to be used in training. Note, there is no early stopping in place.
  • args – Argparser containing several user defined arguments.
  • train_loader – Training data
Returns:

Trained model
vision.uncertainty_test(model, args, test_loader, stochastic_passes=100)[source]

Measure the uncertainty values calculated by the BCNN as the images of the MNIST dataset are rotated through 180 degrees.

Parameters:
  • model (Torch Model) – A trained BCNN
  • args – Arguments object
  • test_loader – Testing dataset
  • stochastic_passes (int) – Number of stochastic passes to maker per image