Welcome to ADV-TRAIN’s documentation!

This is a framework built on top of pytorch to make machine learning training and inference tasks easier. Along with that it also enables easy dataset and network instantiations, visualize boundaries and more. This was created at the Nanoelectronics Research Laboratory at Purdue.

Indices and tables

Modules

class advtrain.utils.normalize.denormalize(mean, std, img_dimensions, device='cpu')

De-normalizes the input and provides a backpropable de-normalization function

Parameters

  • mean (list) – list of the mean for each channel

  • std (list) – list of the std for each channel

  • img_dimensions (list) – list [channels,h,w]

  • device (str) – cpu/cuda

Returns

Returns an object of normalize

class advtrain.utils.normalize.normalize(mean, std, img_dimensions, device='cpu')

Normalizes the input and provides a backpropable normalization function. It perfroms the channel wise normalization using mean and standard deviation

Parameters

  • mean (list) – list of the mean for each channel

  • std (list) – list of the std for each channel

  • img_dimensions (list) – list [channels,h,w]

  • device (str) – cpu/cuda

Returns

Returns an object of denormalize

advtrain.instantiate_model.instantiate_model(dataset='cifar10', num_classes=10, input_quant='FP', arch='resnet', dorefa=False, abit=32, wbit=32, qin=False, qout=False, suffix='', load=False, torch_weights=False, device='cpu')

Initializes/load network with random weight/saved and return auto generated model name ‘dataset_arch_suffix.ckpt’

Parameters

  • dataset – mnists/cifar10/cifar100/imagenet/tinyimagenet/simple dataset the netwoek is trained on. Used in model name

  • num_classes – number of classes in dataset.

  • arch – resnet/vgg/lenet5/basicnet/slpconv model architecture the network to be instantiated with

  • suffix – str appended to the model name

  • load – boolean variable to indicate load pretrained model from ./pretrained/dataset/

  • torch_weights – boolean variable to indicate load weight from torchvision for imagenet dataset

Returns

models with desired weight (pretrained / random ) model_name : str ‘dataset_arch_suffix.ckpt’ used to save/load model in ./pretrained/dataset

Return type

model

class advtrain.boundary_visualization.VisualizeBoundaries(framework=None, net=None, num_classes=10, num_channels=1, img_dim=32, device='cpu')
__init__(framework=None, net=None, num_classes=10, num_channels=1, img_dim=32, device='cpu')

This is a class to visualize the boundaries of a neural net

Parameters

  • framework (utils.Framework) – Object of Framework, If this is passed none of the other parameters are needed

  • net (nn.Module) – network whose decision boundaries are desired

  • num_classes (int) – Number of classes in the dataset

  • num_channels (int) – Number of channels in each image of the dataset

  • img_dim (int) – Assumes a square image, the height of the image in the dataset

  • device (str) – cpu/cuda device to perform the evaluation on

Returns

Returns an object of the VisualizeBoundaries

generate_decision_boundaries(images, labels, explore_range=40, use_random_dir=False)

Uses FGSM adversarial/random direction to obtain the boundaries of the nerual net around the specified images

Parameters

  • images (torch.Tensor) – Tensor of images, expected shape [batch_size, channels, image_dim, image_dim]

  • labels (torch.Tensor) – Ground truth/correct labels for the images, shape [batch_size,]

  • explore_range (int) – Specifes the range around the image explored by the code, and empirical value, also determines the size of the output images

  • values are, 2, 40, 80, 100. Larger the number more the computations hence it takes longer (recommended) –

  • use_random_dir (bool) – Use random direction rather than adversarial direction as the basis vectors for the image

Returns

Returns a Tensor of shape [batch_size, 3, explore_range, explore_range] which contains an RGB image of the boundaries around the input images.

show(image, decision_boundary, index)

Visualizes the image and the decision boundary around it

Parameters

  • image (torch.Tensor) – Tensor of images, expected shape [some non zero size, channels, image_dim, image_dim]

  • decision_boundary (torch.Tensor) – decision boundary a tensor of shape [some non zero size, 3, explore_range, explore_range] which contains an RGB image of the boundaries

  • index (int) – A value less than ‘some non zero size’ and proves an index of the image and decision bounday to display

Returns

Returns a Tensor of shape [batch_size, 3, explore_range, explore_range] which contains an RGB image of the boundaries around the input images.

class advtrain.utils.preprocess.preprocess

This class consists of a forward pass and backward pass function for preprocessing layer.

back_approx(x)

This function forward propogates through the preprocessing layer :param x: Input to be back propagated

Returns

The backward propagated input

forward(x)

This function forward propogates through the preprocessing layer Args:

Returns:

class advtrain.attack_framework.multi_lib_attacks.attack_wrapper(model, device, **params)

This is a LAL (Library Abstraction Layer) Wrapper around foolbox, advertorch and custom attack implementation

Example of attack_params {

‘lib’: ‘foolbox’, ‘attack’:’pgd’, ‘iterations’: 40, ‘epsilon’: 0.01, ‘stepsize’: 0.01, ‘bpda’: True, ‘random’: True ‘preprocess’: <pointer to quantize or hafltone or FP>, ‘custom_norm_func’: <pointer to normalize>, ‘targeted’: True,

}

Example of dataset_params {

‘mean’: [0.5, 0.5, 0.5], ‘std’: [0.5, 0.5, 0.5], ‘num_classes’: 10

}