Modules

Bayes Module

class torchbnn.modules.module.BayesModule

Applies Bayesian Module Currently this module is not being used as base of bayesian modules because it has not many utilies yet, However, it can be used in the near future for convenience.

freeze()

Sets the module in freezed mode. This has effect on bayesian modules. It will fix epsilons, e.g. weight_eps, bias_eps. Thus, bayesian neural networks will return same results with same inputs.

unfreeze()

Sets the module in unfreezed mode. This has effect on bayesian modules. It will unfix epsilons, e.g. weight_eps, bias_eps. Thus, bayesian neural networks will return different results even if same inputs are given.

Bayes Linear

class torchbnn.modules.linear.BayesLinear(prior_mu, prior_sigma, in_features, out_features, bias=True)

Applies Bayesian Linear

Parameters:

• prior_mu (Float) – mean of prior normal distribution.
• prior_sigma (Float) – sigma of prior normal distribution.

Note

other arguments are following linear of pytorch 1.2.0.

https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/linear.py

extra_repr()

Overriden.

forward(input)

Overriden.

Bayes Conv

class torchbnn.modules.conv.BayesConv2d(prior_mu, prior_sigma, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, padding_mode='zeros')

Applies Bayesian Convolution for 2D inputs

Parameters:

• prior_mu (Float) – mean of prior normal distribution.
• prior_sigma (Float) – sigma of prior normal distribution.

Note

other arguments are following conv of pytorch 1.2.0.

https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/conv.py

forward(input)

Overriden.

Bayes Batchnorm

class torchbnn.modules.batchnorm.BayesBatchNorm2d(prior_mu, prior_sigma, num_features, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)

Applies Bayesian Batch Normalization over a 2D input

Parameters:

• prior_mu (Float) – mean of prior normal distribution.
• prior_sigma (Float) – sigma of prior normal distribution.

Note

other arguments are following batchnorm of pytorch 1.2.0.

https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/batchnorm.py

BKLLoss

class torchbnn.modules.loss.BKLLoss(reduction='mean', last_layer_only=False)

Loss for calculating KL divergence of baysian neural network model.

Parameters:

• reduction (string, optional) – Specifies the reduction to apply to the output: 'mean': the sum of the output will be divided by the number of elements of the output. 'sum': the output will be summed.
• last_layer_only (Bool) – True for return only the last layer’s KL divergence.

forward(model)

Parameters:model (nn.Module) – a model to be calculated for KL-divergence.