zennit.torchvision

Specialized Canonizers for models from torchvision.

Classes

`ResNetBasicBlockCanonizer`	Canonizer specifically for BasicBlocks of torchvision.models.resnet* type models.
`ResNetBottleneckCanonizer`	Canonizer specifically for Bottlenecks of torchvision.models.resnet* type models.
`ResNetCanonizer`	Canonizer for torchvision.models.resnet* type models.
`VGGCanonizer`	Canonizer for torchvision.models.vgg* type models.

class zennit.torchvision.VGGCanonizer[source]

Bases: SequentialMergeBatchNorm

Canonizer for torchvision.models.vgg* type models. This is so far identical to a SequentialMergeBatchNorm

class zennit.torchvision.ResNetBottleneckCanonizer[source]

Bases: AttributeCanonizer

Canonizer specifically for Bottlenecks of torchvision.models.resnet* type models.

__init__()[source]

static forward(self, x)[source]

Modified Bottleneck forward for ResNet.

Parameters:

self (torch.nn.Module:) – The module this abstract method will be bound to.
x (torch.Tensor:) – The input tensor.

Returns:

The resulting output tensor.

Return type:

torch.Tensor

class zennit.torchvision.ResNetBasicBlockCanonizer[source]

Bases: AttributeCanonizer

Canonizer specifically for BasicBlocks of torchvision.models.resnet* type models.

__init__()[source]

static forward(self, x)[source]

Modified BasicBlock forward for ResNet.

Parameters:

self (torch.nn.Module:) – The module this abstract method will be bound to.
x (torch.Tensor:) – The input tensor.

Returns:

The resulting output tensor.

Return type:

torch.Tensor

class zennit.torchvision.ResNetCanonizer[source]

Bases: CompositeCanonizer

Canonizer for torchvision.models.resnet* type models. This applies SequentialMergeBatchNorm, as well as add a Sum module to the Bottleneck modules and overload their forward method to use the Sum module instead of simply adding two tensors, such that forward and backward hooks may be applied.

__init__()[source]