Models

Feature Extractors

Basic Conv Nets

class topaz.model.features.basic.BasicConv(layers, units, unit_scaling=1, dropout=0, bn=True, pooling=None, activation=<class 'torch.nn.modules.activation.PReLU'>)

A generic convolutional neural network scaffold.

fill(stride=1)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

unfill()

class topaz.model.features.basic.Conv127(units, **kwargs)

training: bool

class topaz.model.features.basic.Conv63(units, **kwargs)

training: bool

DenseNets

class topaz.model.features.densenet.DenseBlock(nin, ng)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class topaz.model.features.densenet.DenseLayer(nin, ng, bn_size=4, dilation=1, dropout=0, bn=True)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class topaz.model.features.densenet.DenseNet(init_units=64, ng=32, bn_size=4, fc_units=1000, num_layers=12, dropout=0, bn=True)

fill(stride=1)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

unfill()

class topaz.model.features.densenet.MultiscaleDenseNet(base_units=64, ng=48, num_blocks=4)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

MSNets

class topaz.model.features.msnet.DenseBlock(nin, ng)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class topaz.model.features.msnet.DenseLayer(nin, ng, bn_size=4, dilation=1, dropout=0, bn=True)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class topaz.model.features.msnet.DenseNet(init_units=64, ng=32, bn_size=4, fc_units=1000, num_layers=12, dropout=0, bn=True)

fill(stride=1)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

unfill()

class topaz.model.features.msnet.MultiscaleDenseNet(base_units=64, ng=48, num_blocks=4)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class topaz.model.features.msnet.PyramidNet(base_units=64, scales=[2, 4, 8, 16], pooling='max')

fill(*args, **kwargs)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

unfill(*args, **kwargs)

class topaz.model.features.msnet.RMSNet10(units=[64, 64, 128, 128, 256])

fill(*args, **kwargs)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

unfill(*args, **kwargs)

class topaz.model.features.msnet.RMSNet7(units=[64, 64, 128, 128, 256])

fill(*args, **kwargs)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

unfill(*args, **kwargs)

class topaz.model.features.msnet.ResidMultiScaleBlock(nin, ng, nout, scales=[1, 2, 4], pooling='max', padding=False)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class topaz.model.features.msnet.ResidMultiScaleNet(units=[64, 64, 64, 128, 128, 128, 128, 128, 128, 256, 256, 256], scales=[1, 2, 4])

fill(*args, **kwargs)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

unfill(*args, **kwargs)

ResNets

class topaz.model.features.resnet.BasicConv2d(nin, nout, kernel_size, dilation=1, stride=1, bn=False, activation=<class 'torch.nn.modules.activation.ReLU'>)

fill(stride)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

set_padding(pad)

training: bool

unfill()

class topaz.model.features.resnet.MaxPool(kernel_size, stride=1)

fill(stride)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

set_padding(pad)

training: bool

unfill()

class topaz.model.features.resnet.ResNet(*args, **kwargs)

fill(stride=1)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

set_padding(pad)

training: bool

unfill()

class topaz.model.features.resnet.ResNet16(*args, **kwargs)

make_modules(units=[32, 64, 128], bn=True, dropout=0.0, activation=<class 'torch.nn.modules.activation.ReLU'>, pooling=None, **kwargs)

training: bool

class topaz.model.features.resnet.ResNet6(*args, **kwargs)

make_modules(units=[32, 64, 128], bn=True, dropout=0.0, activation=<class 'torch.nn.modules.activation.ReLU'>, **kwargs)

training: bool

class topaz.model.features.resnet.ResNet8(*args, **kwargs)

make_modules(units=[32, 64, 128], bn=True, dropout=0.0, activation=<class 'torch.nn.modules.activation.ReLU'>, pooling=None, **kwargs)

training: bool

class topaz.model.features.resnet.ResidA(nin, nhidden, nout, dilation=1, stride=1, activation=<class 'torch.nn.modules.activation.ReLU'>, bn=False)

fill(stride)

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

set_padding(pad)

training: bool

unfill()

Linear Classifier

class topaz.model.classifier.LinearClassifier(features)

A simple convolutional layer without non-linear activation.

fill(stride=1)

forward(x)

Applies the classifier to an input.

Parameters: x (np.ndarray) – the image from which features are extracted and classified
Returns: output of the classifer
Return type: z (np.ndarray)

property latent_dim

training: bool

unfill()

property width

Model Factory

topaz.model.factory.conv127(*args, **kwargs)

topaz.model.factory.conv31(*args, **kwargs)

topaz.model.factory.conv63(*args, **kwargs)

topaz.model.factory.get_feature_extractor(model, *args, **kwargs)

topaz.model.factory.load_model(path)

Model Generator

class topaz.model.generative.ConvGenerator(nin, units=32, depth=3, activation=<class 'torch.nn.modules.activation.LeakyReLU'>)

forward(z)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

Model Utilities

topaz.model.utils.insize_from_outsize(layers, outsize): calculates in input size of a convolution stack given the layers and output size