lucent.optvis.param package

param.color module

lucent.optvis.param.color.to_valid_rgb(image_f: Callable, decorrelate: Optional[bool] = False) → Callable

Converts image function to one with valid rgb values in [0,1] via sigmoid and potentially decorrelates the color channels

Parameters

  • image_f (Callable) – Original image function

  • decorrelate (Optional[bool], optional) – Whether to decorrelate colors, defaults to False

Returns

New image function

Return type

Callable

param.cppn module

class lucent.optvis.param.cppn.CompositeActivation

Bases: Module

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
lucent.optvis.param.cppn.cppn(size: int, num_output_channels: ~typing.Optional[int] = 3, num_hidden_channels: ~typing.Optional[int] = 24, num_layers: ~typing.Optional[int] = 8, activation_fn: ~typing.Optional[~torch.nn.modules.module.Module] = <class 'lucent.optvis.param.cppn.CompositeActivation'>, normalize: ~typing.Optional[bool] = False) → Tuple[Iterator, Callable]

Creates cppn parameterization.

Parameters

  • size (int) – image size

  • num_output_channels (Optional[int], optional) – number of output channels, defaults to 3

  • num_hidden_channels (Optional[int], optional) – number of hidden channels, defaults to 24

  • num_layers (Optional[int], optional) – number of layers, defaults to 8

  • activation_fn (Optional[torch.nn.Module], optional) – activation function after hidden layers, defaults to CompositeActivation

  • normalize (Optional[bool], optional) – Whether to use instance normalization after each hidden layer, defaults to False

Returns

Iterator over network parameters and a function that returns the network’s output on the grid [-sqrt(3), sqrt(3)] with step size size.

Return type

Tuple[Iterator, Callable]

param.gan module

Copyright 2020 Binxu Wang Use GAN as prior to do feature visualization. This method is inspired by the work

Nguyen, A., Dosovitskiy, A., Yosinski, J., Brox, T., & Clune, J. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks.(2016) NIPS

The GAN model is imported from

A. Dosovitskiy, T. Brox Generating Images with Perceptual Similarity Metrics based on Deep Networks (2016), NIPS. https://lmb.informatik.uni-freiburg.de/people/dosovits/code.html

the author translated the models (pool5-fc8) into pytorch and hosts the weights online.

Jun.4th 2020

class lucent.optvis.param.gan.UpConvGAN(name='fc6', pretrained=True)

Bases: Module

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
visualize(x, scale=1.0)
lucent.optvis.param.gan.load_statedict_from_online(name='fc6')

param.images module

High-level wrapper for paramaterizing images.

lucent.optvis.param.images.image(w: int, h: Optional[int] = None, sd: Optional[float] = 0.01, batch: Optional[int] = 1, decorrelate: Optional[bool] = True, fft: Optional[bool] = True, channels: Optional[int] = None) → Tuple[List[Tensor], Callable]

Creates image parameterization either via rgb or fft.

Parameters

  • w (int) – image width

  • h (Optional[int], optional) – image height, if None will be the same as w, defaults to None

  • sd (Optional[float], optional) – standard deviation of random init, defaults to 0.01

  • batch (Optional[int], optional, defaults to 1.) – [description], defaults to None

  • decorrelate (Optional[bool], optional) – [description], defaults to True

  • fft (Optional[bool], optional) – [description], defaults to True

  • channels (Optional[int], optional) – number of channels - if None will be set to 3, defaults to None

Returns

[description]

Return type

Tuple[List[torch.Tensor], Callable]

param.lowres module

Provides lowres_tensor().

lucent.optvis.param.lowres.lowres_tensor(shape: Union[List, Tuple, Size], underlying_shape: Union[List, Tuple, Size], offset: Optional[Union[bool, int, List]] = None, sd: Optional[float] = 0.01) → Tuple[List[Tensor], Callable]

Produces a tensor paramaterized by a interpolated lower resolution tensor. This is like what is done in a laplacian pyramid, but a bit more general. It can be a powerful way to describe images.

Parameters

  • shape (Union[List, Tuple, torch.Size]) – desired shape of resulting tensor, should be of format (B, C, H, W) #TODO support more shapes

  • underlying_shape (Union[List, Tuple, torch.Size]) – shape of the tensor being resized into final tensor

  • offset (Optional[Union[bool, int, List]], optional) – Describes how to offset the interpolated vector (like phase in a Fourier transform). If None, apply no offset. If int, apply the same offset to each dimension; if a list use each entry for each dimension. If False, do not offset. If True, offset by half the ratio between shape and underlying shape (analogous to 90 degrees), defaults to None

  • sd (Optional[float], optional) – Standard deviation of initial tensor variable., defaults to 0.01

Returns

One-element list containing the low resolution tensor and the corresponding image function returning the tensor on call.

Return type

Tuple[List[torch.Tensor], Callable]

param.spatial module

lucent.optvis.param.spatial.fft_image(shape: Union[List, Tuple, Size], sd: Optional[float] = 0.01, decay_power: Optional[int] = 1) → Tuple[List[Tensor], Callable]

Image parameterization via fft spectrum

Parameters

  • shape (Union[List, Tuple, torch.Size]) – Image shape

  • sd (Optional[float], optional) – Standard deviation for initializing distribution, defaults to 0.01

  • decay_power (Optional[int], optional) – Decay power to dampen scaling, defaults to 1

Returns

One-element list containing image parameters and image function returning the image on call.

Return type

Tuple[List[Tensor], Callable]

lucent.optvis.param.spatial.pixel_image(shape: Union[List, Tuple, Size], sd: Optional[float] = 0.01) → Tuple[List[Tensor], Callable]

Pixel parameterization of an image with given shape.

Init via zero-mean normal distribution with given standard deviation.

Parameters

  • shape (Union[List, Tuple, torch.Size]) – Shape of image

  • sd (Optional[float], optional) – Standard deviation of initializing distribution, defaults to 0.01

Returns

one-element list containing image parameters and image function returning the image on call.

Return type

Tuple[List[torch.Tensor], Callable]

lucent.optvis.param.spatial.rfft2d_freqs(h: int, w: int) → ndarray

Computes 2D spectrum frequencies of an image with given dimensions.

From https://github.com/tensorflow/lucid/blob/master/lucid/optvis/param/spatial.py .

Parameters

  • h (int) – image height

  • w (int) – image width

Returns

2D spectrum frequencies

Return type

np.ndarray

Module contents