Data Learning

class torchelie.data_learning.CorrelateColors

Takes an learnable image and applies the inverse color decorrelation from ImageNet (ie, it correlates the color like ImageNet to ease optimization)

forward(t)

Correlate the color of the image t and return the result

invert(t)

Decorrelate the color of the image t and return the result

class torchelie.data_learning.ParameterizedImg(*shape, init_sd=0.06, init_img=None, space='spectral', colors='uncorr')

A convenient wrapper around PixelImage and SpectralImage and CorrelateColors to make a learnable image.

Parameters:
  • *shape (int) – shape of the image: channel, height, width
  • init_sd (float) – standard deviation for initializing the image if init_img is None
  • init_img (tensor) – an image to use as initialization
  • space (str) – either “pixel” or “spectral” to have the underlying representation be a PixelImage or a SpectralImage
  • colors (str) – either “corr” or “uncorr”, to use a correlated or decorrelated color space
forward()

Return the tensor

render()

Return the tensor on cpu and detached, ready to be transformed to a PIL image

class torchelie.data_learning.PixelImage(shape, sd=0.01, init_img=None)

A learnable image parameterized by its pixel values

Parameters:
  • shape (tuple of int) – a tuple like (channels, height, width)
  • sd (float) – pixels are initialized with a random gaussian value of mean 0.5 and standard deviation sd
  • init_img (tensor, optional) – an image tensor to initialize the pixel values
forward()

Return the image

class torchelie.data_learning.SpectralImage(shape, sd=0.01, decay_power=1, init_img=None)

A learnable image parameterized by its Fourier representation.

See https://distill.pub/2018/differentiable-parameterizations/

Implementation ported from https://github.com/tensorflow/lucid/blob/master/lucid/optvis/param/spatial.py

Parameters:
  • shape (tuple of int) – a tuple like (channels, height, width)
  • sd (float) – amplitudes are initialized with a random gaussian value of mean 0.5 and standard deviation sd
  • init_img (tensor, optional) – an image tensor to initialize the image
forward()

Return the image