Inpainters Module

Interface

class picasso.inpainters.interfaces.HoleInpainter(args, Npix=128, meshgrid=True)

Bases: object

This class provides an interface to the 3 inpainting techniques. One of the key parameters is args importing arguments input by the user in the inpainting scripts.

DPinpaint(reuse)

Set of instructions to inpaint with DeepPrior

GANinpaint(reuse)

Set of instructions to inpaint with ContextualAttention

NNinpaint()

Set of instructions to inpaint with NearestNeighbours

setup_input(fname, rdseed=None)

Pre-process the flat map by renormalizing and reshaping it as it required by the inpainting method

Deep Prior

class picasso.inpainters.deep_prior_inpainter.DeepPrior(input_shape, nd=[16, 32, 64, 128, 128, 128], verbose=False, meshgrid=True)

Bases: object

Inpainting class interface for Deep-Prior

compile(optimizer)

Compile the architecture

downsampling_block(X, f, filters, stage, block, s=2)

Implementation of a down-sampling block

Parameters

• X:

  input tensor of shape (m, n_H_prev, n_W_prev, n_C_prev)

• f: {int}

  specifying the shape of the middle CONV’s window for the main path

• filters :{int}

  defining the number of filters in the CONV layers of the main path

• stage :{int}

  used to name the layers, depending on their position in the network

• block: {string}

  used to name the layers, depending on their position in the network

Returns

• X:

  output of the convolutional block, tensor of shape (n_H, n_W, n_C)

evaluate(z, X)

my_accuracy(y_true, y_pred)

Accuracy defined as the ratio between the 2-norms of the prediction and of the ground-truth.

myloss(y_true, y_pred)

Loss function evaluated outside the masked region and normalized by the ground-truth norm.

predict()

Inpainting with Deep -Prior

rescale_back(v)

rescale inpainted map to physical units .

setup_input(fname_masked)

Preprocessing the corrupted image and setting up the generative images. By default deep-prior will generate pixel values starting from an image with uniformly random distributed pixels (in \([0,1/10]\)). Otherwise if meshgrid==True, generator is a set of 4 images, i.e. a upward, downward, leftward and rightward diagonal gradients_summary. This mitigates the gridding pattern injected in the reconstruction area.

summary()

Returns the model summary

train(z, X, epochs=2000, verbose=False)

Iterate over the epochs

upsampling_block(X, f, filters, stage, block)

Implementation of the up-sampling block

Parameters

• X:

  input tensor of shape (m, n_H_prev, n_W_prev, n_C_prev)

• f: {int}

  specifying the shape of the middle CONV’s window for the main path

• filters :{int}

  defining the number of filters in the CONV layers of the main path

• stage :{int}

  used to name the layers, depending on their position in the network

• block: {string}

  used to name the layers, depending on their position in the network

Returns

• X:

  output of the convolutional block, tensor of shape (n_H, n_W, n_C)

Nearest Neighbours

class picasso.inpainters.nearest_neighbours_inpainter.NearestNeighbours(Npix=128, verbose=False, tol=1e-08)

Bases: object

Nearest-neighbours inpainting by diffusing the average of the nearest pixel values.

Inpainting can be performed on a generic location in the map. It is performed in a loop and it is stopped when the inpainted array in two subsequent iterations is essentially the same for a given tolerance threshold tol. Default tolerance is the one set from numpy.allclose, 1e-8.

predict()

Inpainting the map with Nearest-Neighbours .

setup_input(fname_masked)

Setup the the masked image to be inpainted and the mask

Parameters

• fname_masked:{string}

  loading .npy file of the masked image

GAN

class picasso.inpainters.contextual_attention_gan.ContextualAttention(modeldir=None, verbose=False)

Bases: picasso.inpainters.generative_inpainting_model.InpaintCAModel

Inpainting with GAN.

It requires a pre-trained network with weights stored in modeldir. This is a further interface layer to the Generative inpainting developed by Jiahui Yu (https://github.com/JiahuiYu/generative_inpainting)

postprocess_output(output, sess)

predict(reuse)

Inpainting with GAN

preprocess_input()

rescale_back(v)

MinMax rescale to the map to the physical units

setup_input(fname_masked)

Pre-process the map to be inpainted with GAN.