def plot_reconstructions(self,
X_test,
channel_to_plot=0,
nb_to_plot=10,
plot_input=True,
same_scale_as_input=True):
"""
Plots the original images, as well as their reconstructions by the autoencoder.
Arguments:
X_test: a numpy array of shape (nb_samples, nb_rows, nb_columns, nb_input_channels)
Only the 10 first samples will be plotted.
channel_to_plot: int, the output_channel that will be plotted.
"""
shape_X = X_test.shape
if len(shape_X) == 3:
X_test = X_test.reshape((1, ) + shape_X)
X_rec = self.reconstruction(X_test[:nb_to_plot])
if same_scale_as_input:
try:
v_min = np.min(X_test[:nb_to_plot, :, :, channel_to_plot])
v_max = np.max(X_test[:nb_to_plot, :, :, channel_to_plot])
except IndexError:
v_min = np.min(X_test[:nb_to_plot, :, :, 0])
v_max = np.max(X_test[:nb_to_plot, :, :, 0])
else:
v_min = None
v_max = None
if plot_input:
bastien_utils.plot_all_images(X_test[:nb_to_plot],
channel_to_plot=channel_to_plot)
bastien_utils.plot_all_images(X_rec,
channel_to_plot=channel_to_plot,
v_min=v_min,
v_max=v_max)
def plot_binarized_weighted_atoms(atoms,
h_test_image,
nb_atoms_to_use_for_threshold_computation=20
):
bin_weighted_atoms = binarized_weighted_atoms(
atoms,
h_test_image,
nb_atoms_to_use_for_threshold_computation=
nb_atoms_to_use_for_threshold_computation)
bastien_utils.plot_all_images(bin_weighted_atoms,
same_intensity_scale=True)
def plot_10_most_used_atoms_for_an_image(h_image,
atoms,
same_intensity_scale=True):
"""
Arguments:
h_image: numpy array of shape (N_features,), the encoding of an image
atoms: numpy array of shape (N_features, N_pixels), the atom images of the learned representation
"""
major_atoms = k_most_used_atoms(10, h_image, atoms)
print(
'Atoms associated with the 10 highest code coefficients of the image')
bastien_utils.plot_all_images(major_atoms,
same_intensity_scale=same_intensity_scale)
def plot_atoms_decoder(self,
channel_to_plot=0,
nb_to_plot=-1,
add_bias=False,
normalize=False):
"""
Plot the weights of the decoder.
Arguments:
nb_to_plot: number of basis images to plot, -1 is all, otherwise plot the nb_to_plot first ones.
channel: channel to plot (there are nb_input_channels*nb_atoms atoms)
add_bias: bool, whether to add the bias (784,) to the weights.
normalize: bool. If True each image is normalized, giving the artificial input images that maximize each of the code coefficients (with unity energy).
"""
atoms = self.atom_images_decoder(add_bias=add_bias,
normalize=normalize)
if (nb_to_plot < 0):
n_atoms = self.latent_dim
else:
n_atoms = nb_to_plot
bastien_utils.plot_all_images(atoms[:n_atoms],
channel_to_plot=channel_to_plot,
same_intensity_scale=True)
def plot_weighted_atoms(atoms, h_test_image):
weighted_atoms = atoms_weighted_by_encoding_coefficients(
atoms, h_test_image)
bastien_utils.plot_all_images(weighted_atoms, same_intensity_scale=True)