def __init__(self, layers, init_img, subject_img, style_img,
subject_weights, style_weights, smoothness=0.0):
# Map weights (in convolution indices) to layer indices
self.subject_weights = np.zeros(len(layers))
self.style_weights = np.zeros(len(layers))
layers_len = 0
conv_idx = 0
for l, layer in enumerate(layers):
if isinstance(layer, dp.Activation):
self.subject_weights[l] = subject_weights[conv_idx]
self.style_weights[l] = style_weights[conv_idx]
if subject_weights[conv_idx] > 0 or \
style_weights[conv_idx] > 0:
layers_len = l+1
conv_idx += 1
# Discard unused layers
layers = layers[:layers_len]
# Wrap convolution layers for better performance
self.layers = [Convolution(l) if isinstance(l, dp.Convolution) else l
for l in layers]
# Setup network
x_shape = init_img.shape
self.x = Parameter(init_img)
self.x.setup(x_shape)
for layer in self.layers:
layer.setup(x_shape)
x_shape = layer.y_shape(x_shape)
# Precompute subject features and style Gram matrices
self.subject_feats = [None]*len(self.layers)
self.style_grams = [None]*len(self.layers)
next_subject = ca.array(subject_img)
next_style = ca.array(style_img)
for l, layer in enumerate(self.layers):
next_subject = layer.fprop(next_subject)
next_style = layer.fprop(next_style)
if self.subject_weights[l] > 0:
self.subject_feats[l] = next_subject
if self.style_weights[l] > 0:
gram = gram_matrix(next_style)
# Scale gram matrix to compensate for different image sizes
n_pixels_subject = np.prod(next_subject.shape[2:])
n_pixels_style = np.prod(next_style.shape[2:])
scale = (n_pixels_subject / float(n_pixels_style))
self.style_grams[l] = gram * scale
self.tv_weight = smoothness
kernel = np.array([[0, 1, 0], [1, -4, 1], [0, 1, 0]], dtype=dp.float_)
kernel /= np.sum(np.abs(kernel))
self.tv_kernel = ca.array(kernel[np.newaxis, np.newaxis, ...])
self.tv_conv = ca.nnet.ConvBC01((1, 1), (1, 1))
def __init__(self, layers, subject_img, style_img, subject_weights,
style_weights):
# Map weights (in convolution indices) to layer indices
self.subject_weights = np.zeros(len(layers))
self.style_weights = np.zeros(len(layers))
layers_len = 0
conv_idx = 0
for l, layer in enumerate(layers):
if isinstance(layer, dp.Activation):
self.subject_weights[l] = subject_weights[conv_idx]
self.style_weights[l] = style_weights[conv_idx]
if subject_weights[conv_idx] > 0 or \
style_weights[conv_idx] > 0:
layers_len = l + 1
conv_idx += 1
# Discard unused layers
layers = layers[:layers_len]
# Wrap convolution layers for better performance
self.layers = [
Convolution(l) if isinstance(l, dp.Convolution) else l
for l in layers
]
# Setup network
x_shape = subject_img.shape
self.x = Parameter(subject_img)
self.x._setup(x_shape)
for layer in self.layers:
layer._setup(x_shape)
x_shape = layer.y_shape(x_shape)
# Precompute subject features and style Gram matrices
self.subject_feats = [None] * len(self.layers)
self.style_grams = [None] * len(self.layers)
next_subject = ca.array(subject_img)
next_style = ca.array(style_img)
for l, layer in enumerate(self.layers):
next_subject = layer.fprop(next_subject)
next_style = layer.fprop(next_style)
if self.subject_weights[l] > 0:
self.subject_feats[l] = next_subject
if self.style_weights[l] > 0:
gram = gram_matrix(next_style)
# Scale gram matrix to compensate for different image sizes
n_pixels_subject = np.prod(next_subject.shape[2:])
n_pixels_style = np.prod(next_style.shape[2:])
scale = (n_pixels_subject / float(n_pixels_style))
self.style_grams[l] = gram * scale
