X_train_style = np.array([
load_image(args.style,
size=(height, width),
preprocess_type='vgg19',
verbose=True)
])
print("X_train_style shape:", X_train_style.shape)
print('Loading VGG headless 5')
modelWeights = "%s/%s-%s-%s%s" % (vgg19Dir, 'vgg-19', dim_ordering, K._BACKEND,
'_headless_5_weights.hdf5')
model = VGG_19_headless_5(input_shape,
modelWeights,
trainable=False,
pooling_type=args.pooling_type)
layer_dict, layers_names = get_layer_data(model, 'conv_')
print('Layers found:' + ', '.join(layers_names))
input_layer = model.input
print('Building white noise images')
input_data = create_noise_tensor(height, width, channels)
mean_losses = {}
print('Using optimizer: ' + optimizer)
current_iter = 1
for layer_name in layers_names:
prefix = str(current_iter).zfill(4)
print('Creating labels for ' + layer_name)
out = layer_dict[layer_name].output
X_train_style = np.array([
load_image(args.style,
size=(height, width),
preprocess_type='vgg19',
verbose=True)
])
print("X_train_style shape:", X_train_style.shape)
print('Loading VGG headless 5')
modelWeights = "%s/%s-%s-%s%s" % (vgg19Dir, 'vgg-19', dim_ordering, K._BACKEND,
'_headless_5_weights.hdf5')
model = VGG_19_headless_5(input_shape,
modelWeights,
trainable=False,
pooling_type=args.pooling_type)
layer_dict, layers_names = get_layer_data(
model, 'conv_(1|2|3|4)') # remove conv_5_* layers
print('Layers found:' + ', '.join(layers_names))
input_layer = model.input
print('Building white noise images')
input_data = create_noise_tensor(height, width, channels)
print('Using optimizer: ' + optimizer)
current_iter = 1
alpha = 5e1 # 1e2
for idx_content, lc_name in enumerate(layers_names):
for idx_style, ls_name in enumerate(layers_names):
print('Creating labels for content ' + lc_name + ' and style ' +
ls_name)
out_style = layer_dict[ls_name].output
# We separate the image_freq and the histogram_freq just to avoid dumping everything at the same time
callbacks.append(tb)
callbacks.append(
ModelCheckpointBatch(st_model,
chkp_dir=prefixed_dir + '/chkp',
nb_step_chkp=200))
callbacks.append(HistoryBatch())
print('Loading VGG headless 4')
modelWeights = "%s/%s-%s-%s%s" % (vgg19Dir, 'vgg-19', dim_ordering, K._BACKEND,
'_headless_4_weights.hdf5')
vgg_model = VGG_19_headless_4(input_shape,
modelWeights,
trainable=False,
pooling_type=args.pooling_type)
layer_dict, layers_names = get_layer_data(vgg_model, 'conv_')
style_layers = ['conv_1_2', 'conv_2_2', 'conv_3_4', 'conv_4_2']
style_layers_mask = [name in style_layers for name in layers_names]
# About that: each time you have a pooling you hardly limiting the gradient flow upward
# after the > 2 layers, if you have a to do mini batchs and have a lot of noise
# You won't be able to pick up the gradient to converge
content_layers = ['conv_2_2']
content_layers_mask = [name in content_layers for name in layers_names]
print('Building full model')
mean = load_mean(name='vgg19') # BGR
if K._BACKEND == "tensorflow":
import tensorflow as tf
preprocessed_output = Lambda(
lambda x: tf.reverse(x, [False, False, False, True]) - mean,
name="ltv")(st_model.output) # RGB -> BGR -> BGR - mean