我在网上找了好久。但我什么也没发现。请帮助我或尝试给我一些关于如何实现这一目标的想法。
我已经构建了 3 个输入和 2 个输出模型，如下所示。
the code block is from def build_srgan_model in class SRGANNetwork:
ip = Input(shape=(self.img_width, self.img_height,3), name='x_generator')
ip_gan = Input(shape=(large_width, large_height,3), name='x_discriminator')
ip_vgg = Input(shape=(large_width, large_height,3), name='x_vgg')
sr_output = self.generative_network.create_sr_model(ip)
self.generative_model_ = Model(ip, sr_output)
gan_output = self.discriminative_network.append_gan_network(ip_gan)
self.discriminative_model_ = Model(ip_gan, gan_output)
gan_output = self.discriminative_model_(self.generative_model_.output)
vgg_output = self.vgg_network.append_vgg_network(self.generative_model_.output, ip_vgg)
self.srgan_model_ = Model(inputs=[ip, ip_gan, ip_vgg], outputs=[gan_output, vgg_output])
这是完整的代码，太长了。
VGG网络类：
'''
Helper class to load VGG and its weights to the FastNet model
'''
def __init__(self, img_width=384, img_height=384, vgg_weight=1.0):
self.img_height = img_height
self.img_width = img_width
self.vgg_weight = vgg_weight
self.vgg_layers = None
def append_vgg_network(self, x_in, true_X_input, pre_train=False):
# Append the initial inputs to the outputs of the SRResNet
x = merge([x_in, true_X_input], mode='concat', concat_axis=0)
# Normalize the inputs via custom VGG Normalization layer
x = Normalize(name="normalize_vgg")(x)
# Begin adding the VGG layers
x = Convolution2D(64, 3, 3, activation='relu', name='vgg_conv1_1', border_mode='same')(x)
x = Convolution2D(64, 3, 3, activation='relu', name='vgg_conv1_2', border_mode='same')(x)
x = MaxPooling2D(name='vgg_maxpool1')(x)
x = Convolution2D(128, 3, 3, activation='relu', name='vgg_conv2_1', border_mode='same')(x)
if pre_train:
vgg_regularizer2 = ContentVGGRegularizer(weight=self.vgg_weight)
x = Convolution2D(128, 3, 3, activation='relu', name='vgg_conv2_2', border_mode='same',
activity_regularizer=vgg_regularizer2)(x)
else:
x = Convolution2D(128, 3, 3, activation='relu', name='vgg_conv2_2', border_mode='same')(x)
x = MaxPooling2D(name='vgg_maxpool2')(x)
x = Convolution2D(256, 3, 3, activation='relu', name='vgg_conv3_1', border_mode='same')(x)
x = Convolution2D(256, 3, 3, activation='relu', name='vgg_conv3_2', border_mode='same')(x)
x = Convolution2D(256, 3, 3, activation='relu', name='vgg_conv3_3', border_mode='same')(x)
x = MaxPooling2D(name='vgg_maxpool3')(x)
x = Convolution2D(512, 3, 3, activation='relu', name='vgg_conv4_1', border_mode='same')(x)
x = Convolution2D(512, 3, 3, activation='relu', name='vgg_conv4_2', border_mode='same')(x)
x = Convolution2D(512, 3, 3, activation='relu', name='vgg_conv4_3', border_mode='same')(x)
x = MaxPooling2D(name='vgg_maxpool4')(x)
x = Convolution2D(512, 3, 3, activation='relu', name='vgg_conv5_1', border_mode='same')(x)
x = Convolution2D(512, 3, 3, activation='relu', name='vgg_conv5_2', border_mode='same')(x)
if not pre_train:
vgg_regularizer5 = ContentVGGRegularizer(weight=self.vgg_weight)
x = Convolution2D(512, 3, 3, activation='relu', name='vgg_conv5_3', border_mode='same',
activity_regularizer=vgg_regularizer5)(x)
else:
x = Convolution2D(512, 3, 3, activation='relu', name='vgg_conv5_3', border_mode='same')(x)
x = MaxPooling2D(name='vgg_maxpool5')(x)
return x
def load_vgg_weight(self, model):
# Loading VGG 16 weights
if K.image_dim_ordering() == "th":
weights = get_file('vgg16_weights_th_dim_ordering_th_kernels_notop.h5', THEANO_WEIGHTS_PATH_NO_TOP,
cache_subdir='models')
else:
weights = get_file('vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5', TF_WEIGHTS_PATH_NO_TOP,
cache_subdir='models')
f = h5py.File(weights)
layer_names = [name for name in f.attrs['layer_names']]
if self.vgg_layers is None:
self.vgg_layers = [layer for layer in model.layers
if 'vgg_' in layer.name]
for i, layer in enumerate(self.vgg_layers):
g = f[layer_names[i]] #g is i layer weight of .h5 //dict{'weight_names':wight]}
weights = [g[name] for name in g.attrs['weight_names']]
layer.set_weights(weights)
# Freeze all VGG layers
for layer in self.vgg_layers:
layer.trainable = False
return model
类鉴别器网络：
def __init__(self, img_width=384, img_height=384, adversarial_loss_weight=1, small_model=False):
self.img_width = img_width
self.img_height = img_height
self.adversarial_loss_weight = adversarial_loss_weight
self.small_model = small_model
self.k = 3
self.mode = 2
self.weights_path = "weights/Discriminator weights.h5"
self.gan_layers = None
def append_gan_network(self, true_X_input):
# Normalize the inputs via custom VGG Normalization layer
x = Normalize(type="gan", value=127.5, name="gan_normalize")(true_X_input)
x = Convolution2D(64, self.k, self.k, border_mode='same', name='gan_conv1_1')(x)
x = LeakyReLU(0.3, name="gan_lrelu1_1")(x)
x = Convolution2D(64, self.k, self.k, border_mode='same', name='gan_conv1_2', subsample=(2, 2))(x)
x = LeakyReLU(0.3, name='gan_lrelu1_2')(x)
x = BatchNormalization(mode=self.mode, axis=channel_axis, name='gan_batchnorm1_1')(x)
filters = [128, 256] if self.small_model else [128, 256, 512]
for i, nb_filters in enumerate(filters):
for j in range(2):
subsample = (2, 2) if j == 1 else (1, 1)
x = Convolution2D(nb_filters, self.k, self.k, border_mode='same', subsample=subsample,
name='gan_conv%d_%d' % (i + 2, j + 1))(x)
x = LeakyReLU(0.3, name='gan_lrelu_%d_%d' % (i + 2, j + 1))(x)
x = BatchNormalization(mode=self.mode, axis=channel_axis, name='gan_batchnorm%d_%d' % (i + 2, j + 1))(x)
x = Flatten(name='gan_flatten')(x)
output_dim = 128 if self.small_model else 1024
x = Dense(output_dim, name='gan_dense1')(x)
x = LeakyReLU(0.3, name='gan_lrelu5')(x)
gan_regulrizer = AdversarialLossRegularizer(weight=self.adversarial_loss_weight)
x = Dense(2, activation="softmax", activity_regularizer=gan_regulrizer, name='gan_output')(x)
return x
def set_trainable(self, model, value=True):
if self.gan_layers is None:
disc_model = [layer for layer in model.layers
if 'model' in layer.name][0] # Only disc model is an inner model
self.gan_layers = [layer for layer in disc_model.layers
if 'gan_' in layer.name]
for layer in self.gan_layers:
layer.trainable = value
def load_gan_weights(self, model):
f = h5py.File(self.weights_path)
layer_names = [name for name in f.attrs['layer_names']]
layer_names = layer_names[1:] # First is an input layer. Not needed.
if self.gan_layers is None:
self.gan_layers = [layer for layer in model.layers
if 'gan_' in layer.name]
for i, layer in enumerate(self.gan_layers):
g = f[layer_names[i]]
weights = [g[name] for name in g.attrs['weight_names']]
layer.set_weights(weights)
print("GAN Model weights loaded.")
return model
def save_gan_weights(self, model):
print('GAN Weights are being saved.')
model.save_weights(self.weights_path, overwrite=True)
print('GAN Weights saved.')
类生成网络：
def __init__(self, img_width=96, img_height=96, batch_size=16, nb_upscales=2, small_model=False,
content_weight=1, tv_weight=2e5, gen_channels=64):
self.img_width = img_width
self.img_height = img_height
self.batch_size = batch_size
self.small_model = small_model
self.nb_scales = nb_upscales
self.content_weight = content_weight
self.tv_weight = tv_weight
self.filters = gen_channels
self.mode = 2
self.init = 'glorot_uniform'
self.sr_res_layers = None
self.sr_weights_path = "weights/SRGAN.h5"
self.output_func = None
def create_sr_model(self, ip):
x = Convolution2D(self.filters, 5, 5, activation='linear', border_mode='same', name='sr_res_conv1',
init=self.init)(ip)
x = BatchNormalization(axis=channel_axis, mode=self.mode, name='sr_res_bn_1')(x)
x = LeakyReLU(alpha=0.25, name='sr_res_lr1')(x)
# x = Convolution2D(self.filters, 5, 5, activation='linear', border_mode='same', name='sr_res_conv2')(x)
# x = BatchNormalization(axis=channel_axis, mode=self.mode, name='sr_res_bn_2')(x)
# x = LeakyReLU(alpha=0.25, name='sr_res_lr2')(x)
nb_residual = 5 if self.small_model else 15
for i in range(nb_residual):
x = self._residual_block(x, i + 1)
for scale in range(self.nb_scales):
x = self._upscale_block(x, scale + 1)
scale = 2 ** self.nb_scales
tv_regularizer = TVRegularizer(img_width=self.img_width * scale, img_height=self.img_height * scale,
weight=self.tv_weight) #self.tv_weight)
x = Convolution2D(3, 5, 5, activation='tanh', border_mode='same', activity_regularizer=tv_regularizer,
init=self.init, name='sr_res_conv_final')(x)
x = Denormalize(name='sr_res_conv_denorm')(x)
return x
def _residual_block(self, ip, id):
init = ip
x = Convolution2D(self.filters, 3, 3, activation='linear', border_mode='same', name='sr_res_conv_' + str(id) + '_1',
init=self.init)(ip)
x = BatchNormalization(axis=channel_axis, mode=self.mode, name='sr_res_bn_' + str(id) + '_1')(x)
x = LeakyReLU(alpha=0.25, name="sr_res_activation_" + str(id) + "_1")(x)
x = Convolution2D(self.filters, 3, 3, activation='linear', border_mode='same', name='sr_res_conv_' + str(id) + '_2',
init=self.init)(x)
x = BatchNormalization(axis=channel_axis, mode=self.mode, name='sr_res_bn_' + str(id) + '_2')(x)
m = merge([x, init], mode='sum', name="sr_res_merge_" + str(id))
return m
def _upscale_block(self, ip, id):
'''
As per suggestion from http://distill.pub/2016/deconv-checkerboard/, I am swapping out
SubPixelConvolution to simple Nearest Neighbour Upsampling
'''
init = ip
x = Convolution2D(128, 3, 3, activation="linear", border_mode='same', name='sr_res_upconv1_%d' % id,
init=self.init)(init)
x = LeakyReLU(alpha=0.25, name='sr_res_up_lr_%d_1_1' % id)(x)
x = UpSampling2D(name='sr_res_upscale_%d' % id)(x)
#x = SubPixelUpscaling(r=2, channels=32)(x)
x = Convolution2D(128, 3, 3, activation="linear", border_mode='same', name='sr_res_filter1_%d' % id,
init=self.init)(x)
x = LeakyReLU(alpha=0.3, name='sr_res_up_lr_%d_1_2' % id)(x)
return x
def set_trainable(self, model, value=True):
if self.sr_res_layers is None:
self.sr_res_layers = [layer for layer in model.layers
if 'sr_res_' in layer.name]
for layer in self.sr_res_layers:
layer.trainable = value
def get_generator_output(self, input_img, srgan_model):
if self.output_func is None:
gen_output_layer = [layer for layer in srgan_model.layers
if layer.name == "sr_res_conv_denorm"][0]
self.output_func = K.function([srgan_model.layers[0].input],
[gen_output_layer.output])
return self.output_func([input_img])
SRGAN网络类：
def __init__(self, img_width=96, img_height=96, batch_size=16, nb_scales=2):
self.img_width = img_width
self.img_height = img_height
self.batch_size = batch_size
self.nb_scales = nb_scales #factor 4X using nb_scales Upsampling2D
self.discriminative_network = None # type: DiscriminatorNetwork
self.generative_network = None # type: GenerativeNetwork
self.vgg_network = None # type: VGGNetwork
self.srgan_model_ = None # type: Model
self.generative_model_ = None # type: Model
self.discriminative_model_ = None #type: Model
def build_srgan_pretrain_model(self, use_small_srgan=False):
large_width = self.img_width * 4
large_height = self.img_height * 4
self.generative_network = GenerativeNetwork(self.img_width, self.img_height, self.batch_size, self.nb_scales,
use_small_srgan)
self.vgg_network = VGGNetwork(large_width, large_height)
ip = Input(shape=(3, self.img_width, self.img_height), name='x_generator')
ip_vgg = Input(shape=(3, large_width, large_height), name='x_vgg') # Actual X images
sr_output = self.generative_network.create_sr_model(ip)
self.generative_model_ = Model(ip, sr_output)
vgg_output = self.vgg_network.append_vgg_network(sr_output, ip_vgg, pre_train=True)
self.srgan_model_ = Model(input=[ip, ip_vgg],
output=vgg_output)
self.vgg_network.load_vgg_weight(self.srgan_model_)
srgan_optimizer = Adam(lr=1e-4)
generator_optimizer = Adam(lr=1e-4)
self.generative_model_.compile(generator_optimizer, dummy_loss)
self.srgan_model_.compile(srgan_optimizer, dummy_loss)
return self.srgan_model_
def build_discriminator_pretrain_model(self, use_smal_srgan=False, use_small_discriminator=False):
large_width = self.img_width * 4
large_height = self.img_height * 4
self.generative_network = GenerativeNetwork(self.img_width, self.img_height, self.batch_size, self.nb_scales,
use_small_srgan)
self.discriminative_network = DiscriminatorNetwork(large_width, large_height,
small_model=use_small_discriminator)
ip = Input(shape=(3, self.img_width, self.img_height), name='x_generator')
ip_gan = Input(shape=(3, large_width, large_height), name='x_discriminator') # Actual X images
sr_output = self.generative_network.create_sr_model(ip)
self.generative_model_ = Model(ip, sr_output)
#self.generative_network.set_trainable(self.generative_model_, value=False)
gan_output = self.discriminative_network.append_gan_network(ip_gan)
self.discriminative_model_ = Model(ip_gan, gan_output)
generator_out = self.generative_model_(ip)
gan_output = self.discriminative_model_(generator_out)
self.srgan_model_ = Model(input=ip, output=gan_output)
srgan_optimizer = Adam(lr=1e-4)
generator_optimizer = Adam(lr=1e-4)
discriminator_optimizer = Adam(lr=1e-4)
self.generative_model_.compile(generator_optimizer, loss='mse')
self.discriminative_model_.compile(discriminator_optimizer, loss='categorical_crossentropy', metrics=['acc'])
self.srgan_model_.compile(srgan_optimizer, loss='categorical_crossentropy', metrics=['acc'])
return self.discriminative_model_
def build_srgan_model(self, use_small_srgan=False, use_small_discriminator=False):
large_width = self.img_width * 4
large_height = self.img_height * 4
self.generative_network = GenerativeNetwork(self.img_width, self.img_height, self.batch_size, nb_upscales=self.nb_scales,
small_model=use_small_srgan)
self.discriminative_network = DiscriminatorNetwork(large_width, large_height,
small_model=use_small_discriminator)
self.vgg_network = VGGNetwork(large_width, large_height)
ip = Input(shape=(3, self.img_width, self.img_height), name='x_generator')
ip_gan = Input(shape=(3, large_width, large_height), name='x_discriminator') # Actual X images
ip_vgg = Input(shape=(3, large_width, large_height), name='x_vgg') # Actual X images
sr_output = self.generative_network.create_sr_model(ip)
self.generative_model_ = Model(ip, sr_output)
gan_output = self.discriminative_network.append_gan_network(ip_gan)
self.discriminative_model_ = Model(ip_gan, gan_output)
gan_output = self.discriminative_model_(self.generative_model_.output)
vgg_output = self.vgg_network.append_vgg_network(self.generative_model_.output, ip_vgg)
self.srgan_model_ = Model(input=[ip, ip_gan, ip_vgg], output=[gan_output, vgg_output])
self.vgg_network.load_vgg_weight(self.srgan_model_)
srgan_optimizer = Adam(lr=1e-4)
generator_optimizer = Adam(lr=1e-4)
discriminator_optimizer = Adam(lr=1e-4)
self.generative_model_.compile(generator_optimizer, dummy_loss)
self.discriminative_model_.compile(discriminator_optimizer, loss='categorical_crossentropy', metrics=['acc'])
self.srgan_model_.compile(srgan_optimizer, dummy_loss)
return self.srgan_model_
但是我在使用keras2.0运行代码时遇到了这个错误。
Traceback (most recent call last):
File "<ipython-input-43-f26b38c03792>", line 1, in <module>
runfile('F:/keras_projects/Super-Resolution-using-Generative-Adversarial-Networks-master/models.py', wdir='F:/keras_projects/Super-Resolution-using-Generative-Adversarial-Networks-master')
File "C:\Program Files\Anaconda3\lib\site-packages\spyder\utils\site\sitecustomize.py", line 866, in runfile
execfile(filename, namespace)
File "C:\Program Files\Anaconda3\lib\site-packages\spyder\utils\site\sitecustomize.py", line 102, in execfile
exec(compile(f.read(), filename, 'exec'), namespace)
File "F:/keras_projects/Super-Resolution-using-Generative-Adversarial-Networks-master/models.py", line 778, in <module>
srgan_network.build_srgan_model()
File "F:/keras_projects/Super-Resolution-using-Generative-Adversarial-Networks-master/models.py", line 421, in build_srgan_model
self.srgan_model_ = Model(inputs=[ip, ip_gan, ip_vgg], outputs=[gan_output, vgg_output])
File "C:\Program Files\Anaconda3\lib\site-packages\keras\legacy\interfaces.py", line 88, in wrapper
return func(*args, **kwargs)
File "C:\Program Files\Anaconda3\lib\site-packages\keras\engine\topology.py", line 1676, in __init__
build_map_of_graph(x, finished_nodes, nodes_in_progress)
File "C:\Program Files\Anaconda3\lib\site-packages\keras\engine\topology.py", line 1666, in build_map_of_graph
layer, node_index, tensor_index)
File "C:\Program Files\Anaconda3\lib\site-packages\keras\engine\topology.py", line 1664, in build_map_of_graph
next_node = layer.inbound_nodes[node_index]
AttributeError: 'NoneType' object has no attribute 'inbound_nodes'
以我有限的经验，我认为这是keras版本造成的。很抱歉没有包含太多的表达。
