class Discriminator:
with tf.variable_scope('discriminator'):
inputs = Generator.outputs # will be fed directly if it's a real image
micro = [inputs] # for errors in small scale
wide_micro = tf.layers.conv2d(micro[-1],
filters=8,
kernel_size=(7, 3),
padding='same',
activation=tf.nn.relu)
high_micro = tf.layers.conv2d(micro[-1],
filters=8,
kernel_size=(3, 7),
padding='same',
activation=tf.nn.relu)
micro.append(tf.concat((wide_micro, high_micro), axis=-1))
micro.append(utils.pool(micro[-1], 'MAX', pool_size=4))
micro.append(
tf.layers.conv2d(micro[-1],
filters=4,
kernel_size=5,
padding='same',
activation=tf.nn.relu))
macro = [utils.pool(inputs, 'AVG',
pool_size=4)] # for errors in large scale
wide_macro = tf.layers.conv2d(macro[-1],
filters=8,
kernel_size=(7, 3),
padding='same',
activation=tf.nn.relu)
high_macro = tf.layers.conv2d(macro[-1],
filters=8,
kernel_size=(3, 7),
padding='same',
activation=tf.nn.relu)
macro.append(tf.concat((wide_macro, high_macro), axis=-1))
macro.append(utils.pool(macro[-1], 'MAX', pool_size=4))
macro.append(
tf.layers.conv2d(macro[-1],
filters=4,
kernel_size=5,
padding='same',
activation=tf.nn.relu))
logits = tf.reduce_mean(micro[-1], axis=(1, 2, 3)) + tf.reduce_mean(
macro[-1], axis=(1, 2, 3))
outputs = tf.nn.sigmoid(logits) # 0 for fake, 1 for real
vars = tf.trainable_variables(scope='discriminator')
targets = tf.placeholder(dtype=tf.float32, shape=(None))
loss = tf.losses.sigmoid_cross_entropy(targets, logits)
optimizer = tf.train.AdamOptimizer(5e-4).minimize(loss, var_list=vars)
def bilrp(self, x1, x2, poolsize, gamma_func=None):
# Computing LRP passes for each branch
r1 = self.compute_branch(x1, gamma_func)
r2 = self.compute_branch(x2, gamma_func)
# Computing BiLRP relevances
R = [np.array(r).sum(1) for r in [r1, r2]]
R = np.tensordot(pool(R[0], poolsize),
pool(R[1], poolsize),
axes=(0, 0))
return R
def vgg(inputs, struct, dbb_biases, dbb=False):
def _block(inputs, filters, dbb, dbb_bias):
inputs, mem1 = conv(inputs,
filters,
3,
strides=1,
padding='SAME',
dbb=dbb,
dbb_bias=dbb_bias)
inputs, mem2 = batch_norm(inputs)
#print(inputs)
return inputs, mem1 + mem2
total_mem = 0
inputs, mem1 = _block(inputs, struct[0], dbb, dbb_biases[0])
inputs, mem2 = _block(inputs, struct[1], dbb, dbb_biases[1])
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem3 = _block(inputs, struct[2], dbb, dbb_biases[2])
inputs, mem4 = _block(inputs, struct[3], dbb, dbb_biases[3])
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem5 = _block(inputs, struct[4], dbb, dbb_biases[4])
inputs, mem6 = _block(inputs, struct[5], dbb, dbb_biases[5])
inputs, mem7 = _block(inputs, struct[6], dbb, dbb_biases[6])
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem8 = _block(inputs, struct[7], dbb, dbb_biases[7])
inputs, mem9 = _block(inputs, struct[8], dbb, dbb_biases[8])
inputs, mem10 = _block(inputs, struct[9], dbb, dbb_biases[9])
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem11 = _block(inputs, struct[10], dbb, dbb_biases[10])
inputs, mem12 = _block(inputs, struct[11], dbb, dbb_biases[11])
inputs, mem13 = _block(inputs, struct[12], dbb, dbb_biases[12])
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem_flat = flatten(inputs, struct[13], dbb, dbb_biases[13])
inputs, mem14 = dense(inputs, struct[14], dbb=dbb, dbb_bias=dbb_biases[14])
inputs, mem15 = batch_norm(inputs)
inputs, mem16 = dense(inputs, struct[15], dbb=False)
total_mem = mem1 + mem2 + mem3 + mem4 + \
mem5 + mem6 + mem7 + mem8 + \
mem9 + mem10 + mem11 + mem12 + \
mem13 + mem14 + mem15 + mem16 + mem_flat
return inputs, total_mem
def cnn_model0(IMG_SIZE):
data_input = Input(shape=(IMG_SIZE, IMG_SIZE,3))
layer1 = conv(data_input, 32, (5, 5), padding='same', strides=(1, 1), name='layer1')
layer2 = pool(layer1, pool_size=(3, 3), strides=(2, 2), padding='same', name='layer2', pool_type='max')
layer3 = conv(layer2, 64, (3, 3), padding='same', strides=(1, 1), name='layer3')
layer4 = conv(layer3, 96, (3, 3), padding='same', strides=(1, 1), name='layer4')
layer5 = pool(layer4,pool_size=(3, 3), strides=(2, 2), padding='same',name='layer5',pool_type='max')
layer6 = conv(layer5, 128, (3, 3), padding='same', strides=(1, 1), name='layer6')
layer7 = conv(layer6, 256, (3, 3), padding='same', strides=(1, 1), name='layer7')
layer8 = pool(layer7, pool_size=(3, 3), strides=(2, 2), padding='same', name='layer8', pool_type='max')
layer9 = conv(layer8, 256, (5, 5), padding='same', strides=(1, 1), name='layer9')
x = Flatten()(layer9)
x = Dense(64, activation='relu')(x)
x = Dense(NUM_CLASSES, activation='softmax')(x)
x = Model(data_input, x, name='mnist')
return x
def VGG19(x, n_classes, is_pretrain=True):
with tf.name_scope('VGG16'):
x = utils.conv('conv1_1', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv1_2', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
with tf.name_scope('pool1'):
x = utils.pool('pool1', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = utils.conv('conv2_1', x, 128, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv2_2', x, 128, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
with tf.name_scope('pool2'):
x = utils.pool('pool2', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = utils.conv('conv3_1', x, 256, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv3_2', x, 256, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv3_3', x, 256, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv3_4', x, 256, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
with tf.name_scope('pool3'):
x = utils.pool('pool3', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = utils.conv('conv4_1', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv4_2', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv4_3', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv4_4', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
with tf.name_scope('pool4'):
x = utils.pool('pool4', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = utils.conv('conv5_1', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv5_2', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv5_3', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv5_4', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
with tf.name_scope('pool5'):
x = utils.pool('pool5', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = utils.FC_layer('fc6', x, out_nodes=4096)
with tf.name_scope('batch_norm1'):
x = utils.batch_norm(x)
x = utils.FC_layer('fc7', x, out_nodes=4096)
with tf.name_scope('batch_norm2'):
x = utils.batch_norm(x)
x = utils.FC_layer('fc8', x, out_nodes=n_classes)
return x
def encoder(self, x_ph, isTr_ph):
with tf.variable_scope(self.name):
l = conv('c1', [5,5,1,64], x_ph)
l = conv('c2', [5,5,64,64], l)
l = layer_bn('b1', l, isTr_ph)
l = pool(l)
l = conv('c3', [5,5,64,128], l)
l = conv('c4', [5,5,128,128], l)
l = layer_bn('b2', l, isTr_ph)
l = pool(l)
l = conv('c5', [5,5,128,256], l)
l = conv('c6', [5,5,256,256], l)
l = conv('c7', [5,5,256,256], l)
l = layer_bn('b3', l, isTr_ph)
l = pool(l)
l = conv('c8', [5,5,256,512], l)
l = conv('c9', [5,5,512,512], l)
l = conv('c10', [5,5,512,512], l)
l = layer_bn('b4', l, isTr_ph)
l = pool(l)
return l
def lenet_conv(inputs, struct, dbb=False):
total_mem = 0
if dbb:
inputs, mem1 = conv(inputs, struct[0], 5, strides=1, padding='VALID', \
dbb=dbb, dbb_bias=13)
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem2 = conv(inputs, struct[1], 5, strides=1, padding='VALID', \
dbb=dbb, dbb_bias=25)
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem3 = flatten(inputs, struct[2], dbb=dbb, dbb_bias=156)
inputs, mem4 = dense(inputs, struct[3], dbb=dbb, dbb_bias=54)
inputs, mem5 = dense(inputs, 10, dbb=False)
else:
inputs, mem1 = conv(inputs,
struct[0],
5,
strides=1,
padding='VALID',
dbb=dbb)
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem2 = conv(inputs,
struct[1],
5,
strides=1,
padding='VALID',
dbb=dbb)
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem3 = flatten(inputs, struct[2], dbb=dbb)
inputs, mem4 = dense(inputs, struct[3], dbb=dbb)
inputs, mem5 = dense(inputs, 10, dbb=False)
total_mem = mem1 + mem2 + mem3 + mem4 + mem5
return inputs, total_mem
def execute(self, features):
"""
Executes the layer in convolution->pooling order
:param features : input features
:type features : numpy nd array
"""
# convolution
convolved_features = self.conv_obj.convolution(
features, self.kernels, self.biases)
# pooling
self.out = pool(convolved_features, self.pool_dim)
def _network(self, in_x, isTr, reuse=False, trainable=True):
with tf.variable_scope(self.name):
l = tf.image.resize_images(in_x, [self.wh, self.wh])
l = conv('c1', [3, 3, 3, 64], l, reuse=reuse, trainable=trainable)
l = layer_bn('b1', l, isTr, reuse, trainable)
l = tf.nn.relu(l)
l = pool(l)
l = conv('c2', [3, 3, 64, 64], l, reuse=reuse, trainable=trainable)
l = layer_bn('b2', l, isTr, reuse, trainable)
l = tf.nn.relu(l)
l = pool(l)
l = conv('c3', [3, 3, 64, 64], l, reuse=reuse, trainable=trainable)
l = layer_bn('b3', l, isTr, reuse, trainable)
l = tf.nn.relu(l)
l = pool(l)
l = conv('c4', [3, 3, 64, 64], l, reuse=reuse, trainable=trainable)
l = layer_bn('b4', l, isTr, reuse, trainable)
l = tf.nn.relu(l)
l = pool(l)
l = tf.contrib.layers.flatten(l)
#l = fc('f1', 1024, l, r=reuse, t=trainable)
return l
def execute(self, features):
"""
Executes the layer in convolution->pooling order
:param features : input features
:type features : numpy nd array
"""
# convolution
convolved_features = self.conv_obj.convolution(features,
self.kernels,
self.biases)
# pooling
self.out = pool(convolved_features, self.pool_dim)
def cnn_model2(IMG_SIZE):
data_input = Input(shape=(IMG_SIZE, IMG_SIZE,3))
layer1 = conv(data_input, 32, (5, 5), padding='same', strides=(1, 1), name='layer1')
x = fire_module(layer1, 1, squeeze=16, expand=64, strides=2)
x = BatchNormalization(axis=3, epsilon=1e-06)(x)
x = fire_module(x,2, squeeze=32, expand=128)
x = BatchNormalization(axis=3, epsilon=1e-06)(x)
x = pool(x, pool_size=(3, 3), strides=(2, 2), padding='same', name='layer2', pool_type='max')
x = fire_module(x, 3, squeeze=16, expand=64,strides=1)
x = BatchNormalization(axis=3, epsilon=1e-06)(x)
x = Flatten()(x)
x = Dense(16, activation='relu')(x)
x = Dense(NUM_CLASSES, activation='softmax')(x)
x = Model(data_input, x, name='mnist')
return x
class Generator:
with tf.variable_scope('encoder'):
inputs = tf.placeholder(dtype=tf.float32, shape=(None, None, None, 3))
representations = [inputs]
representations.append(
tf.layers.conv2d(representations[-1],
filters=32,
kernel_size=5,
padding='same',
activation=tf.nn.relu))
representations.append(
utils.pool(representations[-1], 'MAX', pool_size=2))
representations.append(
tf.layers.conv2d(representations[-1],
filters=32,
kernel_size=3,
padding='same',
activation=tf.nn.relu))
representations.append(
utils.pool(representations[-1], 'MAX', pool_size=2))
representations.append(
tf.layers.conv2d(representations[-1],
filters=16,
kernel_size=3,
padding='same',
activation=tf.nn.relu))
representations.append(
utils.pool(representations[-1], 'MAX', pool_size=2))
skip = utils.pool(inputs, 'AVG',
pool_size=8) # skip connection directly to pixels
skip_appended = tf.concat((representations[-1], skip), axis=-1)
representations.append(
tf.layers.conv2d(representations[-1],
filters=16,
kernel_size=3,
padding='same',
activation=tf.nn.relu))
representations.append(
utils.pool(representations[-1], 'MAX', pool_size=2))
encoding = representations[-1]
with tf.variable_scope('decoder'):
representations.append(
utils.zoom_nearest_neighbor(representations[-1], zoom=2))
representations.append(
tf.layers.conv2d(representations[-1],
filters=16,
kernel_size=3,
padding='same',
activation=tf.nn.relu))
representations.append(
utils.zoom_nearest_neighbor(representations[-1], zoom=2))
representations.append(
tf.layers.conv2d(representations[-1],
filters=16,
kernel_size=3,
padding='same',
activation=tf.nn.relu))
representations.append(
utils.zoom_nearest_neighbor(representations[-1], zoom=2))
representations.append(
tf.layers.conv2d(representations[-1],
filters=32,
kernel_size=3,
padding='same',
activation=tf.nn.relu))
representations.append(
utils.zoom_nearest_neighbor(representations[-1], zoom=2))
representations.append(
tf.layers.conv2d(representations[-1],
filters=32,
kernel_size=5,
padding='same',
activation=tf.nn.relu))
logits = tf.layers.conv2d(representations[-1],
filters=3,
kernel_size=5,
padding='same')
representations.append(logits)
outputs = tf.nn.sigmoid(logits) * 255
representations.append(outputs)
vars = tf.trainable_variables(scope='encoder') + tf.trainable_variables(
scope='decoder')
targets = tf.placeholder(dtype=tf.float32, shape=(None, None, None, 3))
mse_loss = tf.reduce_mean(tf.square(outputs - targets)) / (
255**2) # normalized into interval of size 1
ssim_loss = -tf.reduce_mean(tf.image.ssim(
outputs, targets,
max_val=255)) / 2 # also normalized into interval of size 1