def recon(self, features, _recon, _noise, SR, name):
with tf.variable_scope(name):
_recon_h1 = int(_recon.get_shape()[1])
_recon_h2 = int(_recon.get_shape()[2])
delta = tf.get_variable(name='delta', initializer=tf.constant(0.1))
eta = tf.get_variable(name='eta', initializer=tf.constant(0.9))
down = bicubic_interp_2d(
_recon,
[int(_recon_h1 / SR), int(_recon_h2 / SR)])
err1 = bicubic_interp_2d(down - _noise, [_recon_h1, _recon_h2])
err2 = _recon - features
out = _recon - delta * (err1 + eta * err2)
return out
def _transform(V, U, out_size):
with tf.variable_scope('_transform'):
num_batch = tf.shape(U)[0]
height = tf.shape(U)[1]
width = tf.shape(U)[2]
num_channels = tf.shape(U)[3]
# grid of (x_t, y_t, 1), eq (1) in ref [1]
height_f = tf.cast(height, 'float32')
width_f = tf.cast(width, 'float32')
out_height = out_size[0]
out_width = out_size[1]
grid = _meshgrid(out_height, out_width) # [2, h*w]
grid = tf.reshape(grid, [-1]) # [2*h*w]
grid = tf.tile(grid, tf.stack([num_batch])) # [n*2*h*w]
grid = tf.reshape(grid, tf.stack([num_batch, 2, -1])) # [n, 2, h*w]
# transform (x, y)^T -> (x+vx, x+vy)^T
V = bicubic_interp_2d(V, out_size)
V = tf.transpose(V, [0, 3, 1, 2]) # [n, 2, h, w]
V = tf.reshape(V, [num_batch, 2, -1]) # [n, 2, h*w]
T_g = tf.add(V, grid) # [n, 2, h*w]
x_s = tf.slice(T_g, [0, 0, 0], [-1, 1, -1])
y_s = tf.slice(T_g, [0, 1, 0], [-1, 1, -1])
x_s_flat = tf.reshape(x_s, [-1])
y_s_flat = tf.reshape(y_s, [-1])
input_transformed = _interpolate(
U, x_s_flat, y_s_flat, out_size)
output = tf.reshape(
input_transformed,
tf.stack([num_batch, out_height, out_width, num_channels]))
return output
def _transform(V, U, out_size):
with tf.variable_scope('_transform'):
num_batch = tf.shape(U)[0]
height = tf.shape(U)[1]
width = tf.shape(U)[2]
num_channels = tf.shape(U)[3]
height_f = tf.cast(height, 'float32')
width_f = tf.cast(width, 'float32')
out_height = out_size[0]
out_width = out_size[1]
grid = _meshgrid(out_height, out_width)
grid = tf.reshape(grid, [-1])
grid = tf.tile(grid, tf.stack([num_batch]))
grid = tf.reshape(grid, tf.stack([num_batch, 2, -1]))
V = bicubic_interp_2d(V, out_size)
V = tf.transpose(V, [0, 3, 1, 2])
V = tf.reshape(V, [num_batch, 2, -1])
T_g = tf.add(V, grid)
x_s = tf.slice(T_g, [0, 0, 0], [-1, 1, -1])
y_s = tf.slice(T_g, [0, 1, 0], [-1, 1, -1])
x_s_flat = tf.reshape(x_s, [-1])
y_s_flat = tf.reshape(y_s, [-1])
input_transformed = _interpolate(U, x_s_flat, y_s_flat, out_size)
output = tf.reshape(
input_transformed,
tf.stack([num_batch, out_height, out_width, num_channels]))
return output
def build_graph(self, _input, _label):
label_h1 = int(_label.get_shape()[1])
label_h2 = int(_label.get_shape()[2])
x = bicubic_interp_2d(_input, [label_h1, label_h2])
y = _input
with tf.variable_scope("bloclk"): # tf.AUTO_REUSE
encode0, down0 = self.Encoding_block(x, name='Encoding_0')
encode1, down1 = self.Encoding_block(down0, name='Encoding_1')
encode2, down2 = self.Encoding_block(down1, name='Encoding_2')
encode3, down3 = self.Encoding_block(down2, name='Encoding_3')
media_end = self.Encoding_block_end(down3, name='Encoding_end')
decode3 = self.Decoding_block(media_end,
encode3,
name='Decoding_3')
decode2 = self.Decoding_block(decode3, encode2, name='Decoding_2')
decode1 = self.Decoding_block(decode2, encode1, name='Decoding_1')
decode0 = self.Decoding_block(decode1, encode0, name='Decoding_0')
decoding_end = self.feature_decoding_end(decode0, name='end')
conv_out = x + decoding_end
x = self.recon(conv_out, x, y, SR=2, name='recon_1')
for i in xrange(5):
with tf.variable_scope("bloclk", reuse=True): #tf.AUTO_REUSE
encode0, down0 = self.Encoding_block(x, name='Encoding_0')
encode1, down1 = self.Encoding_block(down0, name='Encoding_1')
encode2, down2 = self.Encoding_block(down1, name='Encoding_2')
encode3, down3 = self.Encoding_block(down2, name='Encoding_3')
media_end = self.Encoding_block_end(down3, name='Encoding_end')
decode3 = self.Decoding_block(media_end,
encode3,
name='Decoding_3')
decode2 = self.Decoding_block(decode3,
encode2,
name='Decoding_2')
decode1 = self.Decoding_block(decode2,
encode1,
name='Decoding_1')
decode0 = self.Decoding_block(decode1,
encode0,
name='Decoding_0')
decoding_end = self.feature_decoding_end(decode0, name='end')
conv_out = x + decoding_end
x = self.recon(conv_out, x, y, SR=2, name='recon_%d' % (i + 2))
return x
def _decodeBlock(x, shortcut, rows_odd, cols_odd, cweights, bns, activation=LeakyReLU(alpha=ALPHA)):
#Add zero padding on bottom and right if odd dimension required at output,
#giving an output of one greater than required
x = ZeroPadding2D(padding=((0,rows_odd),(0,cols_odd)))(x)
# x = UpSampling2D(size=(2,2), interpolation=UPSAMPLE_INTERP)(x)
up_size = np.array(x.shape)
up_size[1] *= 2
up_size[2] *= 2
x = bicubic_interp_2d(x,(up_size[1],up_size[2]))
#If padding was added, crop the output to match the target shape
#print(rows_odd)
#print(cols_odd)
x = Cropping2D(cropping=((0,rows_odd),(0,cols_odd)))(x)
x = Concatenate()([shortcut, x])
x = res_Block(x, cweights, bns, activation=LeakyReLU(alpha=ALPHA))
return x
def _transform(V, U, out_size):
num_batch = U.size(0)
height = U.size(2)
width = U.size(3)
num_channels = U.size(1)
# grid of (x_t, y_t, 1), eq (1) in ref [1]
# height_f = tf.cast(height, 'float32')
height_f = float(height)
# width_f = tf.cast(width, 'float32')
width_f = float(width)
out_height = out_size[0]
out_width = out_size[1]
grid = _meshgrid(out_height, out_width) # [2, h*w]
grid = grid.reshape(-1) # [2*h*w]
grid = grid.unsqueeze(0).repeat(num_batch, 1) # [n, 2*h*w]
grid = grid.reshape(num_batch, 2, -1) # [n, 2, h*w]
# Set source position (x+vx, y+vy)^T
V = bicubic_interp_2d(V.permute(0, 2, 3, 1), out_size)
V = V.permute(0, 3, 1, 2) # [n, 2, h, w]
V = V.reshape(num_batch, 2, -1) # [n, 2, h*w]
T_g = V + grid # [n, 2, h*w]
x_s = T_g[:, 0, :]
y_s = T_g[:, 1, :]
x_s_flat = x_s.reshape(-1)
y_s_flat = y_s.reshape(-1)
input_transformed = _interpolate(
U.permute(0, 2, 3, 1), x_s_flat, y_s_flat, out_size)
output = input_transformed.reshape(num_batch, out_height, out_width, num_channels)
return output
import tensorflow as tf
from bicubic_interp import bicubic_interp_2d
import numpy as np
import time
n = 6
x = tf.constant(
np.array(range(n*n), dtype=np.float32).reshape([1, n, n, 1]))
size = [3,3]
t0 = time.time()
y0 = bicubic_interp_2d(x, size)
y1 = bicubic_interp_2d(x, size, endpoint=True)
#print "Gen graph : ", time.time() - t0
z = tf.image.resize_bicubic(x, size)
sess = tf.Session()
x_ = sess.run(x)
t0 = time.time()
y0_ = sess.run(y0)
y1_ = sess.run(y1)
#print "Interp : ", time.time() - t0
z_ = sess.run(z)
print "input : [{}, {}]".format(n, n)
print x_[0,:,:,0]
print ""
print "tf.image.resize_bicubic : {}".format(size)
import tensorflow as tf from bicubic_interp import bicubic_interp_2d import numpy as np x = tf.constant(np.array(range(9), dtype=np.float32).reshape([1, 3, 3, 1])) size = [6, 6] y = bicubic_interp_2d(x, size) z = tf.image.resize_bicubic(x, size) sess = tf.Session() x_ = sess.run(x) y_ = sess.run(y) z_ = sess.run(z) print x_[0, :, :, 0] print y_[0, :, :, 0] print z_[0, :, :, 0]
def generator_model(args, inputs, istrain, reuse):
"""Build generator architecture."""
# inputs: tensor with shape [bn, 256,256, 1]
# inputs = Input(shape=input_shape_generator)
with tf.variable_scope('gen_', reuse=reuse):
x = ReflectionPadding2D((3, 3))(inputs)
x = Conv2D(filters=ngf, kernel_size=(7, 7), padding='valid')(x)
x = batch_norm(x, "bn1", is_train=istrain)
x = Activation('relu')(x)
# x = MaxPooling2D((2, 2), padding='same')(x)e')(x)
# x = Conv2D(filters=ngf, kernel_size=(7,7), padding='same')(x)
# x = batch_norm(x, "bn2", is_train=istrain)
# x = Activation('relu')(x)
n_downsampling = 2
for i in range(n_downsampling):
mult = 2**i
# x = ReflectionPadding2D((2, 2))(x)
if args.max_pooling == False:
x = Conv2D(filters=ngf * mult * 2,
kernel_size=(3, 3),
strides=2,
padding='valid')(x)
else:
x = Conv2D(filters=ngf * mult * 2,
kernel_size=(3, 3),
strides=1,
padding='valid')(x)
x = MaxPooling2D((2, 2), padding='valid')(x)
# x = BatchNormalization()(x, training=istrain)
x = batch_norm(x, "down_bn_" + str(i), is_train=istrain)
tf.summary.histogram('before_active', x)
x = Activation('relu')(x)
tf.summary.histogram('after_activate', x)
mult = 2**n_downsampling
for i in range(n_blocks_gen):
x = res_block(x, ngf * mult, use_dropout=True)
# for i in range(n_downsampling):
# mult = 2**(n_downsampling - i)
# x = UpSampling2D()(x)
# x = Conv2D(filters=int(ngf * mult / 2),kernel_size=(3,3),padding='same')(x)
## x = Conv2DTranspose(filters=int(ngf * mult / 2), kernel_size=(3, 3), strides=2, padding='same')(x)
## x = BatchNormalization()(x, training=istrain)
# x = batch_norm(x, "up_bn_"+str(i), is_train=istrain)
# x = LeakyReLU(alpha=0.3)(x)
x = Conv2D(filters=2, kernel_size=(5, 5), padding='same')(x)
x = batch_norm(x, "final", is_train=istrain)
wrap = Activation('sigmoid')(x)
wrap = tf.multiply(tf.add(wrap, -0.5), 8)
# dense layer
dense = tf.layers.flatten(wrap)
output_size = 128
output_size = args.final_layer # we use the args value here to decide the final layer number
# output_size1 = 16
dense_out = tf.layers.dense(inputs=dense, units=output_size * 2)
# dense_out1 = tf.layers.dense(inputs=dense_out, units=output_size1*2)
x_mean = tf.reshape(dense_out, [-1, output_size, 2])
# x_mean = Conv2D(filters=2, kernel_size=(1,256), padding='valid')(wrap)
# x_layer = wrap[...,0]
# x_mean = tf.reduce_max(wrap, axis=2)
x_mean = tf.expand_dims(x_mean, 2)
wrap = tf.tile(x_mean, multiples=[1, 1, output_size, 1])
wrap = bicubic_interp_2d(wrap, imsize)
outputs = Lambda(WarpST_one,
arguments={
'inputs': inputs,
'name': str(random.random())
})(wrap)
return outputs, wrap[:, :, 0, :]
