def _create_loss1(self):
pass
x = tf.layers.conv2d(self.lr_input,
self.hidden_size,
1,
activation=None,
name='in')
#self.test=tf.reduce_mean(x)
#low resolution
for i in range(6):
x = utils.crop_by_pixel(x, 1) + self.conv(
x, self.hidden_size, self.bottleneck_size, 'lr_conv' + str(i))
temp = tf.nn.relu(x)
#up sampling
x = tf.image.resize_nearest_neighbor(
x,
tf.shape(x)[1:3] * 2) + tf.layers.conv2d_transpose(
temp, self.hidden_size, 2, strides=2, name='up_sampling')
#high resolution
for i in range(4):
x = utils.crop_by_pixel(x, 1) + self.conv(
x, self.hidden_size, self.bottleneck_size, 'hr_conv' + str(i))
x = tf.nn.relu(x)
self.prediction = tf.layers.conv2d(x, 3, 1, name='out')
self.target_crop = utils.crop_center(self.target,
tf.shape(self.prediction)[1:3])
#self.test=tf.reduce_mean(self.prediction)
#self.test=tf.reduce_mean(self.prediction)
self.loss = tf.losses.mean_squared_error(self.target_crop,
self.prediction)
def _create_struct_without_padding(self):
x = tf.layers.conv2d(self.lr_input,
self.hidden_size,
1,
activation=None,
name='in')
#self.test=tf.reduce_mean(x)
#low resolution
for i in range(6):
x = utils.crop_by_pixel(x, 1) + self.conv(
x, self.hidden_size, self.bottleneck_size, 'lr_conv' + str(i))
temp = tf.nn.relu(x)
#up sampling
x = tf.image.resize_nearest_neighbor(
x,
tf.shape(x)[1:3] * 2) + tf.layers.conv2d_transpose(
temp, self.hidden_size, 2, strides=2, name='up_sampling')
#high resolution
for i in range(4):
x = utils.crop_by_pixel(x, 1) + self.conv(
x, self.hidden_size, self.bottleneck_size, 'hr_conv' + str(i))
x = tf.nn.relu(x)
x = tf.layers.conv2d(x, 3, 1, name='out')
bicubic = tf.image.resize_bicubic(self.lr_input + 128,
tf.shape(self.lr_input)[1:3] * 2,
name='bicubic')
bicubic = utils.crop_center(bicubic, tf.shape(x)[1:3])
self.bmp_prediction = x + bicubic
self.bmp_prediction_cast = tf.saturate_cast(self.bmp_prediction,
tf.uint8)
def _create_loss_without_padding(self):
self.target_crop = utils.crop_center(
self.target,
tf.shape(self.bmp_prediction)[1:3])
self.loss = tf.losses.mean_squared_error(self.target_crop,
self.bmp_prediction)
def zero_filled(kspace):
fourier_op = FFT2(np.ones_like(kspace))
im_recon = np.abs(fourier_op.adj_op(kspace))
im_cropped = crop_center(im_recon, 320)
return im_cropped