def lz_training_net(hazed_batch):
x_s = dt.conv('DN_conv1_1',
hazed_batch,
3,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv1_2',
x_s,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('upsampling_1',
x,
64,
128,
kernel_size=[3, 3],
stride=[1, 2, 2, 1])
x = dt.conv('upsampling_2',
x,
128,
128,
kernel_size=[3, 3],
stride=[1, 2, 2, 1])
x1 = dt.conv('DN_conv1_3',
x,
128,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv2_1',
x1,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv_nonacti('DN_conv2_2',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = tf.add(x, x1)
# x = tools.batch_norm(x)
x = dt.acti_layer(x)
# x = tools.conv('DN_conv2_3', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('DN_conv2_4',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x2 = dt.conv('DN_conv3_1',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv3_2',
x2,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv_nonacti('DN_conv3_3',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = tf.add(x, x2)
# x = tools.batch_norm(x)
x = dt.acti_layer(x)
# x = tools.conv('DN_conv3_4', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('DN_conv3_5',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x3 = dt.conv('DN_conv4_1',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv4_2',
x3,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv4_3',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv_nonacti('DN_conv4_4',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = tf.add(x, x3)
# x = tools.batch_norm(x)
x = dt.acti_layer(x)
x = dt.conv('DN_conv4_5',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x4 = dt.conv('DN_conv5_1',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv5_2',
x4,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv5_3',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv5_4',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv_nonacti('DN_conv5_5',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = tf.add(x, x4)
x = dt.acti_layer(x)
x = dt.deconv('DN_deconv1',
x,
64,
64,
output_shape=[35, 112, 112, 64],
kernel_size=[3, 3],
stride=[1, 2, 2, 1])
x = dt.deconv('DN_deconv2',
x,
64,
64,
output_shape=[35, 224, 224, 64],
kernel_size=[3, 3],
stride=[1, 2, 2, 1])
x = dt.conv('DN_conv6_6',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv_nonacti('DN_conv6_7',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = tf.add(x, x_s)
# # x = tools.batch_norm(x)
x = dt.acti_layer(x)
x = dt.conv('DN_conv6_8',
x,
64,
3,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
return x
def inference(hazed_batch):
"""
:param hazed_batch: The hazed training images from get_distorted_image.
Each image is in the form of Images. 4D tensor of [batch_size, height, width, 3] size
Please refer to CIFAR-10 CNN model to design our dehazenet.
:return: A image batch after trained by CNN
"""
# TODO Lida Xu please re-write the CNN model
with tf.name_scope('DehazeNet'):
x = dt.conv('conv1_1', hazed_batch, 3, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('conv1_2', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# with tf.name_scope('pool1'):
# x = tools.pool('pool1', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = dt.conv('conv2_1', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('conv2_2', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# with tf.name_scope('pool2'):
# x = tools.pool('pool2', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = dt.conv('conv3_1', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('conv3_2', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('conv3_3', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# with tf.name_scope('pool3'):
# x = tools.pool('pool3', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = dt.conv('conv4_1', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('conv4_2', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('conv4_3', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# with tf.name_scope('pool4'):
# x = tools.pool('pool4', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = dt.conv('conv5_1', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('conv5_2', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('conv5_3', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# with tf.name_scope('pool5'):
# x = tools.pool('pool5', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = dt.conv('conv6_1', x, 64, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('conv6_2', x, 64, 3, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# x = tools.conv('conv6_3', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# x = tools.conv('conv6_4', x, 3, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# x = tools.FC_layer('fc6', x, out_nodes=4096)
# with tf.name_scope('batch_norm1'):
# x = tools.batch_norm(x)
# x = tools.FC_layer('fc7', x, out_nodes=4096)
# with tf.name_scope('batch_norm2'):
# x = tools.batch_norm(x)
# x = tools.FC_layer('fc8', x, out_nodes=n_classes)
return x
def lz_training_net(hazed_batch):
x_s = dt.conv('DN_conv1_1',
hazed_batch,
3,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv1_2',
x_s,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
# with tf.name_scope('pool1'):
# x = tools.pool('pool1', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
#
# with tf.name_scope('pool2'):
# x = tools.pool('pool2', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = dt.conv('upsampling_1',
x,
64,
128,
kernel_size=[3, 3],
stride=[1, 2, 2, 1])
x = dt.conv('upsampling_2',
x,
128,
128,
kernel_size=[3, 3],
stride=[1, 2, 2, 1])
x1 = dt.conv('DN_conv1_3',
x,
128,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv2_1',
x1,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv_nonacti('DN_conv2_2',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = tf.add(x, x1)
# x = tools.batch_norm(x)
x = dt.acti_layer(x)
# x = tools.conv('DN_conv2_3', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('DN_conv2_4',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x2 = dt.conv('DN_conv3_1',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv3_2',
x2,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv_nonacti('DN_conv3_3',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = tf.add(x, x2)
# x = tools.batch_norm(x)
x = dt.acti_layer(x)
# x = tools.conv('DN_conv3_4', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = dt.conv('DN_conv3_5',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x3 = dt.conv('DN_conv4_1',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv4_2',
x3,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv4_3',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv_nonacti('DN_conv4_4',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = tf.add(x, x3)
# x = tools.batch_norm(x)
x = dt.acti_layer(x)
x = dt.conv('DN_conv4_5',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x4 = dt.conv('DN_conv5_1',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv5_2',
x4,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv5_3',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv('DN_conv5_4',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv_nonacti('DN_conv5_5',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = tf.add(x, x4)
# x = tools.batch_norm(x)
x = dt.acti_layer(x)
x = dt.deconv('DN_deconv1',
x,
64,
64,
output_shape=[35, 112, 112, 64],
kernel_size=[3, 3],
stride=[1, 2, 2, 1])
# x5 = tools.conv('DN_conv6_1', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# x = tools.conv('DN_conv6_2', x5, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# x = tools.conv_nonacti('DN_conv6_3', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# x = tf.add(x, x5)
# # x = tools.batch_norm(x)
# x = tools.acti_layer(x)
x = dt.deconv('DN_deconv2',
x,
64,
64,
output_shape=[35, 224, 224, 64],
kernel_size=[3, 3],
stride=[1, 2, 2, 1])
x = dt.conv('DN_conv6_6',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
x = dt.conv_nonacti('DN_conv6_7',
x,
64,
64,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
# x6 = tools.conv('DN_conv6_4', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# x = tools.conv('DN_conv6_5', x6, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# x = tools.conv_nonacti('DN_conv6_6', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1])
x = tf.add(x, x_s)
# # x = tools.batch_norm(x)
x = dt.acti_layer(x)
x = dt.conv('DN_conv6_8',
x,
64,
3,
kernel_size=[3, 3],
stride=[1, 1, 1, 1])
# x = tools.conv('conv6_4', x, 3, kernel_size=[3, 3], stride=[1, 1, 1, 1])
# x = tools.FC_layer('fc6', x, out_nodes=4096)
# with tf.name_scope('batch_norm1'):
# x = tools.batch_norm(x)
# x = tools.FC_layer('fc7', x, out_nod
#
# es=4096)
# with tf.name_scope('batch_norm2'):
# x = tools.batch_norm(x)
# x = tools.FC_layer('fc8', x, out_nodes=n_classes)
return x