def resnet_conv(input,
kernel_size,
is_training,
stride=1,
num_outputs=None,
normalizer_fn=None,
activation_fn=None,
scope=None):
with variable_scope.variable_scope(scope, 'resnet_conv', [input]) as sc:
last_dim = utils.last_dimension(input.get_shape())
num_outputs = num_outputs or last_dim
output = layers.convolution(inputs=input,
num_outputs=last_dim,
kernel_size=[1, 1, 1])
output = layers.convolution(inputs=output,
num_outputs=last_dim,
kernel_size=[1, 1, kernel_size])
output = layers.convolution(inputs=output,
num_outputs=last_dim,
kernel_size=[1, kernel_size, 1])
output = layers.convolution(
inputs=output,
num_outputs=num_outputs,
kernel_size=[kernel_size, 1, 1],
activation_fn=activation_fn,
normalizer_fn=normalizer_fn,
normalizer_params={'is_training': is_training},
stride=stride)
return output
def brightness_predict_net(inp_img, inp_wb, inp_iso, inp_t):
h = tf.shape(inp_img)[1]
w = tf.shape(inp_img)[2]
inp_img_normed = normalize_var(inp_img, 0, 0.00018782060752181462)
layer_inp_wb = tf.ones([1, h, w, 4]) * normalize_var(
inp_wb, 1.5084804, 0.6111772 * 0.6111772)
layer_inp_iso = tf.ones([1, h, w, 1]) * normalize_var(
inp_iso, 471.72858, 264.23517 * 264.23517)
layer_inp_t = tf.ones([1, h, w, 1]) * normalize_var(
inp_t, 0.08489772, 0.1368367 * 0.1368367)
inp_concated = tf.concat([ inp_img_normed, \
layer_inp_wb, \
layer_inp_iso, \
layer_inp_t],3)
inp_concated.set_shape(shape=(pa.bs, None, None, 4 + 4 + 2))
with tf.variable_scope('brightness_predict_net'):
if pa.ps != 512:
input_BatchRS = tf.image.resize_nearest_neighbor(
inp_concated, (512, 512))
else:
input_BatchRS = inp_concated
c1 = convolution(input_BatchRS, 16, [3, 3], stride=4, \
activation_fn=lrelu,trainable = pa.train_brightness_net) #128*128*3
c2 = convolution(c1, 16, [3, 3], stride=4, \
activation_fn=lrelu,trainable = pa.train_brightness_net) #32*32*8
c3 = convolution(c2, 32, [3, 3], stride=2, \
activation_fn=lrelu,trainable = pa.train_brightness_net) #16*16*16
c4 = convolution(c3, 32, [3, 3], stride=2, \
activation_fn=lrelu,trainable = pa.train_brightness_net) #8*8*32
c5 = convolution(c4, 4, [3, 3], stride=2, \
activation_fn=lrelu,trainable = pa.train_brightness_net) #4*4*4
c5_vec = tf.reshape(c5, shape=[tf.shape(c5)[0], -1])
c6_vec = layer_fc(c5_vec,
4 * 4 * 4,
32,
relu_free=False,
name='dense1',
trainable=pa.train_brightness_net)
gt_time = layer_fc(c6_vec,
32,
1,
relu_free=True,
name='dense2',
trainable=pa.train_brightness_net)
return gt_time
def build_network(input_hold):
net = input_hold
_layers = {'input': net}
# Block 1 : Separable CNN
net = layers.separable_conv2d(
inputs=net,
num_outputs=32,
kernel_size=4,
stride=2,
padding='VALID',
depth_multiplier=4,
)
_layers['sepCNN1'] = net
# Block 2 : Attention (with residual connection)
net, att_layers = non_local_nn_2d(net,
16,
pool=False,
name='non_local',
return_layers=True)
_layers['attention'] = att_layers['attention']
_layers['NLNN'] = net
# Block 3 : Convolution
net = layers.convolution(inputs=net,
num_outputs=64,
kernel_size=3,
stride=2,
padding='VALID')
_layers['CNN1'] = net
net = layers.convolution(inputs=net,
num_outputs=64,
kernel_size=2,
stride=2,
padding='VALID')
_layers['CNN2'] = net
# Block 4 : Feature Vector
net = layers.flatten(net)
_layers['flat'] = net
net = layers.fully_connected(
net,
128,
activation_fn=None,
)
_layers['dense1'] = net
return net, _layers
def conv_model(feature, target, mode):
"""2-layer convolution model."""
# Convert the target to a one-hot tensor of shape (batch_size, 10) and
# with a on-value of 1 for each one-hot vector of length 10.
target = tf.one_hot(tf.cast(target, tf.int32), 10, 1, 0)
# Reshape feature to 4d tensor with 2nd and 3rd dimensions being
# image width and height final dimension being the number of color channels.
feature = tf.reshape(feature, [-1, 28, 28, 1])
# First conv layer will compute 32 features for each 5x5 patch
with tf.variable_scope('conv_layer1'):
h_conv1 = layers.convolution(feature,
32,
kernel_size=[5, 5],
activation_fn=tf.nn.relu)
h_pool1 = max_pool_2x2(h_conv1)
# Second conv layer will compute 64 features for each 5x5 patch.
with tf.variable_scope('conv_layer2'):
h_conv2 = layers.convolution(h_pool1,
64,
kernel_size=[5, 5],
activation_fn=tf.nn.relu)
h_pool2 = max_pool_2x2(h_conv2)
# reshape tensor into a batch of vectors
h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
# Densely connected layer with 1024 neurons.
h_fc1 = layers.dropout(layers.fully_connected(h_pool2_flat,
1024,
activation_fn=tf.nn.relu),
keep_prob=0.5,
is_training=mode == tf.contrib.learn.ModeKeys.TRAIN)
# Compute logits (1 per class) and compute loss.
logits = layers.fully_connected(h_fc1, 10, activation_fn=None)
loss = tf.contrib.losses.softmax_cross_entropy(logits, target)
# Create a tensor for training op.
train_op = layers.optimize_loss(loss,
tf.contrib.framework.get_global_step(),
optimizer='SGD',
learning_rate=0.001)
return tf.argmax(logits, 1), loss, train_op
def down_layer(inputs, depth, model_config):
"""A single encoder layer."""
normalizer, normalizer_params = get_normalizer_and_mode(model_config)
return contrib_layers.convolution(
inputs=inputs,
num_outputs=depth,
kernel_size=3,
stride=2,
padding='SAME',
normalizer_fn=normalizer,
normalizer_params=normalizer_params,
activation_fn=tf.nn.leaky_relu)
def conv_model(feature, target, mode):
"""2-layer convolution model."""
# Convert the target to a one-hot tensor of shape (batch_size, 10) and
# with a on-value of 1 for each one-hot vector of length 10.
target = tf.one_hot(tf.cast(target, tf.int32), 10, 1, 0)
# Reshape feature to 4d tensor with 2nd and 3rd dimensions being
# image width and height final dimension being the number of color channels.
feature = tf.reshape(feature, [-1, 28, 28, 1])
# First conv layer will compute 32 features for each 5x5 patch
with tf.variable_scope('conv_layer1'):
h_conv1 = layers.convolution(feature, 32, kernel_size=[5, 5],
activation_fn=tf.nn.relu)
h_pool1 = max_pool_2x2(h_conv1)
# Second conv layer will compute 64 features for each 5x5 patch.
with tf.variable_scope('conv_layer2'):
h_conv2 = layers.convolution(h_pool1, 64, kernel_size=[5, 5],
activation_fn=tf.nn.relu)
h_pool2 = max_pool_2x2(h_conv2)
# reshape tensor into a batch of vectors
h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
# Densely connected layer with 1024 neurons.
h_fc1 = layers.dropout(
layers.fully_connected(
h_pool2_flat, 1024, activation_fn=tf.nn.relu), keep_prob=0.5,
is_training=mode == tf.contrib.learn.ModeKeys.TRAIN)
# Compute logits (1 per class) and compute loss.
logits = layers.fully_connected(h_fc1, 10, activation_fn=None)
loss = tf.contrib.losses.softmax_cross_entropy(logits, target)
# Create a tensor for training op.
train_op = layers.optimize_loss(
loss, tf.contrib.framework.get_global_step(), optimizer='SGD',
learning_rate=0.001)
return tf.argmax(logits, 1), loss, train_op
def conv_layer(inputs, depth, model_config):
"""A single 3x3 convolutional layer with stride 1."""
normalizer, normalizer_params = get_normalizer_and_mode(model_config)
ibs, ih, iw, ic = inputs.get_shape().as_list()
outputs = contrib_layers.convolution(
inputs=inputs,
num_outputs=depth,
kernel_size=3,
stride=1,
padding='SAME',
normalizer_fn=normalizer,
normalizer_params=normalizer_params,
activation_fn=tf.nn.leaky_relu)
obs, oh, ow, oc = outputs.get_shape().as_list()
log.info(
'Applying conv+relu layer. Input: [%i, %i, %i, %i]. Output: [%i, %i, %i, %i].'
% (ibs, ih, iw, ic, obs, oh, ow, oc))
return outputs
def up_layer(inputs, spatial_dims, depth, model_config):
"""A single decoder layer."""
normalizer, normalizer_params = get_normalizer_and_mode(model_config)
if len(tf.shape(inputs)) != 4:
if len(tf.shape(inputs)) != 5:
raise ValueError('Unexpected input dimensionality: %i' %
len(tf.shape(inputs)))
raise ValueError('3D Upsampling has not been implemented.')
layer = contrib_layers.convolution(
inputs=inputs,
num_outputs=depth,
kernel_size=5,
padding='SAME',
normalizer_fn=normalizer,
normalizer_params=normalizer_params,
activation_fn=tf.nn.leaky_relu)
return tf.image.resize_images(
images=layer, size=spatial_dims, align_corners=True)
def res_net(
features,
res_blocks_size,
num_classes,
res_func,
multi_k=1,
is_training=True,
keep_prob=1.0,
is_add_multiplier=False,
is_double_size=False,
scope=None):
with tf.variable_scope(scope, 'drl_mp', [features]):
# 128 / 1
net = features
with tf.variable_scope('init'):
first_output = np.array(res_blocks_size).flatten()[0]
if(first_output < 16):
raise ValueError("The first output size should be equal or greater than 16: {0}".format(first_output))
net = layers.batch_norm(
inputs=net,
activation_fn=tf.nn.relu,
is_training=is_training)
# 64 / 4
net = layers.convolution(
inputs=net,
num_outputs=4,
kernel_size=[1, 1, 1],
stride=2)
if not is_double_size:
# 32 / 8
net = ops.resnet_reduction(
input=net,
kernel_size=3,
is_training=is_training)
# 16 / 16
net = ops.resnet_reduction(
input=net,
kernel_size=3,
is_training=is_training)
kpc = _KeepProbCalc(
min_keep_prob=keep_prob,
total_block_num=sum(len(x) for x in res_blocks_size))
for block_num, block in enumerate(res_blocks_size):
with tf.variable_scope(str.format('res_{0}', (block_num + 1))):
for b_num, b in enumerate(block):
is_half_size = (block_num != 0 and b_num == 0)
residual_scope = str.format('res_{0}_{1}', (block_num + 1), (b_num + 1))
net = multi_residual(
inputs=net,
num_outputs=b,
multi_k=multi_k,
res_func=res_func,
scope=residual_scope,
keep_prob=kpc.get_next_decr_prob(),
is_training=is_training,
is_half_size=is_half_size,
is_add_multiplier=is_add_multiplier)
with tf.variable_scope('unit_last'):
net = layers.batch_norm(
inputs=net,
activation_fn=tf.nn.relu,
is_training=is_training)
net = tf.reduce_mean(net, [1, 2, 3])
net = layers.fully_connected(
inputs=net,
num_outputs=num_classes,
activation_fn=None,
weights_initializer=tf.uniform_unit_scaling_initializer(factor=1.0),
biases_initializer=tf.constant_initializer())
return net
def quality_pri_net(inp_img, inp_wb, inp_iso, inp_t, gt_t):
h = tf.shape(inp_img)[1]
w = tf.shape(inp_img)[2]
inp_img_normed = normalize_var(inp_img, 0, 0.00018782060752181462)
#inp_img_normed = tf.Print(inp_img_normed,['inp_img_normed:',tf.shape(inp_img)])
layer_inp_wb = tf.ones([1, h, w, 4]) * normalize_var(
inp_wb, 1.5084804, 0.6111772 * 0.6111772)
layer_inp_iso = tf.ones([1, h, w, 1]) * normalize_var(
inp_iso, 471.72858, 264.23517 * 264.23517)
layer_inp_t = tf.ones([1, h, w, 1]) * normalize_var(
inp_t, 0.08489772, 0.1368367 * 0.1368367)
layer_gt_t = tf.ones([1, h, w, 1]) * normalize_var(gt_t, 0.42407602,
0.37255558 * 0.37255558)
inp_concated = tf.concat([ inp_img_normed, \
layer_inp_wb, \
layer_inp_iso, \
layer_inp_t, \
layer_gt_t],3)
inp_concated.set_shape(shape=(pa.bs, None, None, 4 + 4 + 3))
with tf.variable_scope('prefix_net'):
conv0 = convolution(inp_concated,
16, [3, 3],
activation_fn=lrelu,
trainable=pa.train_prefix,
scope='g_conv0_1')
conv0 = convolution(conv0,
16, [3, 3],
activation_fn=lrelu,
trainable=pa.train_prefix,
scope='g_conv0_2')
conv0 = convolution(conv0,
32, [3, 3],
activation_fn=lrelu,
trainable=pa.train_prefix,
scope='g_conv0_3')
conv0 = convolution(conv0,
32, [3, 3],
activation_fn=lrelu,
trainable=pa.train_prefix,
scope='g_conv0_4')
convp1 = convolution(conv0,
32, [3, 3],
activation_fn=lrelu,
trainable=pa.train_prefix,
scope='g_conv0_5')
with tf.variable_scope('quality_pri_net'):
conv1 = convolution(convp1,
32, [3, 3],
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv1_4')
conv1 = convolution(conv1,
32, [3, 3],
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv1_5')
pool1 = pool2d(conv1, [2, 2], padding='SAME')
conv2 = convolution(pool1,
64, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv2_1')
conv2 = convolution(conv2,
64, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv2_2')
pool2 = pool2d(conv2, [2, 2], padding='SAME')
conv3 = convolution(pool2,
128, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv3_1')
conv3 = convolution(conv3,
128, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv3_2')
pool3 = pool2d(conv3, [2, 2], padding='SAME')
conv4 = convolution(pool3,
256, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv4_1')
conv4 = convolution(conv4,
256, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv4_2')
pool4 = pool2d(conv4, [2, 2], padding='SAME')
conv5 = convolution(pool4,
512, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv5_1')
conv5 = convolution(conv5,
512, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv5_2')
up6 = upsample_and_concat(conv5, conv4, 256, 512)
conv6 = convolution(up6,
256, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv6_1')
conv6 = convolution(conv6,
256, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv6_2')
up7 = upsample_and_concat(conv6, conv3, 128, 256)
conv7 = convolution(up7,
128, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv7_1')
conv7 = convolution(conv7,
128, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv7_2')
up8 = upsample_and_concat(conv7, conv2, 64, 128)
conv8 = convolution(up8,
64, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv8_1')
conv8 = convolution(conv8,
64, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv8_2')
up9 = upsample_and_concat(conv8, conv1, 32, 64)
conv9 = convolution(up9,
32, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv9_1')
conv9 = convolution(conv9,
32, [3, 3],
rate=1,
activation_fn=lrelu,
trainable=pa.train_u_net,
scope='g_conv9_2')
conv10 = convolution(conv9,
12, [1, 1],
rate=1,
activation_fn=None,
trainable=pa.train_u_net,
scope='g_conv10')
out = tf.depth_to_space(conv10, 2)
return out, convp1