def dark_net_conv_1(ipt,
i,
name='dark_net_conv_1',
reuse=False,
is_training=True,
norm='batch'):
c1s1k32 = ops.conv2d(ipt,
32,
1,
1,
False,
norm=norm,
activation=ops.leaky_relu,
name=name + str(i),
reuse=reuse,
is_training=is_training)
c3s1k64 = ops.conv2d(c1s1k32,
64,
3,
1,
True,
norm=norm,
activation=ops.leaky_relu,
name=name + str(i + 1),
reuse=reuse,
is_training=is_training)
return tf.add(c3s1k64, ipt, name='shortcut_0')
def dark_net_conv_5(ipt,
i,
j,
name='dark_net_conv_5',
reuse=False,
is_training=True,
norm='batch'):
c1s1k512 = ops.conv2d(ipt,
512,
1,
1,
False,
norm=norm,
activation=ops.leaky_relu,
name=name + str(i),
reuse=reuse,
is_training=is_training)
c3s1k1024 = ops.conv2d(c1s1k512,
1024,
3,
1,
True,
norm=norm,
activation=ops.leaky_relu,
name=name + str(i + 1),
reuse=reuse,
is_training=is_training)
return tf.add(c3s1k1024, ipt, name='shortcut_' + str(j + 19))
def dark_net_conv_2(ipt,
i,
j,
name='dark_net_conv_2',
reuse=False,
is_training=True,
norm='batch'):
c1s1k64 = ops.conv2d(ipt,
64,
1,
1,
False,
norm=norm,
activation=ops.leaky_relu,
name=name + str(i),
reuse=reuse,
is_training=is_training)
c3s1k128 = ops.conv2d(c1s1k64,
128,
3,
1,
True,
norm=norm,
activation=ops.leaky_relu,
name=name + str(i + 1),
reuse=reuse,
is_training=is_training)
if j == 0:
return tf.add(c3s1k128, ipt, name='shortcut_1')
else:
return tf.add(c3s1k128, ipt, name='shortcut_2')
def dark_net_conv_3(ipt,
i,
j,
name='dark_net_conv_3',
reuse=False,
is_training=True,
norm='batch'):
c1s1k128 = ops.conv2d(ipt,
128,
1,
1,
False,
norm=norm,
activation=ops.leaky_relu,
name=name + str(i),
reuse=reuse,
is_training=is_training)
c3s1k256 = ops.conv2d(c1s1k128,
256,
3,
1,
True,
norm=norm,
activation=ops.leaky_relu,
name=name + str(i + 1),
reuse=reuse,
is_training=is_training)
return tf.add(c3s1k256, ipt, name='shotcut_' + str(j + 3))
def inception_module_factorization(inputs, scope=""):
"""
inception module after factorization of n x n convolutions, as illustrated in figure 6 in paper
:param inputs: input tensor with 784 channel
:param scope: scope name
:return: tensor
"""
with tf.variable_scope(scope):
with tf.variable_scope("branch1x1"):
branch1x1 = conv2d(inputs, 192, [1, 1])
with tf.variable_scope("branch7x7"):
branch7x7 = conv2d(inputs, 160, [1, 1])
branch7x7 = conv2d(branch7x7, 160, [1, 7])
branch7x7 = conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope("branch7x7db"):
branch7x7db = conv2d(inputs, 160, [1, 1])
branch7x7db = conv2d(branch7x7db, 160, [7, 1])
branch7x7db = conv2d(branch7x7db, 160, [1, 7])
branch7x7db = conv2d(branch7x7db, 160, [7, 1])
branch7x7db = conv2d(branch7x7db, 192, [1, 7])
with tf.variable_scope("branch_pool"):
branch_pool = tf.nn.avg_pool(inputs, [1, 3, 3, 1], [1, 1, 1, 1],
padding="SAME")
branch_pool = conv2d(branch_pool, 192, [1, 1])
net = tf.concat(
axis=3, values=[branch1x1, branch7x7, branch7x7db, branch_pool])
return net
def inception_module_expanded(inputs, scope=""):
with tf.variable_scope(scope):
with tf.variable_scope('branch1x1'):
branch1x1 = conv2d(inputs, 320, [1, 1])
with tf.variable_scope('branch3x3'):
branch3x3 = conv2d(inputs, 384, [1, 1])
branch3x3 = tf.concat(axis=3,
values=[
conv2d(branch3x3, 384, [1, 3]),
conv2d(branch3x3, 384, [3, 1])
])
with tf.variable_scope('branch3x3db'):
branch3x3db = conv2d(inputs, 448, [1, 1])
branch3x3db = conv2d(branch3x3db, 384, [3, 3])
branch3x3db = tf.concat(axis=3,
values=[
conv2d(branch3x3db, 384, [1, 3]),
conv2d(branch3x3db, 384, [3, 1])
])
with tf.variable_scope("branch_pool"):
branch_pool = tf.nn.avg_pool(inputs, [1, 3, 3, 1], [1, 1, 1, 1],
padding='SAME')
branch_pool = conv2d(branch_pool, 192, [1, 1])
net = tf.concat(
axis=3, values=[branch1x1, branch3x3, branch3x3db, branch_pool])
return net
def inception_module_origin(inputs, scope=""):
"""
The original inception module, change 5x5 to two 3x3
:param inputs: inputs with channel of 192
:param scope: scope name
:return: tensor
"""
# 61 * 61 * 192
with tf.variable_scope(scope):
with tf.variable_scope('branch1x1'):
branch1x1 = conv2d(inputs, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = conv2d(inputs, 48, [1, 1])
branch5x5 = conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3'):
branch3x3 = conv2d(inputs, 64, [1, 1])
branch3x3 = conv2d(branch3x3, 96, [3, 3])
branch3x3 = conv2d(branch3x3, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = tf.nn.avg_pool(inputs, [1, 3, 3, 1], [1, 1, 1, 1],
padding='SAME')
branch_pool = conv2d(branch_pool, 64, [1, 1])
net = tf.concat(axis=3,
values=[branch1x1, branch5x5, branch3x3, branch_pool])
return net
def c3s2k64(ipt, name='c3s2k64', reuse=False, is_training=True, norm='batch'):
return ops.conv2d(ipt,
64,
3,
2,
True,
norm=norm,
activation=ops.leaky_relu,
name=name,
reuse=reuse,
is_training=is_training)
def c3s1k32(ipt, name='c3s1k32', reuse=False, is_training=True, norm='batch'):
return ops.conv2d(ipt,
32,
3,
1,
True,
norm=norm,
activation=ops.leaky_relu,
name=name,
reuse=reuse,
is_training=is_training)
def inception_reduction_v2(inputs, scope=""):
"""
module used for reduce dimension, with different input and output shape
:param inputs: inputs tensor of 768
:param scope: scope name
:return: output tensor of 1280
"""
with tf.variable_scope(scope):
with tf.variable_scope('branch3x3'):
branch3x3 = conv2d(inputs, 192, [1, 1])
branch3x3 = conv2d(branch3x3,
320, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch7x7x3'):
branch7x7x3 = conv2d(inputs, 192, [1, 1])
branch7x7x3 = conv2d(branch7x7x3, 192, [1, 7])
branch7x7x3 = conv2d(branch7x7x3, 192, [7, 1])
branch7x7x3 = conv2d(branch7x7x3,
192, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope("branch_pool"):
branch_pool = tf.nn.max_pool(inputs, [1, 3, 3, 1], [1, 2, 2, 1],
padding='VALID')
net = tf.concat(axis=3, values=[branch3x3, branch7x7x3, branch_pool])
return net
def discriminator(self, image, conditioning, training, reuse=False):
with tf.variable_scope("discriminator", reuse=reuse):
net = conv2d(image, self.d_dim,
training=training) # (14 x 14 x 64)
net = conv2d(net, self.d_dim * 2,
training=training) # (7 x 7 x 128)
net = conv2d(net, self.d_dim * 4,
training=training) # (4 x 4 x 128)
conditioning = tf.reshape(conditioning,
[-1, 1, 1, conditioning.shape[1]])
conditioning = tf.tile(conditioning,
[1, net.shape[1], net.shape[2], 1])
net = tf.concat([conditioning, net], axis=-1) # (4 x 4 x 601)
net = conv2d(net, self.d_dim * 4, kernels=1,
strides=1) # (4 x 4 x 256)
net = layers.conv2d(net, 1, kernel_size=4,
strides=1) # (1 x 1 x 1)
net = tf.nn.sigmoid(net)
net = tf.squeeze(net)
return net
def inception_reduction_v1(inputs, scope=""):
"""
module used for reduce dimension, as illustrated in figure 10 in the paper
:param inputs: tensor of shape 288
:param scope: scope name
:return: tensor
"""
with tf.variable_scope(scope):
with tf.variable_scope("branch3x3"):
branch3x3 = conv2d(inputs, 384, [3, 3], stride=2, padding='VALID')
with tf.variable_scope("branch3x3db"):
branch3x3db = conv2d(inputs, 64, [1, 1])
branch3x3db = conv2d(branch3x3db, 96, [3, 3])
branch3x3db = conv2d(branch3x3db,
96, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope("branch_pool"):
branch_pool = tf.nn.max_pool(inputs, [1, 3, 3, 1], [1, 2, 2, 1],
padding='VALID')
net = tf.concat(axis=3, values=[branch3x3, branch3x3db, branch_pool])
return net
def yolo_3(ipt1,
ipt2,
num_classes,
reuse=False,
is_training=True,
norm='batch',
activation=ops.leaky_relu):
c1s1k128 = ops.conv2d(ipt1,
128,
1,
1,
True,
norm=norm,
activation=activation,
name='conv_67',
reuse=reuse,
is_training=is_training)
with tf.variable_scope('upsample_1'):
upsample = tf.image.resize_nearest_neighbor(
c1s1k128, (c1s1k128.shape[1] * 2, c1s1k128.shape[2] * 2))
with tf.variable_scope('route_1'):
route = tf.concat([upsample, ipt2], axis=3)
c1s1k128 = ops.conv2d(route,
128,
1,
1,
False,
norm=norm,
activation=activation,
name='conv_68',
reuse=reuse,
is_training=is_training)
c3s1k256 = ops.conv2d(c1s1k128,
256,
3,
1,
True,
norm=norm,
activation=activation,
name='conv_69',
reuse=reuse,
is_training=is_training)
c1s1k128 = ops.conv2d(c3s1k256,
128,
1,
1,
False,
norm=norm,
activation=activation,
name='conv_70',
reuse=reuse,
is_training=is_training)
c3s1k256 = ops.conv2d(c1s1k128,
256,
3,
1,
True,
norm=norm,
activation=activation,
name='conv_71',
reuse=reuse,
is_training=is_training)
c1s1k128 = ops.conv2d(c3s1k256,
128,
1,
1,
False,
norm=norm,
activation=activation,
name='conv_72',
reuse=reuse,
is_training=is_training)
c3s1k256 = ops.conv2d(c1s1k128,
256,
3,
1,
True,
norm=norm,
activation=activation,
name='conv_73',
reuse=reuse,
is_training=is_training)
small_obj_raw_detections = ops.conv2d(c3s1k256,
3 * (num_classes + 5),
1,
1,
False,
norm=None,
activation=None,
use_bias=True,
name='conv_74',
reuse=reuse,
is_training=is_training)
return small_obj_raw_detections
def yolo_2(ipt1,
ipt2,
num_classes,
reuse=False,
is_training=True,
norm='batch',
activation=ops.leaky_relu):
c1s1k256 = ops.conv2d(ipt1,
256,
1,
1,
False,
norm=norm,
activation=activation,
name='conv_59',
reuse=reuse,
is_training=is_training)
with tf.variable_scope('upsample_0'):
upsample = tf.image.resize_nearest_neighbor(
c1s1k256, (c1s1k256.shape[1] * 2, c1s1k256.shape[2] * 2))
with tf.variable_scope('route_0'):
route = tf.concat([upsample, ipt2], axis=3)
c1s1k256 = ops.conv2d(route,
256,
1,
1,
False,
norm=norm,
activation=activation,
name='conv_60',
reuse=reuse,
is_training=is_training)
c3s1k512 = ops.conv2d(c1s1k256,
512,
3,
1,
True,
norm=norm,
activation=activation,
name='conv_61',
reuse=reuse,
is_training=is_training)
c1s1k256 = ops.conv2d(c3s1k512,
256,
1,
1,
False,
norm=norm,
activation=activation,
name='conv_62',
reuse=reuse,
is_training=is_training)
c3s1k512 = ops.conv2d(c1s1k256,
512,
3,
1,
True,
norm=norm,
activation=activation,
name='conv_63',
reuse=reuse,
is_training=is_training)
c1s1k256 = ops.conv2d(c3s1k512,
256,
1,
1,
False,
norm=norm,
activation=activation,
name='conv_64',
reuse=reuse,
is_training=is_training)
yolo_route = c1s1k256
c3s1k512 = ops.conv2d(c1s1k256,
512,
3,
1,
True,
norm=norm,
activation=activation,
name='conv_65',
reuse=reuse,
is_training=is_training)
medium_obj_raw_detections = ops.conv2d(c3s1k512,
3 * (num_classes + 5),
1,
1,
False,
norm=None,
activation=None,
use_bias=True,
name='conv_66',
reuse=reuse,
is_training=is_training)
return medium_obj_raw_detections, yolo_route
def yolo_1(ipt,
num_classes,
reuse=False,
is_training=True,
norm='bath',
activation=ops.leaky_relu):
c1s1k512 = ops.conv2d(ipt,
512,
1,
1,
False,
norm=norm,
activation=activation,
name='conv_52',
reuse=reuse,
is_training=is_training)
c3s1k1024 = ops.conv2d(c1s1k512,
1024,
3,
1,
True,
norm=norm,
activation=activation,
name='conv_53',
reuse=reuse,
is_training=is_training)
c1s1k512 = ops.conv2d(c3s1k1024,
512,
1,
1,
False,
norm=norm,
activation=activation,
name='conv_54',
reuse=reuse,
is_training=is_training)
c3s1k1024 = ops.conv2d(c1s1k512,
1024,
3,
1,
True,
norm=norm,
activation=activation,
name='conv_55',
reuse=reuse,
is_training=is_training)
c1s1k512 = ops.conv2d(c3s1k1024,
512,
1,
1,
False,
norm=norm,
activation=activation,
name='conv_56',
reuse=reuse,
is_training=is_training)
yolo_route = c1s1k512
c3s1k1024 = ops.conv2d(c1s1k512,
1024,
3,
1,
True,
norm=norm,
activation=activation,
name='conv_57',
reuse=reuse,
is_training=is_training)
large_obj_raw_detections = ops.conv2d(c3s1k1024,
3 * (num_classes + 5),
1,
1,
False,
norm=None,
activation=None,
use_bias=True,
name='conv_58',
reuse=reuse,
is_training=is_training)
return large_obj_raw_detections, yolo_route
def inception_v3(inputs,
dropout_keep_prob=0.8,
num_classes=2,
is_training=True,
restore_logits=True,
scope=""):
with tf.name_scope(scope, "inception_v3", [inputs]):
# 512 * 512 * 3
conv0 = conv2d(inputs,
32, [3, 3],
stride=2,
scope='conv0',
padding='VALID')
# 255 * 255 * 32
conv1 = conv2d(conv0, 32, [3, 3], scope='conv1', padding='VALID')
# 253 * 253 * 32
conv2 = conv2d(conv1, 64, [3, 3], scope='conv2')
# 253 * 253 * 64
with tf.name_scope('pool1') as scope:
pool1 = tf.nn.max_pool(conv2, [1, 3, 3, 1], [1, 2, 2, 1],
padding='VALID',
name=scope)
# 126 * 126 * 64
conv3 = conv2d(conv2, 80, [1, 1], scope='conv3', padding='VALID')
# 126 * 126 * 80
conv4 = conv2d(conv3, 192, [3, 3], scope='conv4', padding='VALID')
# 124 * 124 * 192
with tf.name_scope('pool2') as scope:
pool2 = tf.nn.max_pool(conv4, [1, 3, 3, 1], [1, 2, 2, 1],
padding='VALID',
name=scope)
net = pool2
for i in range(3):
net = inception_module_origin(net,
scope="inception_origin" +
str(i + 1))
# 61 * 61 * 288
net = inception_reduction_v1(net, scope="reduction_v1")
# 30 * 30 * 768
for i in range(4):
net = inception_module_factorization(
net, scope="inception_factorization" + str(i))
# Auxiliary Head logits, input 30 * 30 * 768
with tf.variable_scope("aux_logits"):
aux_logits = tf.identity(net)
aux_logits = tf.nn.avg_pool(aux_logits, [1, 5, 5, 1], [1, 3, 3, 1],
padding='VALID')
# 9 * 9 * 768
aux_logits = conv2d(aux_logits, 128, [1, 1], scope='proj')
shape = aux_logits.get_shape()
aux_logits = conv2d(aux_logits,
768,
shape[1:3],
stddev=0.01,
padding='VALID')
dims = aux_logits.get_shape()[1:]
k = dims.num_elements()
aux_logits = tf.reshape(aux_logits, [-1, k])
aux_logits = fc(aux_logits,
num_classes,
activation=None,
stddev=0.001)
net = inception_reduction_v2(net, scope="reduction_v2")
# 14 * 14 * 1280
for i in range(2):
net = inception_module_expanded(net,
scope="inception_expanded" +
str(i))
with tf.variable_scope("logits"):
shape = net.get_shape()
net = tf.nn.avg_pool(net, [1, shape[1], shape[2], 1], [1, 1, 1, 1],
padding='VALID')
net = tf.nn.dropout(net, dropout_keep_prob, name="dropout")
dims = net.get_shape()[1:].num_elements()
net = tf.reshape(net, [-1, dims])
logits = fc(net, num_classes, activation=None, scope="logits")
return logits, aux_logits