def __call__(self, inputs):
"""Generate Computation Graph"""
with tf.variable_scope(self.scope):
if self.data_format == 'channels_first':
inputs = tf.transpose(inputs, [0, 3, 1, 2])
inputs = inputs / 255
inputs, route2, route4 = darknet53(inputs, data_format=self.data_format)
inputs, route1 = feature_pyramid_network(inputs, filters=512, data_format=self.data_format)
detect1 = yolo_layer(inputs,
n_classes=self.n_classes,
anchors=_ANCHORS[6:],
img_size=self.input_size,
data_format=self.data_format)
inputs = Conv2D(route1, filters=256, kernel_size=1, data_format=self.data_format)
inputs = BatchNormalization(inputs, data_format=self.data_format)
inputs = LeakyReLU(inputs)
upsample_size = route2.get_shape().as_list()
inputs = upsample(inputs, out_shape=upsample_size, data_format=self.data_format)
axis = 1 if self.data_format == 'channels_first' else 3
inputs = tf.concat([inputs, route2], axis=axis)
inputs, route3 = feature_pyramid_network(inputs, filters=256, data_format=self.data_format)
detect2 = yolo_layer(inputs,
n_classes=self.n_classes,
anchors=_ANCHORS[3:6],
img_size=self.input_size,
data_format=self.data_format)
inputs = Conv2D(route3, filters=128, kernel_size=1, data_format=self.data_format)
inputs = BatchNormalization(inputs, data_format=self.data_format)
inputs = LeakyReLU(inputs)
upsample_size = route4.get_shape().as_list()
inputs = upsample(inputs, out_shape=upsample_size, data_format=self.data_format)
axis = 1 if self.data_format == 'channels_first' else 3
inputs = tf.concat([inputs, route4], axis=axis)
inputs, _ = feature_pyramid_network(inputs, filters=128, data_format=self.data_format)
detect3 = yolo_layer(inputs,
n_classes=self.n_classes,
anchors=_ANCHORS[:3],
img_size=self.input_size,
data_format=self.data_format)
inputs = tf.concat([detect1, detect2, detect3], axis=1)
inputs = build_boxes(inputs)
boxes_dicts = non_max_suppression(inputs,
n_classes=self.n_classes,
max_output_size=self.max_output_size,
iou_threshold=self.iou_threshold,
confidence_threshold=self.confidence_threshold)
return boxes_dicts
def _pnet(styles, data_format="channels_last"):
with tf.variable_scope("pnet", reuse=tf.AUTO_REUSE):
pnet_outs = {}
inputs = ((styles / 255.) - 0.5) * 2
pnet_outs["conv1"] = \
conv1 = conv(inputs, name="conv1", data_format=data_format, filters=32, kernel_size=9, strides=1)
pnet_outs["conv2"] = \
conv2 = conv(conv1, name="conv2", data_format=data_format, filters=64, kernel_size=3, strides=2)
pnet_outs["conv3"] = \
conv3 = conv(conv2, name="conv3", data_format=data_format, filters=128, kernel_size=3, strides=2)
res1 = residual_block(conv3,
name="res1",
data_format=data_format,
filters=128,
kernel_size=3)
res2 = residual_block(res1,
name="res2",
data_format=data_format,
filters=128,
kernel_size=3)
res3 = residual_block(res2,
name="res3",
data_format=data_format,
filters=128,
kernel_size=3)
res4 = residual_block(res3,
name="res4",
data_format=data_format,
filters=128,
kernel_size=3)
res5 = residual_block(res4,
name="res5",
data_format=data_format,
filters=128,
kernel_size=3)
pnet_outs["up1"] \
= up1 = upsample(res5, name="up1", data_format=data_format, filters=64, kernel_size=3, strides=2)
pnet_outs["up2"] \
= _ = upsample(up1, name="up2", data_format=data_format, filters=32, kernel_size=3, strides=2)
gammas, betas = _pnet_fc(pnet_outs)
return gammas, betas
def model_fn(inputs):
with tf.variable_scope("tnet", reuse=tf.AUTO_REUSE):
inputs = ((inputs / 255.) - 0.5) * 2
conv1 = conv(inputs,
name="conv1",
data_format=data_format,
filters=32,
kernel_size=9,
strides=1)
conv2 = conv(conv1,
name="conv2",
data_format=data_format,
filters=64,
kernel_size=3,
strides=2)
conv3 = conv(conv2,
name="conv3",
data_format=data_format,
filters=128,
kernel_size=3,
strides=2)
res1 = residual_block(conv3,
name="res1",
data_format=data_format,
filters=128,
kernel_size=3)
res2 = residual_block(res1,
name="res2",
data_format=data_format,
filters=128,
kernel_size=3)
res3 = residual_block(res2,
name="res3",
data_format=data_format,
filters=128,
kernel_size=3)
res4 = residual_block(res3,
name="res4",
data_format=data_format,
filters=128,
kernel_size=3)
res5 = residual_block(res4,
name="res5",
data_format=data_format,
filters=128,
kernel_size=3)
up1 = upsample(res5,
name="up1",
data_format=data_format,
filters=64,
kernel_size=3,
strides=2)
up2 = upsample(up1,
name="up2",
data_format=data_format,
filters=32,
kernel_size=3,
strides=2)
conv4 = conv(up2,
name="conv4",
data_format=data_format,
filters=3,
kernel_size=9,
strides=1,
with_bn=False,
with_relu=False)
out = tf.clip_by_value(conv4, 0., 255.)
return out
def build_nework(self, input_data):
route_1, route_2, input_data = backbone.DarkNet53(
input_data, self.trainable)
# 对darknet的最后一层输出继续做两次33卷积，并生成大分辨率的预测输出
input_data = layers.convolution_layer(input_data, (1, 1, 1024, 512),
self.trainable, 'conv52')
input_data = layers.convolution_layer(input_data, (3, 3, 512, 1024),
self.trainable, 'conv53')
input_data = layers.convolution_layer(input_data, (1, 1, 1024, 512),
self.trainable, 'conv54')
input_data = layers.convolution_layer(input_data, (3, 3, 512, 1024),
self.trainable, 'conv55')
input_data = layers.convolution_layer(input_data, (1, 1, 1024, 512),
self.trainable, 'conv56')
conv_lobj_branch = layers.convolution_layer(input_data,
(3, 3, 512, 1024),
self.trainable,
'conv_lobj_branch')
#输出大分辨率的输出 (13,13,256)
conv_lbbox = layers.convolution_layer(
conv_lobj_branch,
(1, 1, 1024, self.anchor_per_scale * (self.num_class + 5)),
self.trainable,
'conv_lbbox',
activate=False,
bn=False)
# route1和darknet output相结合，生成中分辨率的预测输出
input_data = layers.convolution_layer(input_data, (1, 1, 512, 256),
self.trainable, 'conv57')
input_data = layers.upsample(
input_data, name='upsample0',
method=self.upsample_method) # (26,26,256)
with tf.variable_scope('route_1'):
input_data = tf.concat([input_data, route_2],
axis=-1) # (26, 26, 768)
input_data = layers.convolution_layer(input_data, (1, 1, 768, 256),
self.trainable, 'conv58')
input_data = layers.convolution_layer(input_data, (3, 3, 256, 512),
self.trainable, 'conv59')
input_data = layers.convolution_layer(input_data, (1, 1, 512, 256),
self.trainable, 'conv60')
input_data = layers.convolution_layer(input_data, (3, 3, 256, 512),
self.trainable, 'conv61')
input_data = layers.convolution_layer(input_data, (1, 1, 512, 256),
self.trainable, 'conv62')
conv_mobj_branch = layers.convolution_layer(input_data,
(3, 3, 256, 512),
self.trainable,
'conv_mobj_branch')
# 输出中分辨率的输出 (26, 26, 255)
conv_mbbox = layers.convolution_layer(
conv_mobj_branch,
(1, 1, 512, self.anchor_per_scale * (self.num_class + 5)),
self.trainable,
'conv_mbbox',
activate=False,
bn=False)
# route1和route2结合生成小分辨率的预测
input_data = layers.convolution_layer(input_data, (1, 1, 256, 128),
self.trainable, 'conv63')
input_data = layers.upsample(
input_data, name='upsample1',
method=self.upsample_method) # (52, 52, 128)
with tf.variable_scope('route_2'):
input_data = tf.concat([input_data, route_1],
axis=-1) # (52, 52, 384)
input_data = layers.convolution_layer(input_data, (1, 1, 384, 128),
self.trainable, 'conv64')
input_data = layers.convolution_layer(input_data, (3, 3, 128, 256),
self.trainable, 'conv65')
input_data = layers.convolution_layer(input_data, (1, 1, 256, 128),
self.trainable, 'conv66')
input_data = layers.convolution_layer(input_data, (3, 3, 128, 256),
self.trainable, 'conv67')
input_data = layers.convolution_layer(input_data, (1, 1, 256, 128),
self.trainable, 'conv68')
conv_sobj_branch = layers.convolution_layer(input_data,
(3, 3, 128, 256),
self.trainable,
'conv_sobj_branch')
# 输出小分辨率的输出 (52, 52, 255)
conv_sbbox = layers.convolution_layer(
conv_sobj_branch,
(1, 1, 256, self.anchor_per_scale * (self.num_class + 5)),
self.trainable,
'conv_sbbox',
activate=False,
bn=False)
return conv_lbbox, conv_mbbox, conv_sbbox
def _tnet(contents, gammas, betas, data_format="channels_last"):
with tf.variable_scope("tnet", reuse=tf.AUTO_REUSE):
inputs = ((contents / 255.) - 0.5) * 2
conv1 = conv(inputs,
name="conv1",
data_format=data_format,
filters=32,
kernel_size=9,
strides=1,
bn_gamma=gammas["conv1"],
bn_beta=betas["conv1"])
conv2 = conv(conv1,
name="conv2",
data_format=data_format,
filters=64,
kernel_size=3,
strides=2,
bn_gamma=gammas["conv2"],
bn_beta=betas["conv2"])
conv3 = conv(conv2,
name="conv3",
data_format=data_format,
filters=128,
kernel_size=3,
strides=2,
bn_gamma=gammas["conv3"],
bn_beta=betas["conv3"])
res1 = residual_block(conv3,
name="res1",
data_format=data_format,
filters=128,
kernel_size=3,
bn_gamma1=gammas["res1_1"],
bn_beta1=betas["res1_1"],
bn_gamma2=gammas["res1_2"],
bn_beta2=betas["res1_2"])
res2 = residual_block(res1,
name="res2",
data_format=data_format,
filters=128,
kernel_size=3,
bn_gamma1=gammas["res2_1"],
bn_beta1=betas["res2_1"],
bn_gamma2=gammas["res2_2"],
bn_beta2=betas["res2_2"])
res3 = residual_block(res2,
name="res3",
data_format=data_format,
filters=128,
kernel_size=3,
bn_gamma1=gammas["res3_1"],
bn_beta1=betas["res3_1"],
bn_gamma2=gammas["res3_2"],
bn_beta2=betas["res3_2"])
res4 = residual_block(res3,
name="res4",
data_format=data_format,
filters=128,
kernel_size=3,
bn_gamma1=gammas["res4_1"],
bn_beta1=betas["res4_1"],
bn_gamma2=gammas["res4_2"],
bn_beta2=betas["res4_2"])
res5 = residual_block(res4,
name="res5",
data_format=data_format,
filters=128,
kernel_size=3,
bn_gamma1=gammas["res5_1"],
bn_beta1=betas["res5_1"],
bn_gamma2=gammas["res5_2"],
bn_beta2=betas["res5_2"])
up1 = upsample(res5,
name="up1",
data_format=data_format,
filters=64,
kernel_size=3,
strides=2,
bn_gamma=gammas["up1"],
bn_beta=betas["up1"])
up2 = upsample(up1,
name="up2",
data_format=data_format,
filters=32,
kernel_size=3,
strides=2,
bn_gamma=gammas["up2"],
bn_beta=betas["up2"])
conv4 = conv(up2,
name="conv4",
data_format=data_format,
filters=3,
kernel_size=9,
strides=1,
with_bn=False,
with_relu=False)
out = tf.clip_by_value(conv4, 0., 255.)
return out