def BasicBlock(self,
inputs,
filters,
strides=1,
is_training=False,
use_bn=True):
expansion = 1
conv1_bn_relu = _conv(inputs,
filters, [3, 3],
strides,
'same',
activation=tf.nn.relu,
is_training=self.is_training,
use_bn=self.use_bn)
conv2_bn = _conv(conv1_bn_relu,
filters, [3, 3],
1,
'same',
activation=None,
is_training=self.is_training,
use_bn=self.use_bn)
if strides != 1 or self.inplanes != filters * expansion:
inputs = _conv(inputs,
filters, [1, 1],
strides,
'valid',
activation=None,
is_training=self.is_training,
use_bn=self.use_bn)
self.inplanes = filters * expansion
out = tf.nn.relu(conv2_bn + inputs)
return out
def _build_model(self, inputs):
with tf.variable_scope('mobilenet'):
# mobilenet v2
# c2, c3, c4, c5 = mobilenet.MobileNetV2(inputs=inputs, is_training=self.is_training).forward()
# mobilenet v3
c2, c3, c4, c5, cls = mobilenet_v3.mobilenet_v3_small(inputs=inputs, is_training=self.is_training)
p5 = _conv(c5, 24, [1, 1], is_training=self.is_training)
up_p5 = upsampling(p5, method='resize')
reduce_dim_c4 = _conv(c4, 24, [1, 1], is_training=self.is_training)
p4 = 0.5 * up_p5 + 0.5 * reduce_dim_c4
up_p4 = upsampling(p4, method='resize')
reduce_dim_c3 = _conv(c3, 24, [1, 1], is_training=self.is_training)
p3 = 0.5 * up_p4 + 0.5 * reduce_dim_c3
up_p3 = upsampling(p3, method='resize')
reduce_dim_c2 = _conv(c2, 24, [1, 1], is_training=self.is_training)
p2 = 0.5 * up_p3 + 0.5 * reduce_dim_c2
features = _conv(p2, 24, [3, 3], is_training=self.is_training)
with tf.variable_scope('detector'):
hm = _conv(features, 24, [3, 3], is_training=self.is_training)
hm = tf.layers.conv2d(hm, cfgs.NUM_CLASS, 1, 1, padding='valid', activation=tf.nn.sigmoid,
bias_initializer=tf.constant_initializer(-np.log(99.)), name='hm')
wh = _conv(features, 24, [3, 3], is_training=self.is_training)
wh = tf.layers.conv2d(wh, 2, 1, 1, padding='valid', activation=None, name='wh')
reg = _conv(features, 24, [3, 3], is_training=self.is_training)
reg = tf.layers.conv2d(reg, 2, 1, 1, padding='valid', activation=None, name='reg')
return hm, wh, reg, cls
def _build_model(self, inputs):
with tf.variable_scope('resnet'):
c2, c3, c4, c5 = resnet.resnet18(is_training=self.is_training).forward(inputs)
p5 = _conv(c5, 128, [1,1], is_training=self.is_training, name="conv_p5")
up_p5 = upsampling(p5)
reduce_dim_c4 = _conv(c4, 128, [1,1], is_training=self.is_training, name="conv_c4")
p4 = 0.5*up_p5 + 0.5*reduce_dim_c4
up_p4 = upsampling(p4)
reduce_dim_c3 = _conv(c3, 128, [1,1], is_training=self.is_training, name="conv_c3")
p3 = 0.5*up_p4 + 0.5*reduce_dim_c3
up_p3 = upsampling(p3)
reduce_dim_c2 = _conv(c2, 128, [1,1], is_training=self.is_training, name="conv_c2")
p2 = 0.5*up_p3 + 0.5*reduce_dim_c2
features = _conv(p2, 128, [3,3], is_training=self.is_training, name="conv_p2")
# IDA-up
# p2 = _conv(c2, 128, [1,1], is_training=self.is_training)
# p3 = _conv(c3, 128, [1,1], is_training=self.is_training)
# p4 = _conv(c4, 128, [1,1], is_training=self.is_training)
# p5 = _conv(c5, 128, [1,1], is_training=self.is_training)
# up_p3 = upsampling(p3, method='resize')
# p2 = _conv(p2+up_p3, 128, [3,3], is_training=self.is_training)
# up_p4 = upsampling(upsampling(p4, method='resize'), method='resize')
# p2 = _conv(p2+up_p4, 128, [3,3], is_training=self.is_training)
# up_p5 = upsampling(upsampling(upsampling(p5, method='resize'), method='resize'), method='resize')
# features = _conv(p2+up_p5, 128, [3,3], is_training=self.is_training)
with tf.variable_scope('detector'):
print('feature shape: ', features.shape)
hm = _conv(features, 64, [3,3], is_training=self.is_training, name="hm_conv_1")
hm = _conv_nn(hm, cfg.num_classes, [1, 1], padding='VALID', activation = tf.nn.relu, name='hm_conv')
#hm = _conv(hm, cfg.num_classes, [1, 1], padding="valid", name="hm")
wh = _conv(features, 64, [3,3], is_training=self.is_training, name="wh_conv_1")
#wh = tf.layers.conv2d(wh, 2, 1, 1, padding='valid', activation = None, bias_initializer=tf.constant_initializer(-np.log(99.)), name='wh')
wh = _conv_nn(wh, 2, [1, 1], padding='VALID', activation = tf.nn.relu, name="wh_conv")
#wh = tf.reshape(wh, [-1, wh.shape[1], wh.shape[2], wh.shape[3]])
#wh = tf.layers.conv2d(wh, 2, 1, 1, padding='valid', activation = None, name='wh')
#wh = _conv(wh, 2, [1, 1], padding="valid", name="wh")
reg = _conv(features, 64, [3,3], is_training=self.is_training, name="reg_conv_1")
#reg = tf.layers.conv2d(reg, 2, 1, 1, padding='valid', activation = None, bias_initializer=tf.constant_initializer(-np.log(99.)), name='reg')
reg = _conv_nn(reg, 2, [1, 1], padding='VALID', activation = tf.nn.relu, name="reg_conv")
#reg = tf.reshape(reg, [-1, reg.shape[1], reg.shape[2], reg.shape[3]])
#reg = _conv(reg, 2, [1, 1], padding="valid", name="reg")
return hm, wh, reg
def _layer0(self, inputs, filters, kernel_size=(7, 7)):
outputs = _conv(inputs,
filters, [7, 7],
2,
'same',
activation=tf.nn.relu,
is_training=self.is_training,
use_bn=self.use_bn)
outputs = tf.layers.max_pooling2d(outputs,
pool_size=3,
strides=2,
padding='same')
return outputs
def Bottleneck(self,
inputs,
filters,
strides=1,
is_training=False,
use_bn=True):
expansion = 4
conv1_bn_relu = _conv(inputs,
filters, [1, 1],
1,
'VALID',
activation=tf.nn.relu,
is_training=self.is_training,
use_bn=self.use_bn)
conv2_bn_relu = _conv(conv1_bn_relu,
filters, [3, 3],
strides,
'SAME',
activation=tf.nn.relu,
is_training=self.is_training,
use_bn=self.use_bn)
conv3_bn = _conv(conv2_bn_relu,
filters * expansion, [1, 1],
1,
'VALID',
activation=None,
is_training=self.is_training,
use_bn=self.use_bn)
if strides != 1 or self.inplanes != filters * expansion:
inputs = _conv(inputs,
filters * expansion, [1, 1],
strides,
'VALID',
activation=None,
is_training=self.is_training,
use_bn=self.use_bn)
self.inplanes = filters * expansion
out = tf.nn.relu(conv3_bn + inputs)
return out
def BasicBlock(self,
inputs,
filters,
strides=1,
is_training=False,
use_bn=True):
expansion = 1
conv1_bn_relu = _conv(inputs,
filters, [3, 3],
strides,
'SAME',
activation=tf.nn.relu,
is_training=self.is_training,
use_bn=self.use_bn,
name="conv" + str(self.counter))
self.counter += 1
conv2_bn = _conv(conv1_bn_relu,
filters, [3, 3],
1,
'SAME',
activation=None,
is_training=self.is_training,
use_bn=self.use_bn,
name="conv" + str(self.counter))
self.counter += 1
if strides != 1 or self.inplanes != filters * expansion:
inputs = _conv(inputs,
filters, [1, 1],
strides,
'VALID',
activation=None,
is_training=self.is_training,
use_bn=self.use_bn,
name="conv" + str(self.counter))
self.counter += 1
self.inplanes = filters * expansion
out = tf.nn.relu(conv2_bn + inputs)
return out
def _build_model(self, inputs):
with tf.variable_scope('resnet'):
c2, c3, c4, c5 = resnet.resnet34(is_training=self.is_training).forward(inputs)
p5 = _conv(c5, 128, [1,1], is_training=self.is_training)
up_p5 = upsampling(p5, method='resize')
reduce_dim_c4 = _conv(c4, 128, [1,1], is_training=self.is_training)
p4 = 0.5*up_p5 + 0.5*reduce_dim_c4
up_p4 = upsampling(p4, method='resize')
reduce_dim_c3 = _conv(c3, 128, [1,1], is_training=self.is_training)
p3 = 0.5*up_p4 + 0.5*reduce_dim_c3
up_p3 = upsampling(p3, method='resize')
reduce_dim_c2 = _conv(c2, 128, [1,1], is_training=self.is_training)
p2 = 0.5*up_p3 + 0.5*reduce_dim_c2
features = _conv(p2, 128, [3,3], is_training=self.is_training)
# IDA-up
# p2 = _conv(c2, 128, [1,1], is_training=self.is_training)
# p3 = _conv(c3, 128, [1,1], is_training=self.is_training)
# p4 = _conv(c4, 128, [1,1], is_training=self.is_training)
# p5 = _conv(c5, 128, [1,1], is_training=self.is_training)
# up_p3 = upsampling(p3, method='resize')
# p2 = _conv(p2+up_p3, 128, [3,3], is_training=self.is_training)
# up_p4 = upsampling(upsampling(p4, method='resize'), method='resize')
# p2 = _conv(p2+up_p4, 128, [3,3], is_training=self.is_training)
# up_p5 = upsampling(upsampling(upsampling(p5, method='resize'), method='resize'), method='resize')
# features = _conv(p2+up_p5, 128, [3,3], is_training=self.is_training)
with tf.variable_scope('detector'):
hm = _conv(features, 64, [3,3], is_training=self.is_training)
hm = tf.layers.conv2d(hm, cfg.num_classes, 1, 1, padding='valid', activation = tf.nn.sigmoid, bias_initializer=tf.constant_initializer(-np.log(99.)), name='hm')
wh = _conv(features, 64, [3,3], is_training=self.is_training)
wh = tf.layers.conv2d(wh, 2, 1, 1, padding='valid', activation = None, name='wh')
reg = _conv(features, 64, [3,3], is_training=self.is_training)
reg = tf.layers.conv2d(reg, 2, 1, 1, padding='valid', activation = None, name='reg')
return hm, wh, reg
def _build_model(self, inputs):
with tf.variable_scope('dla'):
assert self.net_name in ["dla_34"]
if self.net_name == "dla_34":
features = dla_net.DLAnet(
is_training=self.is_training).forward(inputs)
with tf.variable_scope('detector'):
hm = _conv(features, 64, [3, 3], is_training=self.is_training)
hm = tf.layers.conv2d(
hm,
cfg.num_classes,
1,
1,
padding='valid',
activation=tf.nn.sigmoid,
bias_initializer=tf.constant_initializer(-np.log(99.)),
name='hm')
wh = _conv(features, 64, [3, 3], is_training=self.is_training)
wh = tf.layers.conv2d(wh,
2,
1,
1,
padding='valid',
activation=None,
name='wh')
reg = _conv(features, 64, [3, 3], is_training=self.is_training)
reg = tf.layers.conv2d(reg,
2,
1,
1,
padding='valid',
activation=None,
name='reg')
return hm, wh, reg
def _layer0(self, inputs, filters, kernel_size=(7, 7)):
outputs = _conv(inputs,
filters, [7, 7],
2,
'SAME',
activation=tf.nn.relu,
is_training=self.is_training,
use_bn=self.use_bn,
name="conv" + str(self.counter))
self.counter += 1
outputs = tf.layers.max_pooling2d(outputs,
pool_size=3,
strides=2,
padding='SAME')
print('Counter: ', self.counter)
self.first_layer = outputs
return outputs