def _block(self, inputs, filter, layer_name):
with tf.name_scope(layer_name):
inputs = common._conv2d_fixed_padding(inputs, filter[0], 1, 1,
layer_name + '_conv1')
inputs = common._conv2d_fixed_padding(inputs, filter[1], 3, 1,
layer_name + '_conv2')
inputs = common._conv2d_fixed_padding(inputs, 64, 3, 2,
layer_name + '_conv3')
return inputs
def _extract_block(self, inputs, layer_name):
with tf.name_scope(layer_name):
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=32,
kernel_size=3,
strides=1,
layer_name=layer_name +
'_conv1')
for i in range(2):
name = layer_name + 'res_block1' + str(i)
inputs = self._darknet53_block(inputs, 16, name)
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=128,
kernel_size=3,
strides=2,
layer_name=layer_name +
'_conv2')
for i in range(4):
name = layer_name + 'res_block2' + str(i)
inputs = self._darknet53_block(inputs, 64, name)
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=512,
kernel_size=3,
strides=2,
layer_name=layer_name +
'_conv3')
for i in range(4):
name = layer_name + 'res_block3' + str(i)
inputs = self._darknet53_block(inputs, 256, name)
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=128,
kernel_size=3,
strides=2,
layer_name=layer_name +
'_conv4')
for i in range(2):
name = layer_name + 'res_block4' + str(i)
inputs = self._darknet53_block(inputs, 64, name)
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=4,
kernel_size=3,
strides=2,
layer_name=layer_name +
'_conv5')
inputs = tf.reshape(inputs, (-1, 1, 1, 64))
inputs = tf.tile(inputs, multiples=[1, 16, 16, 1])
return inputs
def _darknet53_block(self, inputs, filters, name=None):
"""
implement residuals block in darknet53
"""
shortcut = inputs
inputs = common._conv2d_fixed_padding(inputs, filters * 1, 1, 1,
name + '_conv1')
inputs = common._conv2d_fixed_padding(inputs, filters * 2, 3, 1,
name + '_conv2')
inputs = inputs + shortcut
return inputs
def transform_layer(self, inputs, scope):
with tf.name_scope(scope):
inputs = common._conv2d_fixed_padding(inputs,
DEEPTH,
1,
strides=1,
layer_name="trans_1")
inputs = common._conv2d_fixed_padding(inputs,
DEEPTH,
3,
strides=1,
layer_name="trans_2")
return inputs
def transition_layer(self, inputs, filters, scope):
with tf.name_scope(scope):
inputs = common._conv2d_fixed_padding(inputs,
filters,
1,
strides=1,
layer_name="trans_3")
return inputs
def _yolo_block(self, inputs: tf.Tensor, filters: int, layer_name: str):
"""
:param inputs:
:param filters:
:param layer_name:
:return:
"""
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=filters * 1,
kernel_size=1,
layer_name=layer_name + "_a")
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=filters * 2,
kernel_size=3,
layer_name=layer_name + "_b")
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=filters * 1,
kernel_size=1,
layer_name=layer_name + "_c")
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=filters * 2,
kernel_size=3,
layer_name=layer_name + "_d")
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=filters * 1,
kernel_size=1,
layer_name=layer_name + "_e")
route = inputs
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=filters * 2,
kernel_size=3,
layer_name=layer_name + "_f")
return route, inputs
def forward(
self,
room_zid: tf.Tensor,
room_cid: tf.Tensor,
room_mid: tf.Tensor,
room_scid: tf.Tensor,
room_distance: tf.Tensor,
zone_zid: tf.Tensor,
zone_cid: tf.Tensor,
zone_mid: tf.Tensor,
):
"""
Creates YOLO v3 model.
"""
batch_norm_params = {
'decay': self._BATCH_NORM_DECAY,
'epsilon': 1e-05,
'scale': True,
'is_training': self.trainable,
'fused': None,
}
with slim.arg_scope(
[slim.conv2d, slim.batch_norm, common._fixed_padding]):
with slim.arg_scope(
[slim.conv2d],
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params,
weights_regularizer=slim.l2_regularizer(0.0005),
biases_initializer=None,
activation_fn=lambda x: tf.nn.leaky_relu(
x, alpha=self._LEAKY_RELU)):
with tf.variable_scope('darknet-53'):
route_1, route_2, inputs = darknet53(
room_zid=room_zid,
room_cid=room_cid,
room_mid=room_mid,
room_scid=room_scid,
zone_zid=zone_zid,
zone_cid=zone_cid,
zone_mid=zone_mid,
room_distance=room_distance,
furniture_nums=self.furniture_nums,
furniture_embedding_size=self.furniture_embedding_size
).outputs
print("yolo:{0}".format(inputs))
with tf.variable_scope('yolo-v3'):
route, inputs = self._yolo_block(inputs=inputs,
filters=512,
layer_name="route1")
inputs = common._conv2d_fixed_padding(
inputs, 512, 1, layer_name="route1_conv")
upsample_size = route_2.get_shape().as_list()
inputs = self._upsample(inputs, upsample_size)
inputs = tf.concat([inputs, route_2], axis=3)
route, inputs = self._yolo_block(inputs=inputs,
filters=256,
layer_name="route2")
inputs = common._conv2d_fixed_padding(
inputs, 64, 1, layer_name="route2_conv")
upsample_size = route_1.get_shape().as_list()
inputs = self._upsample(inputs, upsample_size)
inputs = tf.concat([inputs, route_1],
axis=3,
name='concat_2')
route, inputs = self._yolo_block(inputs=inputs,
filters=4,
layer_name="route3")
feature_map_3 = tf.identity(route, name='feature_map_3')
return feature_map_3
def forward(self,
room_zid: tf.Tensor,
room_cid: tf.Tensor,
room_mid: tf.Tensor,
room_scid: tf.Tensor,
zone_zid: tf.Tensor,
zone_cid: tf.Tensor,
zone_mid: tf.Tensor,
room_distance: tf.Tensor = None):
room_embedding_list = EmbeddingLayers(
input_tensor_list=[
room_zid, room_cid, room_mid, room_scid, room_distance
],
vocab_size_list=[200, self.furniture_nums, 200, 200, 200],
num_unit_list=[self.furniture_embedding_size] * 6,
scopes=[
"zid_table", "cid_table", "mid_table", "scid_table",
"distance_table"
],
trainable=True,
dtype=tf.float32)
room_embedding = tf.concat(values=room_embedding_list, axis=-1)
zone_embedding_list = EmbeddingLayers(
input_tensor_list=[zone_zid, zone_cid, zone_mid],
vocab_size_list=[200, self.furniture_nums, 200],
num_unit_list=[self.furniture_embedding_size] * 3,
scopes=["zid_table", "cid_table", "scid_table"],
trainable=True,
dtype=tf.float32)
zone_embedding = tf.concat(values=zone_embedding_list, axis=-1)
zone_feature = self._extract_block(zone_embedding, 'zone_extract')
room_feature = self._block(room_embedding, [16, 32], 'room')
for i in range(2):
name = 'res_block1' + str(i)
room_feature = self._darknet53_block(room_feature, 32, name)
room_feature = common._conv2d_fixed_padding(inputs=room_feature,
filters=64,
kernel_size=3,
strides=2,
layer_name="forward_a")
inputs = tf.concat([zone_feature, room_feature],
axis=3,
name='concat_zone_room')
for i in range(8):
name = 'res_block2' + str(i)
inputs = self._darknet53_block(inputs, 64, name)
route_1 = inputs
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=512,
kernel_size=3,
strides=2,
layer_name="forward_b")
for i in range(8):
name = 'res_block3' + str(i)
inputs = self._darknet53_block(inputs, 256, name)
route_2 = inputs
inputs = common._conv2d_fixed_padding(inputs=inputs,
filters=1024,
kernel_size=3,
strides=2,
layer_name="forward_c")
for i in range(4):
name = 'res_block4' + str(i)
inputs = self._darknet53_block(inputs, 512, name)
return route_1, route_2, inputs