def yolo_convolutional(inputs, filters, trainable, data_format, name):
with tf.variable_scope(name):
inputs = convolutional(inputs=inputs,
filters=filters,
kernel_size=1,
trainable=trainable,
name='conv0',
data_format=data_format)
inputs = convolutional(inputs=inputs,
filters=2 * filters,
kernel_size=3,
trainable=trainable,
name='conv1',
data_format=data_format)
inputs = convolutional(inputs=inputs,
filters=filters,
kernel_size=1,
trainable=trainable,
name='conv2',
data_format=data_format)
inputs = convolutional(inputs=inputs,
filters=2 * filters,
kernel_size=3,
trainable=trainable,
name='conv3',
data_format=data_format)
inputs = convolutional(inputs=inputs,
filters=filters,
kernel_size=1,
trainable=trainable,
name='conv4',
data_format=data_format)
route = inputs
inputs = convolutional(inputs=inputs,
filters=2 * filters,
kernel_size=3,
trainable=trainable,
name='conv5',
data_format=data_format)
return route, inputs
def __build(self, inputs):
with tf.variable_scope('yolo_v3_model'):
if self.data_format == 'channels_first':
inputs = tf.transpose(inputs, [0, 3, 1, 2])
# mark this step
#inputs = inputs / 255
route1, route2, inputs = darknet53(inputs=inputs, trainable=self.trainable,
data_format=self.data_format)
route, inputs = yolo_convolutional(inputs=inputs, filters=512,
trainable=self.trainable, data_format=self.data_format, name='yolo_conv0')
conv_lbbox, pred_lbbox, xy_offset_l = yolo_detection(inputs=inputs,
n_classes=self.n_classes,
anchors=_ANCHORS[6:9],
img_size=self.model_size, trainable=self.trainable,
data_format=self.data_format, name='conv_lbbox')
inputs = convolutional(inputs=route, filters=256, kernel_size=1,
trainable=self.trainable, name='conv57',
data_format=self.data_format)
upsample_size = route2.get_shape().as_list()
inputs = upsample(inputs=inputs, out_shape=upsample_size,
data_format=self.data_format, name='upsample0')
if self.data_format=='channels_first':
axis = 1
else:
axis = 3
with tf.variable_scope('route_1'):
inputs = tf.concat([inputs, route2], axis=axis)
route, inputs = yolo_convolutional(inputs=inputs, filters=256,
trainable=self.trainable, data_format=self.data_format, name='yolo_conv1')
conv_mbbox, pred_mbbox, xy_offset_m = yolo_detection(inputs=inputs,
n_classes=self.n_classes,
anchors=_ANCHORS[3:6],
img_size=self.model_size, trainable=self.trainable,
data_format=self.data_format, name='conv_mbbox')
inputs = convolutional(inputs=route, filters=128, kernel_size=1,
trainable=self.trainable, name='conv63',
data_format=self.data_format)
upsample_size = route1.get_shape().as_list()
inputs = upsample(inputs, out_shape=upsample_size,
data_format=self.data_format, name='upsample1')
with tf.variable_scope('route_2'):
inputs = tf.concat([inputs, route1], axis=axis)
route, inputs = yolo_convolutional(inputs=inputs, filters=128,
trainable=self.trainable, data_format=self.data_format, name='yolo_conv2')
conv_sbbox, pred_sbbox, xy_offset_s = yolo_detection(inputs=inputs,
n_classes=self.n_classes,
anchors=_ANCHORS[0:3],
img_size=self.model_size, trainable=self.trainable,
data_format=self.data_format, name='conv_sbbox')
return [conv_lbbox, conv_mbbox, conv_sbbox],\
[pred_lbbox, pred_mbbox, pred_sbbox],\
[xy_offset_l, xy_offset_m, xy_offset_s]
def darknet53(inputs, trainable, data_format):
with tf.variable_scope('darknet'):
inputs = convolutional(inputs=inputs,
filters=32,
kernel_size=3,
trainable=trainable,
name='conv0',
data_format=data_format)
inputs = convolutional(inputs=inputs,
filters=64,
kernel_size=3,
trainable=trainable,
name='conv1',
strides=2,
data_format=data_format)
for i in range(1):
inputs = residual(inputs=inputs,
filters=32,
trainable=trainable,
data_format=data_format,
name='residual%d' % (i + 0))
inputs = convolutional(inputs=inputs,
filters=128,
kernel_size=3,
trainable=trainable,
name='conv4',
strides=2,
data_format=data_format)
for i in range(2):
inputs = residual(inputs=inputs,
filters=64,
trainable=trainable,
data_format=data_format,
name='residual%d' % (i + 1))
inputs = convolutional(inputs=inputs,
filters=256,
kernel_size=3,
trainable=trainable,
name='conv9',
strides=2,
data_format=data_format)
for i in range(8):
inputs = residual(inputs=inputs,
filters=128,
trainable=trainable,
data_format=data_format,
name='residual%d' % (i + 3))
route1 = inputs
inputs = convolutional(inputs=inputs,
filters=512,
kernel_size=3,
trainable=trainable,
name='conv26',
strides=2,
data_format=data_format)
for i in range(8):
inputs = residual(inputs=inputs,
filters=256,
trainable=trainable,
data_format=data_format,
name='residual%d' % (i + 11))
route2 = inputs
inputs = convolutional(inputs=inputs,
filters=1024,
kernel_size=3,
trainable=trainable,
name='conv43',
strides=2,
data_format=data_format)
for i in range(4):
inputs = residual(inputs=inputs,
filters=512,
trainable=trainable,
data_format=data_format,
name='residual%d' % (i + 19))
return route1, route2, inputs
n_channel = 1
n_filter = 32
HF = 5
HW = 5
n_input = 2
n_hidden = 100
n_output = 10
X = graph.Placeholder(name='inputs') #to feed with attributes
Y = graph.Placeholder(name='labels') #to feed with labels
#convolutional(X, channel_in, channel_out, filter_height = 3, filter_width = 3, stride = 1, pad = 1):
conv_1 = layers.convolutional(X,
1,
16,
filter_height=3,
filter_width=3,
stride=1,
pad=1)
conv_2 = layers.convolutional(conv_1,
16,
32,
filter_height=3,
filter_width=3,
stride=3,
pad=1)
flat = op.Flatten(conv_2)
fc_1 = layers.fully_connected(flat,
10 * 10 * 32,
n_hidden,
activation='sigmoid')
def yolo_detection(inputs, n_classes, anchors, img_size, trainable,
data_format, name):
'''
Args:
inputs: tensor input
n_classes: number of labels
anchors: a list of anchor sizes
img_size: the input size of the model
data_format: input format
'''
n_anchors = len(anchors)
filters = n_anchors * (5 + n_classes)
inputs = convolutional(inputs=inputs,
filters=filters,
kernel_size=1,
trainable=trainable,
name=name,
data_format=data_format,
act=False,
bn=False)
# raw output of detection conv layer
raw_output = inputs
shape = inputs.get_shape().as_list()
# channels_first: NCHW
# channels_last: NHWC
if data_format == 'channels_first':
grid_shape = shape[2:4]
# reshape to NHWC
inputs = tf.transpose(inputs, [0, 2, 3, 1])
raw_output = tf.transpose(raw_output, [0, 2, 3, 1])
else:
grid_shape = shape[1:3]
inputs = tf.reshape(
inputs, [-1, n_anchors * grid_shape[0] * grid_shape[1], 5 + n_classes])
strides = (img_size[0] // grid_shape[0], img_size[1] // grid_shape[1])
# split & get the 4 components of output
box_centers, box_shapes, confidence, classes = \
tf.split(inputs, [2, 2, 1, n_classes], axis=-1)
x = tf.range(grid_shape[0], dtype=tf.float32)
y = tf.range(grid_shape[1], dtype=tf.float32)
x_offset, y_offset = tf.meshgrid(x, y)
x_offset = tf.reshape(x_offset, (-1, 1))
y_offset = tf.reshape(y_offset, (-1, 1))
x_y_offset = tf.concat([x_offset, y_offset], axis=-1)
xy_offset_output = tf.identity(x_y_offset)
xy_offset_output = tf.reshape(xy_offset_output,
[grid_shape[0], grid_shape[1], 1, 2])
x_y_offset = tf.tile(x_y_offset, [1, n_anchors])
x_y_offset = tf.reshape(x_y_offset, [1, -1, 2])
box_centers = tf.nn.sigmoid(box_centers)
box_centers = (box_centers + x_y_offset) * strides
anchors = tf.tile(anchors, [grid_shape[0] * grid_shape[1], 1])
box_shapes = tf.exp(box_shapes) * tf.cast(anchors, tf.float32)
confidence = tf.nn.sigmoid(confidence)
classes = tf.nn.sigmoid(classes)
inputs = tf.concat([box_centers, box_shapes, confidence, classes], axis=-1)
inputs = tf.reshape(
inputs, [-1, grid_shape[0], grid_shape[1], n_anchors, 5 + n_classes])
return raw_output, inputs, xy_offset_output