import tensorflow as tf
from tensorflow.keras import Input
from pytwovision.models.layers.conv2d_bn_leaky_relu_layer import conv2d_bn_leaky_relu_layer
from pytwovision.models.layers.upsample_layer import UpsampleLayer
[docs]class BuildYoloV3(tf.keras.Model):
"""Build YoloV3 model given a backbone.
Args:
backbone: an object with a backbone network.
num_class: an integer with the quantity of classes.
Returns:
A list where the first one is used to predict large-sized objects,
the second one is used to predict medium-sized objects, the third one is used to small objects
and the last one is the input shape returned.
"""
def __init__(self, backbone, num_class):
super().__init__()
if not isinstance(num_class, int):
raise ValueError('num_class has to be an integer')
self.base_outputs = backbone
self.num_class = num_class
self.first_stack_filters = [(1, 1, 1024, 512), (3, 3, 512, 1024), (1, 1, 1024, 512),
(3, 3, 512, 1024), (1, 1, 1024, 512)]
self.second_stack_filters = [(1, 1, 768, 256), (3, 3, 256, 512), (1, 1, 512, 256),
(3, 3, 256, 512), (1, 1, 512, 256)]
self.third_stack_filters = [(1, 1, 384, 128), (3, 3, 128, 256), (1, 1, 256, 128),
(3, 3, 128, 256), (1, 1, 256, 128)]
[docs] def call(self, input_shape):
if len(input_shape) != 3: raise ValueError("input shape should have a len == 3")
input_shape = Input([input_shape[0], input_shape[1], input_shape[2]])
try:
route_1, route_2, x = self.base_outputs.build_model(input_shape)
except ValueError:
raise Exception('Backbone output shape mismatch with yolov3 input shape')
for n, filters in enumerate(self.first_stack_filters):
x = conv2d_bn_leaky_relu_layer(x, filters)
conv_lobj_branch = conv2d_bn_leaky_relu_layer(x, (3, 3, 512, 1024))
# conv_lbbox is used to predict large-sized objects , Shape = [None, 13, 13, 255] (if num_class == 80 like coco dataset)
conv_lbbox = conv2d_bn_leaky_relu_layer(conv_lobj_branch, (1, 1, 1024, 3*(self.num_class + 5)), activate=False, bn=False)
x = conv2d_bn_leaky_relu_layer(x, (1, 1, 512, 256))
# upsample here uses the nearest neighbor interpolation method, which has the advantage that the
# upsampling process does not need to learn, thereby reducing the network parameter
x = UpsampleLayer()(x)
x = tf.concat([x, route_2], axis=-1)
for n, filters in enumerate(self.second_stack_filters):
x = conv2d_bn_leaky_relu_layer(x, filters)
conv_mobj_branch = conv2d_bn_leaky_relu_layer(x, (3, 3, 256, 512))
# conv_mbbox is used to predict medium-sized objects, shape = [None, 26, 26, 255] (if num_class == 80 like coco dataset)
conv_mbbox = conv2d_bn_leaky_relu_layer(conv_mobj_branch, (1, 1, 512, 3*(self.num_class + 5)), activate=False, bn=False)
x = conv2d_bn_leaky_relu_layer(x, (1, 1, 256, 128))
x = UpsampleLayer()(x)
x = tf.concat([x, route_1], axis=-1)
for n, filters in enumerate(self.third_stack_filters):
x = conv2d_bn_leaky_relu_layer(x, filters)
conv_sobj_branch = conv2d_bn_leaky_relu_layer(x, (3, 3, 128, 256))
# conv_sbbox is used to predict small size objects, shape = [None, 52, 52, 255] (if num_class == 80 like coco dataset)
conv_sbbox = conv2d_bn_leaky_relu_layer(conv_sobj_branch, (1, 1, 256, 3*(self.num_class +5)), activate=False, bn=False)
return [conv_sbbox, conv_mbbox, conv_lbbox, input_shape]