def build_class_head(width, depth, num_classes=20, num_anchors=9):
options = {
'kernel_size': 3,
'strides': 1,
'padding': 'same',
# 'kernel_initializer': initializers.normal(mean=0.0, stddev=0.01, seed=None),
}
inputs = layers.Input(shape=(None, None, width))
outputs = inputs
for i in range(depth):
outputs = layers.Conv2D(filters=width,
activation='relu',
kernel_initializer=initializers.RandomNormal(
mean=0.0, stddev=0.01, seed=None),
bias_initializer='zeros',
**options)(outputs)
# outputs = layers.Conv2D(num_anchors * num_classes, **options)(outputs)
outputs = layers.Conv2D(
filters=num_classes * num_anchors,
kernel_initializer=initializers.RandomNormal(mean=0.0,
stddev=0.01,
seed=None),
bias_initializer=PriorProbability(probability=0.01),
name='pyramid_classification',
**options)(outputs)
# (b, num_anchors_this_feature_map, 4)
outputs = layers.Reshape((-1, num_classes))(outputs)
outputs = layers.Activation('sigmoid')(outputs)
return models.Model(inputs=inputs, outputs=outputs, name='class_head')
def build_class_head(width, depth, num_classes=20):
options = {
'kernel_size':
3,
'strides':
1,
'padding':
'same',
'kernel_initializer':
initializers.RandomNormal(mean=0.0, stddev=0.01, seed=None),
}
inputs = layers.Input(shape=(None, None, width))
outputs = inputs
for i in range(depth):
outputs = layers.Conv2D(filters=width,
activation='relu',
bias_initializer='zeros',
**options)(outputs)
outputs = layers.Conv2D(
filters=num_classes,
bias_initializer=PriorProbability(probability=0.01),
activation='sigmoid',
**options)(outputs)
outputs = layers.Reshape((-1, num_classes),
name='class_head_reshape')(outputs)
return models.Model(inputs=inputs, outputs=outputs, name='class_head')
def __init__(self, width, depth, num_classes=20, num_anchors=9, separable_conv=True, freeze_bn=False, **kwargs):
super(ClassNet, self).__init__(**kwargs)
self.width = width
self.depth = depth
self.num_classes = num_classes
self.num_anchors = num_anchors
self.separable_conv = separable_conv
options = {
'kernel_size': 3,
'strides': 1,
'padding': 'same',
}
if self.separable_conv:
kernel_initializer = {
'depthwise_initializer': initializers.VarianceScaling(),
'pointwise_initializer': initializers.VarianceScaling(),
}
options.update(kernel_initializer)
self.convs = [layers.SeparableConv2D(filters=width, bias_initializer='zeros', name=f'{self.name}/class-{i}',
**options)
for i in range(depth)]
self.head = layers.SeparableConv2D(filters=num_classes * num_anchors,
bias_initializer=PriorProbability(probability=0.01),
name=f'{self.name}/class-predict', **options)
else:
kernel_initializer = {
'kernel_initializer': initializers.RandomNormal(mean=0.0, stddev=0.01, seed=None)
}
options.update(kernel_initializer)
self.convs = [layers.Conv2D(filters=width, bias_initializer='zeros', name=f'{self.name}/class-{i}',
**options)
for i in range(depth)]
self.head = layers.Conv2D(filters=num_classes*num_classes * num_anchors,
bias_initializer=PriorProbability(probability=0.01),
name='class-predict', **options)
self.bns = [
[layers.BatchNormalization(momentum=MOMENTUM, epsilon=EPSILON, name=f'{self.name}/class-{i}-bn-{j}') for j
in range(3, 8)]
for i in range(depth)]
# self.bns = [[BatchNormalization(freeze=freeze_bn, name=f'{self.name}/class-{i}-bn-{j}') for j in range(3, 8)]
# for i in range(depth)]
self.relu = layers.Lambda(lambda x: tf.nn.swish(x))
self.reshape = layers.Reshape((-1, num_classes))
self.activation = layers.Activation('sigmoid')
self.level = 0
def classification_coco(fpn_features, w_head, d_head, num_anchors,
num_classes):
options = {
'kernel_size': 3,
'strides': 1,
'padding': 'same',
'depthwise_initializer': initializers.VarianceScaling(),
'pointwise_initializer': initializers.VarianceScaling(),
}
cls_convs = [
layers.SeparableConv2D(filters=w_head,
bias_initializer='zeros',
name=f'class_net/class-{i}',
**options) for i in range(d_head)
]
cls_head_conv = layers.SeparableConv2D(
filters=num_classes * num_anchors,
bias_initializer=PriorProbability(probability=3e-4),
name='class_net/class-predict',
**options)
cls_bns = [[
layers.BatchNormalization(momentum=MOMENTUM,
epsilon=EPSILON,
name=f'class_net/class-{i}-bn-{j}')
for j in range(3, 8)
] for i in range(d_head)]
cls_relu = layers.Lambda(lambda x: tf.nn.swish(x))
classification = []
cls_reshape = layers.Reshape((-1, num_classes))
cls_activation = layers.Activation('sigmoid')
for i, feature in enumerate(fpn_features):
for j in range(d_head):
feature = cls_convs[j](feature)
feature = cls_bns[j][i](feature)
feature = cls_relu(feature)
feature = cls_head_conv(feature)
feature = cls_reshape(feature)
feature = cls_activation(feature)
classification.append(feature)
classification = layers.Concatenate(axis=1,
name='classification')(classification)
return classification
def retinanet(self, num_classes, backbone='resnet50', **kwargs):
""" Returns a retinanet model using the correct backbone.
"""
inputs_base = keras.layers.Input(shape=(None, None, 3))
input_shape = (None, None, 3)
resnet = keras.applications.ResNet50(weights="imagenet", include_top=False,input_shape=input_shape,classes=num_classes,input_tensor=inputs_base)
layer_names = ["conv3_block4_out", "conv4_block6_out", "conv5_block3_out"]
layer_outputs = [resnet.get_layer(name).output for name in layer_names]
num_anchors = 9
pyramid_feature_size = 256
regression_feature_size = 256
name = 'regression_submodel'
options = {
'kernel_size': 3,
'strides': 1,
'padding': 'same',
'kernel_initializer': keras.initializers.RandomNormal(mean=0.0, stddev=0.01, seed=None),
'bias_initializer': 'zeros'
}
inputs = keras.layers.Input(shape=(None, None, pyramid_feature_size)) # None None 256
outputs = inputs
for i in range(4):
outputs = keras.layers.Conv2D(filters=regression_feature_size, activation='relu',
name='pyramid_regression_{}'.format(i), **options)(outputs)
outputs = keras.layers.Conv2D(num_anchors * 4, name='pyramid_regression', **options)(outputs)
outputs = keras.layers.Reshape((-1, 4), name='pyramid_regression_reshape')(outputs)
default_regression_model = keras.models.Model(inputs=inputs, outputs=outputs, name=name)
#num_classes = 1
prior_probability = 0.01
classification_feature_size = 256
name = 'classification_submodel'
options = {
'kernel_size': 3,
'strides': 1,
'padding': 'same',
}
inputs = keras.layers.Input(shape=(None, None, pyramid_feature_size))
outputs = inputs
for i in range(4):
outputs = keras.layers.Conv2D(filters=classification_feature_size, activation='relu',
name='pyramid_classification_{}'.format(i),
kernel_initializer=keras.initializers.RandomNormal(mean=0.0, stddev=0.01,
seed=None),
bias_initializer='zeros',
**options
)(outputs)
outputs = keras.layers.Conv2D(filters=num_classes * num_anchors, kernel_initializer=keras.initializers.zeros(),
bias_initializer=PriorProbability(probability=prior_probability),
name='pyramid_classification',
**options
)(outputs)
outputs = keras.layers.Reshape((-1, num_classes), name='pyramid_classification_reshape')(outputs)
outputs = keras.layers.Activation('sigmoid', name='pyramid_classification_sigmoid')(outputs)
default_classification_model = keras.models.Model(inputs=inputs, outputs=outputs, name=name)
backbone_layers = layer_outputs
name = 'retinanet'
submodels = [('regression', default_regression_model),
('classification', default_classification_model)
]
C3, C4, C5 = backbone_layers
feature_size = 256
# upsample C5 to get P5 from the FPN paper
P5 = keras.layers.Conv2D(feature_size, kernel_size=1, strides=1, padding='same', name='C5_reduced')(C5)
P5_upsampled = UpsampleLike(name='P5_upsampled')([P5, C4])
P5 = keras.layers.Conv2D(feature_size, kernel_size=3, strides=1, padding='same', name='P5')(P5)
# add P5 elementwise to C4
P4 = keras.layers.Conv2D(feature_size, kernel_size=1, strides=1, padding='same', name='C4_reduced')(C4)
P4 = keras.layers.Add(name='P4_merged')([P5_upsampled, P4])
P4_upsampled = UpsampleLike(name='P4_upsampled')([P4, C3])
P4 = keras.layers.Conv2D(feature_size, kernel_size=3, strides=1, padding='same', name='P4')(P4)
# add P4 elementwise to C3
P3 = keras.layers.Conv2D(feature_size, kernel_size=1, strides=1, padding='same', name='C3_reduced')(C3)
P3 = keras.layers.Add(name='P3_merged')([P4_upsampled, P3])
P3 = keras.layers.Conv2D(feature_size, kernel_size=3, strides=1, padding='same', name='P3')(P3)
# "P6 is obtained via a 3x3 stride-2 conv on C5"
P6 = keras.layers.Conv2D(feature_size, kernel_size=3, strides=2, padding='same', name='P6')(C5)
# "P7 is computed by applying ReLU followed by a 3x3 stride-2 conv on P6"
P7 = keras.layers.Activation('relu', name='C6_relu')(P6)
P7 = keras.layers.Conv2D(feature_size, kernel_size=3, strides=2, padding='same', name='P7')(P7)
features = [P3, P4, P5, P6, P7]
pyramids = []
for n, m in submodels:
list_models = []
for f in features:
list_models.append(m(f))
pyramids.append(keras.layers.Concatenate(axis=1, name=n)(list_models))
backbone_retinanet = keras.models.Model(inputs=inputs_base, outputs=pyramids, name='retinanet')
return backbone_retinanet