def __init__(self,
num_classes,
backbone_resnet='resnet50',
input_size=512,
max_objects=100,
score_threshold=0.1,
nms=True,
flip_test=False,
mode="train"):
self.nms = nms
self.flip_test = flip_test
self.score_threshold = score_threshold
self.max_objects = max_objects
self.output_size = input_size // 4
self.image_input = Input(shape=(None, None, 3))
self.heatmap_input = Input(shape=(self.output_size, self.output_size,
num_classes))
self.width_height_input = Input(shape=(max_objects, 2))
self.offset_input = Input(shape=(max_objects, 2))
self.offset_mask_input = Input(shape=(max_objects, ))
self.index_input = Input(shape=(max_objects, ))
self.num_classes = num_classes
self.backbone = resnet_models.ResNet50(
self.image_input, include_top=False
) if backbone_resnet == 'resnet50' else resnet_models.ResNet101(
image_input, include_top=False)
def resnet_retinanet(num_classes, backbone='resnet50', inputs=None):
# choose default input
if inputs is None:
inputs = keras.layers.Input(shape=(224, 224, 3))
# create the resnet backbone
if backbone == 'resnet50':
resnet = Resmodel.ResNet50(inputs, include_top=False, freeze_bn=True)
elif backbone == 'resnet101':
resnet = Resmodel.ResNet101(inputs, include_top=False, freeze_bn=True)
elif backbone == 'resnet152':
resnet = Resmodel.ResNet152(inputs, include_top=False, freeze_bn=True)
# create the full model
model = retinanet(inputs=inputs, num_classes=num_classes, backbone=resnet)
model.summary()
plot_model(model, to_file='retinanet_model.png', show_shapes=True)
return model
def centernet(num_classes,
backbone='resnet18',
input_size=512,
max_objects=100,
score_threshold=0.1,
nms=True,
flip_test=False,
freeze_bn=False):
assert backbone in [
'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152'
]
output_size = input_size // 4
image_input = Input(shape=(None, None, 3))
hm_input = Input(shape=(output_size, output_size, num_classes))
wh_input = Input(shape=(max_objects, 2))
reg_input = Input(shape=(max_objects, 2))
reg_mask_input = Input(shape=(max_objects, ))
index_input = Input(shape=(max_objects, ))
print('\nBACKBONE IS {}\n'.format(backbone))
if backbone == 'resnet18':
resnet = resnet_models.ResNet18(image_input,
include_top=False,
freeze_bn=freeze_bn)
elif backbone == 'resnet34':
resnet = resnet_models.ResNet34(image_input,
include_top=False,
freeze_bn=freeze_bn)
elif backbone == 'resnet50':
resnet = resnet_models.ResNet50(image_input,
include_top=False,
freeze_bn=freeze_bn)
# resnet = ResNet50(input_tensor=image_input, include_top=False)
elif backbone == 'resnet101':
resnet = resnet_models.ResNet101(image_input,
include_top=False,
freeze_bn=freeze_bn)
else:
resnet = resnet_models.ResNet152(image_input,
include_top=False,
freeze_bn=freeze_bn)
# (b, 16, 16, 2048)
C5 = resnet.outputs[-1]
# C5 = resnet.get_layer('activation_49').output
x = Dropout(rate=0.5)(C5)
# x = Conv2D(64, 1, padding='same', use_bias=True, kernel_initializer='he_normal', kernel_regularizer=l2(5e-4))(
# x)
# decoder
num_filters = 256
for i in range(3):
num_filters = num_filters // pow(2, i)
# x = Conv2D(num_filters, 3, padding='same', use_bias=False, kernel_initializer='he_normal', kernel_regularizer=l2(5e-4))(
# x)
# x = BatchNormalization()(x)
# x = ReLU()(x)
x = Conv2DTranspose(num_filters, (4, 4),
strides=2,
use_bias=False,
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(x)
x = BatchNormalization()(x)
x = ReLU()(x)
# hm header
y1 = Conv2D(64,
3,
padding='same',
use_bias=True,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(x)
y1 = BatchNormalization()(y1)
y1 = ReLU()(y1)
y1 = Conv2D(num_classes,
1,
use_bias=True,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4),
activation='sigmoid')(y1)
# wh header
y2 = Conv2D(64,
3,
padding='same',
use_bias=True,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(x)
y2 = BatchNormalization()(y2)
y2 = ReLU()(y2)
y2 = Conv2D(2,
1,
use_bias=True,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(y2)
# reg header
y3 = Conv2D(64,
3,
padding='same',
use_bias=True,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(x)
y3 = BatchNormalization()(y3)
y3 = ReLU()(y3)
y3 = Conv2D(2,
1,
use_bias=True,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(y3)
loss_ = Lambda(loss, name='centernet_loss')([
y1, y2, y3, hm_input, wh_input, reg_input, reg_mask_input, index_input
])
model = Model(inputs=[
image_input, hm_input, wh_input, reg_input, reg_mask_input, index_input
],
outputs=[loss_])
# detections = decode(y1, y2, y3)
detections = Lambda(lambda x: decode(*x,
max_objects=max_objects,
score_threshold=score_threshold,
nms=nms,
flip_test=flip_test,
num_classes=num_classes))(
[y1, y2, y3])
prediction_model = Model(inputs=image_input, outputs=detections)
debug_model = Model(inputs=image_input, outputs=[y1, y2, y3])
return model, prediction_model, debug_model
def centernet(num_classes,
backbone='resnet50',
input_size=512,
max_objects=100,
score_threshold=0.1,
nms=True,
flip_test=False,
freeze_bn=True):
assert backbone in [
'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152'
]
output_size = input_size // 4
image_input = Input(shape=(None, None, 3))
if backbone == 'resnet18':
resnet = resnet_models.ResNet18(image_input,
include_top=False,
freeze_bn=freeze_bn)
elif backbone == 'resnet34':
resnet = resnet_models.ResNet34(image_input,
include_top=False,
freeze_bn=freeze_bn)
elif backbone == 'resnet50':
resnet = resnet_models.ResNet50(image_input,
include_top=False,
freeze_bn=freeze_bn)
elif backbone == 'resnet101':
resnet = resnet_models.ResNet101(image_input,
include_top=False,
freeze_bn=freeze_bn)
else:
resnet = resnet_models.ResNet152(image_input,
include_top=False,
freeze_bn=freeze_bn)
# (b, 16, 16, 2048)
C5 = resnet.outputs[-1]
x = Dropout(rate=0.5)(C5)
# decoder
num_filters = 256
for i in range(3):
num_filters = num_filters // pow(2, i)
x = Conv2DTranspose(num_filters, (4, 4),
strides=2,
use_bias=False,
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(x)
x = BatchNormalization()(x)
x = ReLU()(x)
# hm header
y1 = Conv2D(64,
3,
padding='same',
use_bias=False,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(x)
y1 = BatchNormalization()(y1)
y1 = ReLU()(y1)
y1 = Conv2D(num_classes,
1,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4),
activation='sigmoid')(y1)
# wh header
y2 = Conv2D(64,
3,
padding='same',
use_bias=False,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(x)
y2 = BatchNormalization()(y2)
y2 = ReLU()(y2)
y2 = Conv2D(2,
1,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(y2)
# reg header
y3 = Conv2D(64,
3,
padding='same',
use_bias=False,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(x)
y3 = BatchNormalization()(y3)
y3 = ReLU()(y3)
y3 = Conv2D(2,
1,
kernel_initializer='he_normal',
kernel_regularizer=l2(5e-4))(y3)
detections = Lambda(lambda x: decode(*x,
max_objects=max_objects,
score_threshold=score_threshold,
nms=nms,
flip_test=flip_test,
num_classes=num_classes))(
[y1, y2, y3])
prediction_model = Model(inputs=image_input, outputs=detections)
return prediction_model