def __init__(self,
input_shape,
class_num,
anchors,
batch_size=5,
is_predict=False,
train_taegets=None):
self.__input_shape = input_shape
if train_taegets is None:
train_taegets = []
train_backbone = TrainTarget.BACKBONE in train_taegets
train_rpn = TrainTarget.RPN in train_taegets
train_head = TrainTarget.HEAD in train_taegets
inputs = []
outputs = []
inputs_images = Input(shape=self.__input_shape)
inputs += [inputs_images]
resnet = ResNet(inputs_images.get_shape(),
input_layers=inputs_images,
trainable=train_backbone).get_residual_network()
rpn = RegionproposalNet(resnet.get_shape(),
anchors,
input_layers=inputs_images,
image_shape=self.__input_shape,
prev_layers=resnet,
batch_size=batch_size,
is_predict=is_predict,
trainable=train_rpn).get_network()
rpn_cls_probs, rpn_regions, rpn_prop_regs = rpn
if train_rpn and not is_predict:
inputs_rp_cls = Input(shape=[None, 1], dtype='int32')
inputs_rp_reg = Input(shape=[None, 4], dtype='float32')
inputs += [inputs_rp_cls, inputs_rp_reg]
rp_cls_losses = RPClassLoss()([inputs_rp_cls, rpn_cls_probs])
rp_reg_losses = RPRegionLoss()(
[inputs_rp_cls, inputs_rp_reg, rpn_regions])
outputs += [rp_cls_losses, rp_reg_losses]
if train_head and not is_predict:
inputs_cls = Input(shape=[None, 1], dtype='int32')
inputs_reg = Input(shape=[None, 4], dtype='float32')
inputs += [inputs_cls, inputs_reg]
dtr = DetectionTargetRegion(
positive_threshold=0.5,
positive_ratio=0.33,
image_shape=self.__input_shape,
batch_size=batch_size,
exclusion_threshold=0.1,
count_per_batch=64)([inputs_cls, inputs_reg, rpn_prop_regs])
dtr_cls_labels, dtr_offsets_labels, dtr_regions = dtr
clsses, offsets = self.__head_net(resnet,
dtr_regions,
class_num,
batch_size=batch_size)
cls_losses = ClassLoss()([dtr_cls_labels, clsses])
reg_losses = RegionLoss()(
[dtr_cls_labels, dtr_offsets_labels, offsets])
outputs += [cls_losses, reg_losses]
if is_predict:
clsses, offsets = self.__head_net(resnet,
rpn_prop_regs,
class_num,
batch_size=batch_size)
outputs = [rpn_prop_regs, clsses, offsets]
self.__network = outputs
self.__model = Model(inputs=inputs, outputs=outputs)
for output in outputs:
self.__model.add_loss(tf.reduce_mean(output))
