def backbone_resnet152(cfg):
resnet152 = resnet_v1_152(batchnorm_training=cfg.SOLVER.BN_TRAIN,
weight_decay=cfg.SOLVER.WEIGHT_DECAY,
classes=None,
weights=None,
include_top=False)
backbone = Model(inputs=resnet152.input,
outputs=resnet152.get_layer('conv4_block36_out').output)
head_to_tail = model_util.extract_submodel(
model=resnet152,
inputs=resnet152.get_layer('conv4_block36_out').output,
outputs=resnet152.get_layer('conv5_block3_out').output)
return backbone, head_to_tail
def backbone_vgg16(cfg):
vgg16 = VGG16(include_top=True, weights='imagenet')
# 不要最后的一个池化层
backbone = Model(inputs=vgg16.input,
outputs=vgg16.get_layer('block5_conv3').output)
# conv3_1之前的层不训练
for layer in backbone.layers[:7]:
layer.trainable = False
# 获取vgg16最后分类的那一部分
head_to_tail = model_util.extract_submodel(
model=vgg16,
inputs=vgg16.get_layer('block5_pool').output,
outputs=vgg16.get_layer('fc2').output)
return backbone, head_to_tail