pretrained_models.py 文件源码

def resnet_v1_50_fn(input_tensor: tf.Tensor, is_training=False, blocks=4, weight_decay=0.0001, renorm=True) -> tf.Tensor:
    with slim.arg_scope(nets.resnet_v1.resnet_arg_scope(weight_decay=weight_decay, batch_norm_decay=0.999)), \
         slim.arg_scope([layers.batch_norm], renorm_decay=0.95, renorm=renorm):
        input_tensor = mean_substraction(input_tensor)
        assert 0 < blocks <= 4
        blocks_list = [
              nets.resnet_v1.resnet_v1_block('block1', base_depth=64, num_units=3, stride=2),
              nets.resnet_v1.resnet_v1_block('block2', base_depth=128, num_units=4, stride=2),
              nets.resnet_v1.resnet_v1_block('block3', base_depth=256, num_units=6, stride=2),
              nets.resnet_v1.resnet_v1_block('block4', base_depth=512, num_units=3, stride=1),
        ]
        net, endpoints = nets.resnet_v1.resnet_v1(input_tensor,
                                                  blocks=blocks_list[:blocks],
                                                  num_classes=None,
                                                  is_training=is_training,
                                                  global_pool=False,
                                                  output_stride=None,
                                                  include_root_block=True,
                                                  reuse=None,
                                                  scope='resnet_v1_50')

        desired_endpoints = ['resnet_augmented/resnet_v1_50/conv1',
                             'resnet_v1_50/block1/unit_2/bottleneck_v1',
                             'resnet_v1_50/block2/unit_3/bottleneck_v1',
                             'resnet_v1_50/block3/unit_5/bottleneck_v1',
                             'resnet_v1_50/block4/unit_2/bottleneck_v1'
                             ]

        intermediate_layers = list()
        for d in desired_endpoints:
            intermediate_layers.append(endpoints[d])

        return net, intermediate_layers