AttributeError: 'Node' object has no attribute 'output_masks'

错误:

Traceback (most recent call last):
  File "/home/nianxiongdi/algorithm/deform-conv/scripts/scaled_mnist1.py", line 94, in <module>
    inputs, outputs, model = get_deform_cnn(use_cpu=False, print_summary=True)
  File "/home/nianxiongdi/algorithm/deform-conv/deform_conv/cnn.py", line 405, in get_deform_cnn
    conv_block_1 = buildConv2DBlock(inputs, 64, 1, 2)
  File "/home/nianxiongdi/algorithm/deform-conv/deform_conv/cnn.py", line 251, in buildConv2DBlock
    conv2d = ConvOffset2D(filters, name='conv12_offset')(conv2d)
  File "/home/nianxiongdi/anaconda3/envs/py36/lib/python3.6/site-packages/keras/engine/topology.py", line 583, in __call__
    previous_mask = _collect_previous_mask(inputs)
  File "/home/nianxiongdi/anaconda3/envs/py36/lib/python3.6/site-packages/keras/engine/topology.py", line 2737, in _collect_previous_mask
    mask = node.output_masks[tensor_index]

 

大概代码：

 A.py

from tensorflow.keras import Model, Input
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Conv2DTranspose, Lambda, Layer, BatchNormalization, Activation
from tensorflow.keras import backend as K

B.py

from keras.layers import Conv2D

class ConvOffset2D(Conv2D):
    """ConvOffset2D"""

    def __init__(self, filters, init_normal_stddev=0.01, **kwargs):
        """Init"""

        self.filters = filters # 32  fea
        super(ConvOffset2D, self).__init__(
            self.filters * 2, (3, 3), padding='same', use_bias=False,
            # TODO gradients are near zero if init is zeros
            kernel_initializer='zeros',
            # kernel_initializer=RandomNormal(0, init_normal_stddev),
            **kwargs
        )

 

A.py 调用了B.py  出错的原因是因为 类型不同：

调用到： self.assert_input_compatibility(inputs)  Checks compatibility between the layer and provided inputs.

 

解决问题：

     把B.py与A.py文件中的import的类型进行统一

 

from tensorflow.keras.layers import Conv2D

class ConvOffset2D(Conv2D):
    """ConvOffset2D"""

    def __init__(self, filters, init_normal_stddev=0.01, **kwargs):
        """Init"""

        self.filters = filters # 32  fea
        super(ConvOffset2D, self).__init__(
            self.filters * 2, (3, 3), padding='same', use_bias=False,
            # TODO gradients are near zero if init is zeros
            kernel_initializer='zeros',
            # kernel_initializer=RandomNormal(0, init_normal_stddev),
            **kwargs
        )