带有关键字“插值”的 UpSampling2D 引发错误答案

【问题标题】：UpSampling2D throwing error with keyword 'interpolation'带有关键字“插值”的 UpSampling2D 引发错误
【发布时间】：2019-07-21 00:23:11
【问题描述】：

当我尝试使用上采样层创建网络时，当我手动将 interpolate 关键字设置为双线性时，我遇到了一个奇怪的错误。如果我忽略它，并使用默认的“最近邻居”；它工作正常。有谁知道怎么回事？

模型代码。在“up1”层抛出错误

def build_model(self):

    chnl4_input = Input(shape=(368, 256, 4))
    chnl3_input = Input(shape=(736, 512, 3))

    conv1 = Conv2D(26, self.kernel_size, activation='relu', padding='same')(chnl4_input)
    conv2 = Conv2D(26, self.kernel_size, strides=(2, 2), activation='relu', padding='same')(conv1)

    conv5 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(conv2)
    conv6 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(conv5)

    up1 = concatenate([UpSampling2D(size=(2, 2), interpolation='bilinear')(conv6), conv1], axis=-1)
    conv7 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(up1)

    conv8 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(conv7)
    conv9 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(conv8)

    conv11 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(conv9)
    conv12 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(conv11)

    up3 = concatenate([UpSampling2D(size=(2, 2), interpolation='bilinear')(conv12), chnl3_input], axis=-1)
    conv13 = Conv2D(67, self.kernel_size, activation='relu', padding='same')(up3)

    conv14 = Conv2D(67, self.kernel_size, activation='relu', padding='same')(conv13)
    conv15 = Conv2D(32, self.kernel_size, activation='relu', padding='same')(conv14)
    conv16 = Conv2D(3, self.kernel_size, activation='relu', padding='same')(conv15)

    out = conv16

    self.model = Model(inputs=[chnl4_input, chnl3_input], outputs=[out])

    self.model.compile(optimizer=self.optimizer_func, loss=self.loss_func)
    self.model.name = 'UNET'

    return self.modele here

错误：TypeError：（'关键字参数不理解：'，'插值'）

 ~/MastersWork/Fergal/Scripts/models.py in build_model(self)
     29         conv6 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(conv5)
     30 
---> 31         up1 = concatenate([UpSampling2D(size=(2, 2), interpolation='bilinear')(conv6), conv1], axis=-1)
     32         conv7 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(up1)
     33 

~/anaconda3/envs/rhys_tensorflow/lib/python3.6/site-packages/keras/legacy/interfaces.py in wrapper(*args, **kwargs)
     89                 warnings.warn('Update your `' + object_name +
     90                               '` call to the Keras 2 API: ' + signature, stacklevel=2)
---> 91             return func(*args, **kwargs)
     92         wrapper._original_function = func
     93         return wrapper

~/anaconda3/envs/rhys_tensorflow/lib/python3.6/site-packages/keras/layers/convolutional.py in __init__(self, size, data_format, **kwargs)
   1804     @interfaces.legacy_upsampling2d_support
   1805     def __init__(self, size=(2, 2), data_format=None, **kwargs):
-> 1806         super(UpSampling2D, self).__init__(**kwargs)
   1807         self.data_format = conv_utils.normalize_data_format(data_format)
   1808         self.size = conv_utils.normalize_tuple(size, 2, 'size')

~/anaconda3/envs/rhys_tensorflow/lib/python3.6/site-packages/keras/engine/topology.py in __init__(self, **kwargs)
    291         for kwarg in kwargs:
    292             if kwarg not in allowed_kwargs:
--> 293                 raise TypeError('Keyword argument not understood:', kwarg)
    294         name = kwargs.get('name')
    295         if not name:

作为参考，关于 upSampling2D 的 Keras 页面 https://www.tensorflow.org/api_docs/python/tf/keras/layers/UpSampling2D

【问题讨论】：

您检查过，您使用的是最新版本的 tensorflow/keras 吗？
嘿，是的。我从星期一开始（02/25）开始运行 tf nightly build，但我也回到了 tf1.9，我看到了同样的错误。编辑，也是1.12。这么长时间没人注意到这一点，我觉得很奇怪，所以这可能是我的环境中的一个问题？
我使用的是 Keras 的 2.2.2 和 Tensorflow 的 1.9.0 版本。我有同样的问题。也许是一个错误
有趣。在这种情况下，我将在 keras GitHub 上提出问题。干杯
与 Tensorflow 1.12.0 相同的问题，但 Keras 2.2.4 的工作原理如文档中所述。

标签： tensorflow keras

【解决方案1】：

def bilinear_upsameple(tensor, size):
    y = tf.image.resize_bilinear(images=tensor, size=size)
    return y
dims = K.int_shape(input_tensor)
y_scaled = Lambda(lambda x : bilinear_upsameple(tensor=x, size=(dims[1]*scale, dims[2]*scale)))(input_tensor)

这里是双线性上采样的解决方法，使用 lambda 层和 tf.image.resize_bilinear 在 tf 1.12.0 上工作正常

【讨论】：