Question

我试图创建一个能够区分10个图像类的keras模型。我为每个对象创建了12个观察点，然后将它们传递给模型。

input_shape = (12, 1, 32, 32)
model = get_test_shared_model()
input = keras.layers.Input(shape=input_shape)
views = SplitLayer(num_views)(input)  # list of keras-tensors
processed_views = []  # empty list
for view in views:
    x = model(view)
    processed_views.append(x)

在这个阶段，我自己对每个观察点进行了计算，现在我尝试使用最大层将它们合并在一起。 Conv2D线路上发生错误。关于如何解决这个问题的任何想法？

pooled_views = keras.layers.Maximum()(processed_views)
prediction = Conv2D(32, kernel_size=(3, 3), activation='relu')(pooled_views)

Split Layer源代码是：

class SplitLayer(keras.layers.Layer):
    """
    Layer expects a tensor (multi-dimensonal array) of shape (samples, views, 
    ...)
    and returns a list of #views elements, each of shape (samples, ...)
    """

    def __init__(self, num_splits, **kwargs):
        self.num_splits = num_splits
        super(SplitLayer, self).__init__(**kwargs)

    def call(self, x):
        a = [x[:, i] for i in range(self.num_splits)]
        return a

    def compute_output_shape(self, input_shape):
        return [(input_shape[0],) + input_shape[2:]]*self.num_splits


def get_test_shared_model():
    num_channels = 4  # for example
    cnn = keras.models.Sequential()
    cnn.add(keras.layers.Conv2D(num_channels, kernel_size=(3, 3),
                            activation='relu',
                            input_shape=(32, 32, 1)))
    cnn.add(keras.layers.MaxPooling2D(pool_size=(2, 2)))
    cnn.add(keras.layers.Flatten())
    cnn.add(keras.layers.Dense(128, activation='relu'))
    return cnn

Keras ValueError：输入0与图层conv2d_2不兼容：预期ndim = 4，发现ndim = 2

0 个答案: