残差网络：操作数无法与形状一起广播 (128, 128, 16) (126, 126, 16)

Question

我正在尝试根据 paper 在 Keras 中编写 ResNet-12 代码。 但是我在第8层有一个错误，在我下面的代码中，problelem在函数Layer_Type3中。

我看不出问题出在哪里，有人可以帮忙吗？提前致谢。

错误是：

ValueError：操作数无法与形状一起广播（128， 128, 16) (126, 126, 16)

def Layer_Type1(n_output):
    # n_output: number of feature maps in the block
    # upscale: should we use the 1x1 conv2d mapping for shortcut or not

    # keras functional api: return the function of type
    # Tensor -> Tensor
    def f(x):

        # convolution
        h = Conv2D(kernel_size=3, filters=n_output,  strides=1, padding='SAME',kernel_regularizer=regularizers.l2(0.01))(x)

        # second pre-activation
        h = BatchNormalization()(h)
        h = Activation(relu)(h)

        return h

    return f


def Layer_Type2(n_output):
    # n_output: number of feature maps in the block
    # upscale: should we use the 1x1 conv2d mapping for shortcut or not

    # keras functional api: return the function of type
    # Tensor -> Tensor
    def f(x):


        # first convolution
        h = Layer_Type1(n_output)(x)

        # second convolution
        h = Conv2D(kernel_size=3, filters=n_output , strides=1, padding='SAME',kernel_regularizer=regularizers.l2(0.01))(h)

        # second pre-activation
        h = BatchNormalization()(h)


        # F_l(x) = f(x) + H_l(x):
        return add([x, h])

    return f


def Layer_Type3(n_output):

    def f(x):

        # first convolution
        h = Layer_Type1(n_output)(x)

        # second convolution
        h = Conv2D(kernel_size=3 ,filters=n_output, strides=1,kernel_regularizer=regularizers.l2(0.01))(h)

        # second pre-activation
        h = BatchNormalization()(h)
        h = AveragePooling2D(pool_size=(3,3), strides=2)(h)

        # short cut
        d = Conv2D(kernel_size=1, filters=n_output, strides=2)(x)
        d =BatchNormalization()(d)

        return add([d, h])

    return f


def Layer_Type4(n_output):

    def f(x):

        # first convolution
        h = Layer_Type1(n_output)(x)

        # second convolution
        h = Conv2D(kernel_size=3, filters=n_output, strides=1, kernel_regularizer=regularizers.l2(0.01))(h)

        # second pre-activation
        h = BatchNormalization()(h)
        h = GlobalAveragePooling2D()(h)




        return h

    return f


input_tensor = Input((256,256,1))
## 2 Layers of type1 :
x= Layer_Type1(64)(input_tensor)
x= Layer_Type1(16)(x)

# 5 layers pf type 2:
for i in range(5) :
  x = Layer_Type2(16)(x)

## 4 layers of type 3
x= Layer_Type3(16)(x) #1  # ########## Error here 
x=Layer_Type3(64)(x) #2
x=Layer_Type3(128)(x) #3
x=Layer_Type3(256)(x) # 4

# 1 layer of type 4 :
x=Layer_Type4(512)(x)


x = Dropout(0.2)(x)
# last softmax layer
x = Dense(units=2, kernel_regularizer=regularizers.l2(0.01))(x)
x = Activation(softmax)(x)

model = Model(inputs=input_tensor, outputs=x)
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

Answer 1

错误消息来自Numpy 库。也许在对 numpy 数组进行操作时会出现，但它们的形状不兼容。从您的代码中，我猜当您尝试添加两个卷积图（特征图）（一个是（126,126,126），另一个是（128,128,1））时会发生这种情况。尝试仔细检查具有填充、池大小和步幅的图层。