Keras中的一维CNN过滤器可视化

Question

我想使用梯度上升方法（https://blog.keras.io/how-convolutional-neural-networks-see-the-world.html）可视化我的1D-CNN的1D过滤器。下面给出的是我的模型摘要和我尝试过的代码。

    Model.summary() #after removing the fully connected layers. 

    Layer (type)                 Output Shape              Param #   
    =================================================================
    conv1d_5_input (InputLayer)  (None, 289, 2493)         0         
    _________________________________________________________________
    conv1d_5 (Conv1D)            (None, 131, 22)           1535710   
    _________________________________________________________________
    batch_normalization_12 (Batc (None, 131, 22)           88        
    _________________________________________________________________
    activation_12 (Activation)   (None, 131, 22)           0         
    _________________________________________________________________
    max_pooling1d_5 (MaxPooling1 (None, 65, 22)            0         
    _________________________________________________________________
    dropout_14 (Dropout)         (None, 65, 22)            0         
    _________________________________________________________________
    conv1d_6 (Conv1D)            (None, 30, 30)            23790     
    _________________________________________________________________
    batch_normalization_13 (Batc (None, 30, 30)            120       
    _________________________________________________________________
    activation_13 (Activation)   (None, 30, 30)            0         
    _________________________________________________________________
    max_pooling1d_6 (MaxPooling1 (None, 15, 30)            0         
    =================================================================
    Total params: 1,559,708
    Trainable params: 1,559,604
    Non-trainable params: 104



    layer_dict = dict([(layer.name, layer) for layer in model2.layers])

    input_img = model2.input

    layer_name = 'conv1d_5'
    filter_index = 18        

    layer_output = layer_dict[layer_name].output

    loss_fxn = K.mean(layer_output[ :, :, filter_index])

    # Find gradient of loss_fxn w.r.t. the input_img
    gradient = K.gradients(loss_fxn, input_img)[0]

    # Normalize the gradient

    gradient = gradient / (K.sqrt(K.mean(K.square(gradient))) + 1e-05)
    iterate = K.function([input_img], [loss_fxn, gradient])
    w = 256
    h = 256

    input_img_data = np.random.random((1, w, h, 3))*20 + 128  ##NOISY IMAGE 


    step_size = 1

    steps = 100
    for i in range(steps):
        loss, grads = iterate([input_img_data])
        input_img_data = input_img_data + grads*step_size
        print("Step: {}/{}, Loss: {}".format(i+1, steps, loss))

这是我得到的错误：

  

InvalidArgumentError：转置期望大小为5的向量。但是   input（1）是大小为4 [[{{node   conv1d_5_1 / convolution / Conv2D-0-TransposeNHWCToNCHW-LayoutOptimizer}}]]]      [[{{node Mean_10}}]]

如何可视化一维滤镜？