修正input_img_data = np.random.random((1, 150, 150, 3)) * 20 + 128.

原始程式碼:

def generate_pattern(layer_name, filter_index, size=150):
    # 構建一個最大化激活的 損失函數  loss function
    # 考慮的層的 第n個 filter
    layer_output = model.get_layer(layer_name).output
    loss = K.mean(layer_output[:, :, :, filter_index])

    # 計算這種損失的輸入圖像的梯度
    grads = K.gradients(loss, model.input)[0]

    # Normalization gradient 梯度
    grads /= (K.sqrt(K.mean(K.square(grads))) + 1e-5)

    # 函數返回 給定 輸入圖片的 損失 和 梯度
    iterate = K.function([model.input], [loss, grads])
   
    # 帶有一些噪音的灰色圖像
    input_img_data = np.random.random((1, size, size, 3)) * 20 + 128.

    # Run 梯度上升 40步
    step = 1.
    for i in range(40):
        loss_value, grads_value = iterate([input_img_data])
        input_img_data += grads_value * step
       
    img = input_img_data[0]
    return deprocess_image(img)

出現錯誤

Cell In[34], line 3
    # 考慮的層的 第n個 filter
    ^
SyntaxError: invalid non-printable character U+00A0

修正成這樣就好了

def generate_pattern(layer_name, filter_index, size=150):
    # Construct a loss function that maximizes the activation for the nth filter in the considered layer
    layer_output = model.get_layer(layer_name).output
    loss = K.mean(layer_output[:, :, :, filter_index])

    # Compute the gradient of this loss with respect to the input image
    grads = K.gradients(loss, model.input)[0]

    # Normalize the gradient
    grads /= (K.sqrt(K.mean(K.square(grads))) + 1e-5)

    # Define a function to return the loss and gradient given an input image
    iterate = K.function([model.input], [loss, grads])
    
    # Generate a grayscale image with some noise
    input_img_data = np.random.random((1, size, size, 3)) * 20 + 128.

    # Run gradient ascent for 40 steps
    step = 1.
    for i in range(40):
        loss_value, grads_value = iterate([input_img_data])
        input_img_data += grads_value * step
        
    img = input_img_data[0]
    return deprocess_image(img)