在 keras cnn 中可视化过滤器

Question

def build_model(network):
    model = Sequential()
    model.add(Conv2D(6, (5,5), padding='valid', activation = 'relu', kernel_initializer='he_normal', input_shape=(32,32,3)))
    print(np.asarray(model.get_weights())[0].shape)
    model.add(MaxPooling2D((2, 2), strides=(2, 2)))
    model.add(Conv2D(16, (5,5), padding='valid', activation = 'relu', kernel_initializer='he_normal'))
    model.add(MaxPooling2D((2, 2), strides=(2, 2)))

    model.add(Flatten())
    model.add(Dense(120, activation = 'relu', kernel_initializer='he_normal'))
    model.add(Dense(84, activation = 'relu', kernel_initializer='he_normal'))
    model.add(Dense(10, activation = 'softmax', kernel_initializer='he_normal'))
    sgd = optimizers.SGD(lr=learning_rate)
    model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
    return model

这是 Keras 中的模型。我想在第一个 conv 层中可视化过滤器。我想绘制过滤器本身，而不是当我们将梯度一直反向传播到图像时出现的模式。

我找到了一种获取权重的方法 - 使用 model.get_weights()

如何绘制这些权重？ np.asarray(model.get_weights())[0] 的形状是 (5,5,3,6)。

如何用它制作六个尺寸为 5x5x3 的滤镜？

Answer 1

使用 matplotlib 可以绘制切片。

你应该自己将切片分开，例如：

import matplotlib.pyplot as plt

#normalize these filters first, otherwise they won't be in a suitable range for plotting:

maxVal = filters.max()
minVal = filters.min()
absMax = max(abs(minVal),abs(maxVal))

filters = (filters / absMax)*255


for outputChannel in range(6):
    for inputChannel in range(3):
        filt = filters[:,:,inputChannel,outputChannel]

      #a trick to see negatives as blue and positives as red
        imageRed = np.array(filt)
        imageBlue = np.array(filt)
        imageRed[imageRed<0] = 0
        imageBlue[imageBlue>0]= 0

        print(imageRed)
        print(imageBlue)

        finalImage = np.zeros((filt.shape[0],filt.shape[1],3))
        finalImage[:,:,0] = imageRed
        finalImage[:,:,2] = -imageBlue

        #plot image here
        plt.figure()  
        plt.imshow(finalImage)