卷积网络与特征可视化

人们常说神经网络的解释性不强，即神经网络模型是一个“黑盒”，它学到的经验很难用人类可以理解的方式呈现（反例是树模型，可解释性强）。这种说法不完全正确，卷积神经网络学习到的“经验”就非常适合可视化，因为很大程度上它们是视觉概念的表示。

可视化中间激活方法

可视化中间激活（层的输出通常被称为该层的激活，即激活函数的输出），是指对于给定输入，展示网络各个卷积层和池化层输出的特征图。
首先我们找一张可爱的猫咪镇楼......

然后将该图片读取，并处理成张量格式

from keras.preprocessing import image  # 将图像处理为4D张量形式
import matplotlib.pyplot as plt
import numpy as np
import os
# 忽略硬件加速的警告信息
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
# 获取当前目录地址
FILE_DIR = os.path.dirname(os.path.abspath(__file__))
# 设置图像参数尺寸
target_size = (224, 224, 3)

def path_to_tensor(img_path):
    '''图片格式处理'''
    img = image.load_img(img_path, target_size=target_size)
    img_tensor = image.img_to_array(img)
    img_tensor = np.expand_dims(img_tensor, axis=0).astype('float32')/255
    return img_tensor

if __name__ == '__main__':
    # 读取图片并进行格式处理
    img_path = os.path.join(FILE_DIR, 'cat.jpg')
    img_tensor = path_to_tensor(img_path)

卷积网络使用了keras自带的VGG16，提取特征

# 模型初始化
model = vgg16.VGG16(weights='imagenet', include_top=False)
model.summary()

然后抽取中间层输出，主要有两种方式

# 采用K.function抽取中间层
layer_1 = K.function([model.layers[0].input], [model.layers[1].output])
layer_2 = K.function([model.layers[0].input], [model.get_layer('block1_conv2').output])   
# 构造一个新模型提取输出
activation_model = Model(inputs=model.layers[0].input, outputs=model.layers[3].output)

feature_maps1 = layer_1([img_tensor])[0]
feature_maps2 = layer_2([img_tensor])[0]
feature_maps3 = activation_model.predict([img_tensor])[0]

#plt.imshow(feature_maps1[0,:,:,3], cmap='viridis')
#plt.imshow(feature_maps2[0,:,:,3], cmap='viridis')
plt.imshow(feature_maps3[:,:,60], cmap='viridis')    # 可以改变数字以切换通道查看不同的特征图
plt.show()

接下来我们将中间层激活的所有通道可视化

#-*- coding:utf-8 -*-
from keras import backend as K
from keras.models import Model
from keras.applications import vgg16
from keras.preprocessing import image  # 将图像处理为4D张量形式
import matplotlib.pyplot as plt
import numpy as np
import os
# 忽略硬件加速的警告信息
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
# 获取当前目录地址
FILE_DIR = os.path.dirname(os.path.abspath(__file__))
# 设置图像参数尺寸
target_size = (224, 224, 3)

def path_to_tensor(img_path):
    '''图片格式处理'''
    img = image.load_img(img_path, target_size=target_size)
    img_tensor = image.img_to_array(img)
    img_tensor = np.expand_dims(img_tensor, axis=0).astype('float32')/255
    return img_tensor

if __name__ == '__main__':
    # 读取图片并进行格式处理
    img_path = os.path.join(FILE_DIR, 'cat.jpg')
    img_tensor = path_to_tensor(img_path)

    # 模型初始化
    model = vgg16.VGG16(weights='imagenet', include_top=False)
    # model.summary()

    # 构造一个新模型提取输出
    layer_outputs = [layer.output for layer in model.layers[1:8]]
    activation_model = Model(inputs=[model.layers[0].input], outputs=layer_outputs)
    activations = activation_model.predict([img_tensor])

    layer_names = []
    for layer in model.layers[1:8]:
        layer_names.append(layer.name)

    # 每行显示通道数量
    images_per_row = 16
    
    # 循环打印每一层的特征图
    for layer_name, layer_activation in zip(layer_names, activations):
        print(layer_name)
        print(layer_activation.shape)
        # 特征图中通道个数
        n_features = layer_activation.shape[-1]
        # 特征图形状为(1, width, height, array_len)
        size = layer_activation.shape[1]
        # 将激活通道平铺
        n_cols = n_features // images_per_row  # 需要多少行才能排满

        display_grid = np.zeros((n_cols*size, images_per_row*size))

        for col in range(n_cols):
            for row in range(images_per_row):
                # 定位特征通道
                channel_image = layer_activation[:,:,:,(col*images_per_row+row)]
                # 对特征进行后处理使其更美观
                channel_image -= channel_image.mean()
                channel_image *= 64
                channel_image += 128
                channel_image = np.clip(channel_image, 0, 255).astype('uint8')
                display_grid[col * size : (col + 1) * size, row * size : (row + 1) * size] = channel_image
            
        plt.title(layer_name)
        plt.imshow(display_grid)
        plt.show()

block1_conv1
block1_conv2
block2_conv1
block3_conv1

随着模型越来越深，提取的通道数越来越多，特征也更为抽象。