keras可视化特征响应

参考：https://blog.csdn.net/C_chuxin/article/details/85265082

以目标检测基于VGG16的SSD为例，先建立SSD模型model

示例图片如下。若最大概率检测到火车train的anchor为6856号(0开始)，即19*19特征图(即fc7)的(9,9),aspect ratio=1.，则以下实验观察该anchor的train类在fc7特征图上的响应。该示例图片在batch中为1号，令i=1

from keras.models import Model
import cv2 
import matplotlib.pyplot as plt
import numpy as np
import keras.backend as K

i = 1
model1 = Model(model.input, model.get_layer('fc7_mbox_conf').output)
pred1 = model1.predict(processed_images)#(batch, height, width, channel)
r = np.argmax(pred1[i,:,:,19])//19#9, 该图片位于batch的1号
c = np.argmax(pred1[i,:,:,19])-r*19#9
most_act = model1.output[i,r,c,19]

grads = K.gradients(most_act, model.get_la