Keras gradCAM

#######a

加载有权重的模型

model = resnet_18_res2net(input_shape=(256, 256, 1), nclass=2)
print(model.summary())

model.compile(keras.optimizers.Adam(lr=0.0001), loss='categorical_crossentropy', metrics=['accuracy'])
model_path = "train_weights/"
model.load_weights(os.path.join(model_path, "resnet_18_res2net.h5"))#resnet18_se

#####输入图像

image = np.array(Image.open('C:/Users/xian/Desktop/grad_CAM/2.bmp').convert("L"))#RGB
image = np.expand_dims(image, axis=0)
image = np.expand_dims(image, axis=-1)
image = image.astype('float32') / 255
print(image.shape)

###调用函数

cam,heatmap=grad_cam(model,image,1,'concatenate_5')#  conv5_block16_concat  concatenate_5

#叠加在原图上

print(cam.shape)
image_s = np.array(Image.open('C:/Users/xian/Desktop/grad_CAM/2.bmp').convert("RGB"))#RGB
print(image_s.shape)
img_add = cv2.addWeighted(image_s, 0.7, cam, 0.3, 0)

plt.figure("cam")
plt.imshow(cam)
plt.show()
cv2.imwrite('C:/Users/xian/Desktop/grad_CAM/2_cam.jpg', cam)



####计算CAM的函数

def grad_cam(model, x, category_index, layer_name):
    """
    Args:
       model: model
       x: image input
       category_index: category index
       layer_name: last convolution layer name
    """
    # 取得目标分类的CNN输出值，也就是loss
    class_output = model.output[:, category_index]
    print(class_output)
    # 取得想要算出梯度的层的输出
    convolution_output = model.get_layer(layer_name).output
    print(convolution_output)
    # 利用gradients函数，算出梯度公式
    grads = K.gradients(class_output, convolution_output)[0]
    print(grads)
    # 定义计算函数（tensorflow的常见做法，与一般开发语言不同，先定义计算逻辑图，之后一起计算。）
    gradient_function = K.function([model.input], [convolution_output, grads])
    print("开始问题")
    print(x.shape)

    # 根据实际的输入图像得出梯度张量(返回是一个tensor张量，VGG16 是一个7X7X512的三维张量)
    output, grads_val = gradient_function([x])
    print("输出")
    print(output)
    print("梯度")
    print(grads_val)
    output, grads_val = output[0], grads_val[0]

    # 取得所有梯度的平均值(维度降低：7X7X512 -> 512)
    weights = np.mean(grads_val, axis=(0, 1))
    # 把所有平面的平均梯度乘到最后卷积层(vgg16最后一层是池化层)上，得到一个影响输出的梯度权重图
    cam = np.dot(output, weights)

    # 把梯度权重图RGB化
    cam = cv2.resize(cam, (x.shape[1], x.shape[2]), cv2.INTER_LINEAR)
    cam = np.maximum(cam, 0)
    heatmap = cam / np.max(cam)

    # Return to BGR [0..255] from the preprocessed image
    image_rgb = x[0, :]
    image_rgb -= np.min(image_rgb)
    image_rgb = np.minimum(image_rgb, 255)

    cam = cv2.applyColorMap(np.uint8(255 * heatmap), cv2.COLORMAP_JET)
    cam = np.float32(cam) + np.float32(image_rgb)
    cam = 255 * cam / np.max(cam)
    return np.uint8(cam), heatmap