【Pytorch学习笔记】通过实现Grad-CAM学习module中的forward_hook和backward_hook函数

一、pandas是什么？

1.引入库

代码如下（示例）：

import cv2
import os
import numpy as np
from PIL import Image
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.transforms as transforms

2.定义网络

代码如下（示例）：

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool1 = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.pool2 = nn.MaxPool2d(2, 2)
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool1(F.relu(self.conv1(x)))
        x = self.pool1(F.relu(self.conv2(x)))
        x = x.view(-1, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

3.将img进行预处理，并转换成模型输入所需的形式—— BCH*W

def img_transform(img_in, transform):
    """
    将img进行预处理，并转换成模型输入所需的形式—— B*C*H*W
    :param img_roi: np.array
    :return:
    """
    img = img_in.copy()
    img = Image.fromarray(np.uint8(img))
    img = transform(img)
    img = img.unsqueeze(0)    # C*H*W --> B*C*H*W
    return img

4.读取图片，转为模型可读的形式

def img_preprocess(img_in):
    """
    读取图片，转为模型可读的形式
    :param img_in: ndarray, [H, W, C]
    :return: PIL.image
    """
    img = img_in.copy()
    img = cv2.resize(img,(32, 32))
    img = img[:, :, ::-1]   # BGR --> RGB
    transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize([0.4948052, 0.48568845, 0.44682974], [0.24580306, 0.24236229, 0.2603115])
    ])
    img_input = img_transform(img, transform)
    return img_input

ef backward_hook(module, grad_in, grad_out):
    grad_block.append(grad_out[0].detach())


def farward_hook(module, input, output):
    fmap_block.append(output)


def show_cam_on_image(img, mask, out_dir):
    heatmap = cv2.applyColorMap(np.uint8(255*mask), cv2.COLORMAP_JET)
    heatmap = np.float32(heatmap) / 255
    cam = heatmap + np.float32(img)
    cam = cam / np.max(cam)

    path_cam_img = os.path.join(out_dir, "cam.jpg")
    path_raw_img = os.path.join(out_dir, "raw.jpg")
    if not os.path.exists(out_dir):
        os.makedirs(out_dir)
    cv2.imwrite(path_cam_img, np.uint8(255 * cam))
    cv2.imwrite(path_raw_img, np.uint8(255 * img))

5.计算类向量

def comp_class_vec(ouput_vec, index=None):
    """
    计算类向量
    :param ouput_vec: tensor
    :param index: int，指定类别
    :return: tensor
    """
    if not index:
        index = np.argmax(ouput_vec.cpu().data.numpy())
    else:
        index = np.array(index)
    index = index[np.newaxis, np.newaxis]
    index = torch.from_numpy(index)
    one_hot = torch.zeros(1, 10).scatter_(1, index, 1)
    one_hot.requires_grad = True
    class_vec = torch.sum(one_hot * output)  # one_hot = 11.8605

    return class_vec

6.依据梯度和特征图，生成cam

def gen_cam(feature_map, grads):
    """
    依据梯度和特征图，生成cam
    :param feature_map: np.array， in [C, H, W]
    :param grads: np.array， in [C, H, W]
    :return: np.array, [H, W]
    """
    cam = np.zeros(feature_map.shape[1:], dtype=np.float32)  # cam shape (H, W)

    weights = np.mean(grads, axis=(1, 2))  #

    for i, w in enumerate(weights):
        cam += w * feature_map[i, :, :]

    cam = np.maximum(cam, 0)
    cam = cv2.resize(cam, (32, 32))
    cam -= np.min(cam)
    cam /= np.max(cam)

    return cam

7.主函数

if __name__ == '__main__':

    BASE_DIR = os.path.dirname(os.path.abspath(__file__))
    path_img = os.path.join(BASE_DIR, "..", "..", "Data", "cam_img", "test_img_8.png")
    path_net = os.path.join(BASE_DIR, "..", "..", "Data", "net_params_72p.pkl")
    output_dir = os.path.join(BASE_DIR, "..", "..", "Result", "backward_hook_cam")

    classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
    fmap_block = list()
    grad_block = list()

    # 图片读取；网络加载
    img = cv2.imread(path_img, 1)  # H*W*C
    img_input = img_preprocess(img)
    net = Net()
    net.load_state_dict(torch.load(path_net))

    # 注册hook
    net.conv2.register_forward_hook(farward_hook)
    net.conv2.register_backward_hook(backward_hook)

    # forward
    output = net(img_input)
    idx = np.argmax(output.cpu().data.numpy())
    print("predict: {}".format(classes[idx]))

    # backward
    net.zero_grad()
    class_loss = comp_class_vec(output)
    class_loss.backward()

    # 生成cam
    grads_val = grad_block[0].cpu().data.numpy().squeeze()
    fmap = fmap_block[0].cpu().data.numpy().squeeze()
    cam = gen_cam(fmap, grads_val)

    # 保存cam图片
    img_show = np.float32(cv2.resize(img, (32, 32))) / 255
    show_cam_on_image(img_show, cam, output_dir)

二、完整程序

# coding: utf-8
"""
通过实现Grad-CAM学习module中的forward_hook和backward_hook函数
"""
import cv2
import os
import numpy as np
from PIL import Image
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.transforms as transforms


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool1 = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.pool2 = nn.MaxPool2d(2, 2)
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool1(F.relu(self.conv1(x)))
        x = self.pool1(F.relu(self.conv2(x)))
        x = x.view(-1, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x


def img_transform(img_in, transform):
    """
    将img进行预处理，并转换成模型输入所需的形式—— B*C*H*W
    :param img_roi: np.array
    :return:
    """
    img = img_in.copy()
    img = Image.fromarray(np.uint8(img))
    img = transform(img)
    img = img.unsqueeze(0)    # C*H*W --> B*C*H*W
    return img


def img_preprocess(img_in):
    """
    读取图片，转为模型可读的形式
    :param img_in: ndarray, [H, W, C]
    :return: PIL.image
    """
    img = img_in.copy()
    img = cv2.resize(img,(32, 32))
    img = img[:, :, ::-1]   # BGR --> RGB
    transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize([0.4948052, 0.48568845, 0.44682974], [0.24580306, 0.24236229, 0.2603115])
    ])
    img_input = img_transform(img, transform)
    return img_input


def backward_hook(module, grad_in, grad_out):
    grad_block.append(grad_out[0].detach())


def farward_hook(module, input, output):
    fmap_block.append(output)


def show_cam_on_image(img, mask, out_dir):
    heatmap = cv2.applyColorMap(np.uint8(255*mask), cv2.COLORMAP_JET)
    heatmap = np.float32(heatmap) / 255
    cam = heatmap + np.float32(img)
    cam = cam / np.max(cam)

    path_cam_img = os.path.join(out_dir, "cam.jpg")
    path_raw_img = os.path.join(out_dir, "raw.jpg")
    if not os.path.exists(out_dir):
        os.makedirs(out_dir)
    cv2.imwrite(path_cam_img, np.uint8(255 * cam))
    cv2.imwrite(path_raw_img, np.uint8(255 * img))


def comp_class_vec(ouput_vec, index=None):
    """
    计算类向量
    :param ouput_vec: tensor
    :param index: int，指定类别
    :return: tensor
    """
    if not index:
        index = np.argmax(ouput_vec.cpu().data.numpy())
    else:
        index = np.array(index)
    index = index[np.newaxis, np.newaxis]
    index = torch.from_numpy(index)
    one_hot = torch.zeros(1, 10).scatter_(1, index, 1)
    one_hot.requires_grad = True
    class_vec = torch.sum(one_hot * output)  # one_hot = 11.8605

    return class_vec


def gen_cam(feature_map, grads):
    """
    依据梯度和特征图，生成cam
    :param feature_map: np.array， in [C, H, W]
    :param grads: np.array， in [C, H, W]
    :return: np.array, [H, W]
    """
    cam = np.zeros(feature_map.shape[1:], dtype=np.float32)  # cam shape (H, W)

    weights = np.mean(grads, axis=(1, 2))  #

    for i, w in enumerate(weights):
        cam += w * feature_map[i, :, :]

    cam = np.maximum(cam, 0)
    cam = cv2.resize(cam, (32, 32))
    cam -= np.min(cam)
    cam /= np.max(cam)

    return cam


if __name__ == '__main__':

    BASE_DIR = os.path.dirname(os.path.abspath(__file__))
    path_img = os.path.join(BASE_DIR, "..", "..", "Data", "cam_img", "test_img_8.png")
    path_net = os.path.join(BASE_DIR, "..", "..", "Data", "net_params_72p.pkl")
    output_dir = os.path.join(BASE_DIR, "..", "..", "Result", "backward_hook_cam")

    classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
    fmap_block = list()
    grad_block = list()

    # 图片读取；网络加载
    img = cv2.imread(path_img, 1)  # H*W*C
    img_input = img_preprocess(img)
    net = Net()
    net.load_state_dict(torch.load(path_net))

    # 注册hook
    net.conv2.register_forward_hook(farward_hook)
    net.conv2.register_backward_hook(backward_hook)

    # forward
    output = net(img_input)
    idx = np.argmax(output.cpu().data.numpy())
    print("predict: {}".format(classes[idx]))

    # backward
    net.zero_grad()
    class_loss = comp_class_vec(output)
    class_loss.backward()

    # 生成cam
    grads_val = grad_block[0].cpu().data.numpy().squeeze()
    fmap = fmap_block[0].cpu().data.numpy().squeeze()
    cam = gen_cam(fmap, grads_val)

    # 保存cam图片
    img_show = np.float32(cv2.resize(img, (32, 32))) / 255
    show_cam_on_image(img_show, cam, output_dir)

---

总结

以上就是今天要讲的内容，本文仅仅简单介绍了通过实现Grad-CAM学习module中的forward_hook和backward_hook函数。可供学习参考~