pytorch的hook函数怎么使用

"""
@brief      : pytorch的hook函数
"""

import torch
import torch.nn as nn
from tools.common_tools2 import set_seed

set_seed(1)

# ----------------------------------- 1 tensor hook 1
flag = 0
# flag = 1
if flag:
    w = torch.tensor([1.], requires_grad=True)
    x = torch.tensor([2.], requires_grad=True)
    a = torch.add(w, x)
    b = torch.add(w, 1)
    y = torch.mul(a, b)

    a_grad = list()


    def grad_hook(grad):
        a_grad.append(grad)


    handle = a.register_hook(grad_hook)

    y.backward()

    # 查看梯度
    print("gradient:", w.grad, x.grad, a.grad, b.grad, y.grad)
    print("a_grad[0]:", a_grad[0])

    handle.remove()

# ----------------------------------- 2 tensor hook 2
flag = 0
# flag = 1
if flag:
    w = torch.tensor([1.], requires_grad=True)
    x = torch.tensor([2.], requires_grad=True)
    a = torch.add(w, x)
    b = torch.add(w, 1)
    y = torch.mul(a, b)

    a_grad = list()


    def grad_hook(grad):
        grad *= 2
        return grad * 3


    handle = w.register_hook(grad_hook)

    y.backward()

    print("w.grad:", w.grad)
    handle.remove()

# --------------------------- 3 Module.register_forward_hook and pre hook
# flag = 0
flag = 1
if flag:
    class Net(nn.Module):
        def __init__(self):
            super(Net, self).__init__()
            self.conv1 = nn.Conv2d(1, 2, 3)
            self.pool1 = nn.MaxPool2d(2, 2)

        def forward(self, x):
            x = self.conv1(x)
            x = self.pool1(x)
            return x


    def forward_hook(module, data_input, data_output):
        fmap_block.append(data_output)
        input_block.append(data_input)


    def forward_pre_hook(module, data_input):
        print("forward_pre_hook input:{}".format(data_input))


    def backward_hook(module, grad_input, grad_output):
        print("backward hook input:{}".format(grad_input))
        print("backward hook output:{}".format(grad_output))


    # 初始化网络
    net = Net()
    net.conv1.weight[0].detach().fill_(1)
    net.conv1.weight[1].detach().fill_(2)
    net.conv1.bias.data.detach().zero_()

    # 注册hook
    fmap_block = list()
    input_block = list()
    net.conv1.register_forward_hook(forward_hook)
    net.conv1.register_forward_pre_hook(forward_pre_hook)
    net.conv1.register_backward_hook(backward_hook)

    # inference
    fake_img = torch.ones((1, 1, 4, 4))  # batch size * channel * H * W
    output = net(fake_img)  # 前向传播

    loss_fnc = nn.L1Loss()
    target = torch.randn_like(output)
    loss = loss_fnc(target, output)
    loss.backward()

    # 观察
    print("output shape: {}\noutput value: {}\n".format(output.shape, output))
    print("feature maps shape: {}\noutput value: {}\n".format(fmap_block[0].shape, fmap_block[0]))
    print("input shape: {}\ninput value: {}".format(input_block[0][0].shape, input_block[0]))

# -*- coding:utf-8 -*-
"""
@brief      : 采用hook函数可视化特征图
"""

import torch.nn as nn
import numpy as np
from PIL import Image
import torchvision.transforms as transforms
import torchvision.utils as vutils
from torch.utils.tensorboard import SummaryWriter
from tools.common_tools2 import set_seed
import torchvision.models as models

set_seed(1)  # 设置随机种子

# ----------------------------------- feature map visualization
# flag = 0
flag = 1
if flag:
    writer = SummaryWriter(comment='test_your_comment', filename_suffix="_test_your_filename_suffix")

    # 数据
    path_img = "./lena.png"  # your path to image
    normMean = [0.49139968, 0.48215827, 0.44653124]
    normStd = [0.24703233, 0.24348505, 0.26158768]

    norm_transform = transforms.Normalize(normMean, normStd)
    img_transforms = transforms.Compose([
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        norm_transform
    ])

    img_pil = Image.open(path_img).convert('RGB')
    if img_transforms is not None:
        img_tensor = img_transforms(img_pil)
    img_tensor.unsqueeze_(0)  # chw --> bchw

    # 模型
    alexnet = models.alexnet(pretrained=True)

    # 注册hook
    fmap_dict = dict()
    for name, sub_module in alexnet.named_modules():

        if isinstance(sub_module, nn.Conv2d):
            key_name = str(sub_module.weight.shape)
            fmap_dict.setdefault(key_name, list())

            n1, n2 = name.split(".")


            def hook_func(m, i, o):
                key_name = str(m.weight.shape)
                fmap_dict[key_name].append(o)


            alexnet._modules[n1]._modules[n2].register_forward_hook(hook_func)

    # forward
    output = alexnet(img_tensor)

    # add image
    for layer_name, fmap_list in fmap_dict.items():
        fmap = fmap_list[0]
        fmap.transpose_(0, 1)

        nrow = int(np.sqrt(fmap.shape[0]))
        fmap_grid = vutils.make_grid(fmap, normalize=True, scale_each=True, nrow=nrow)
        writer.add_image('feature map in {}'.format(layer_name), fmap_grid, global_step=322)