0、参考
https://oldpan.me/archives/pytorch-autograd-hook
https://pytorch.org/docs/stable/search.html?q=hook&check_keywords=yes&area=default
https://github.com/pytorch/pytorch/issues/598
https://github.com/sksq96/pytorch-summary
https://github.com/allensll/test/blob/591c7ce3671dbd9687b3e84e1628492f24116dd9/net_analysis/viz_lenet.py
1、背景
在神经网络的反向传播当中个，流程只保存叶子节点的梯度，对于中间变量的梯度没有进行保存。
import torch
x = torch.tensor([1,2],dtype=torch.float32,requires_grad=True)
y = x+2
z = torch.mean(torch.pow(y, 2))
lr = 1e-3
z.backward()
x.data -= lr*x.grad.data
print(y.grad)此时输出就是：None，这个时候hook的作用就派上，hook可以通过自定义一些函数，从而完成中间变量的输出，比如中间特征图、中间层梯度修正等。
在pytorch docs搜索hook，可以发现有四个hook相关的函数，分别为register_hook，register_backward_hook，register_forward_hook，register_forward_pre_hook。其中register_hook属于tensor类，而后面三个属于moudule类。
- register_hook函数属于torch.tensor类，函数在tensor梯度计算的时候就会执行，这个函数主要处理梯度相关的数据，表现形式$hook(grad) rightarrow Tensor or None$.
import torch
x = torch.tensor([1,2],dtype=torch.float32,requires_grad=True)
y = x * 2
y.register_hook(print)
<torch.utils.hooks.RemovableHandle at 0x7f765e876f60>
z = torch.mean(y)
z.backward()
tensor([ 0.5000, 0.5000])- Register_backward_hook等三个属于torch.nn，属于moudule中的方法。
hook(module, grad_input, grad_output) -> Tensor or None写个demo，参考:
下面的计算为
import torch
import torch.nn as nn
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")def print_hook(grad):print ("register hook:", grad)return gradclass TestNet(nn.Module):def __init__(self):super(TestNet, self).__init__()self.f1 = nn.Linear(4, 1, bias=True)self.weights_init()def weights_init(self):self.f1.weight.data.fill_(4)self.f1.bias.data.fill_(0.1)def forward(self, input):self.input = inputout = input * 0.75out = self.f1(out)out = out / 4return outdef back_hook(self, moudle, grad_input, grad_output):print ("back hook in:", grad_input)print ("back hook out:", grad_output)# 修改梯度# grad_input = list(grad_input)# grad_input[0] = grad_input[0] * 100# print (grad_input)return tuple(grad_input)if __name__ == '__main__':input = torch.tensor([1, 2, 3, 4], dtype=torch.float32, requires_grad=True).to(device)net = TestNet()net.to(device)net.register_backward_hook(net.back_hook)ret = net(input)print ("result", ret)ret.backward()print('input.grad:', input.grad)for param in net.parameters():print('{}:grad->{}'.format(param, param.grad))输出：
result tensor([7.5250], grad_fn=<DivBackward0>)
back hook in: (tensor([0.2500]), None)
back hook out: (tensor([1.]),)
input.grad: tensor([0.7500, 0.7500, 0.7500, 0.7500])
Parameter containing:
tensor([[4., 4., 4., 4.]], requires_grad=True):grad->tensor([[0.1875, 0.3750, 0.5625, 0.7500]])
Parameter containing:
tensor([0.1000], requires_grad=True):grad->tensor([0.2500])输出结果以及梯度都很明显，简单分析一下w权重的梯度，
另外，hook中有个bug，假设我们bug，假设我们注释掉out = out / 4这行，可以发现输出变成back hook in: (tensor([1.]), tensor([1.]))。这种情况就不符合上面我们的梯度计算公式，是因为这个时候：
则此时的偏导只是对
- register_forward_hook跟Register_backward_hook差不多，就不过多复述。
- register_forward_pre_hook，可以发现其输入只有
hook(module, input) -> None
其主要是针对推理时的hook.
2、应用
2.1 特征图打印
直接利用pytorch已有的resnet18进行特征图打印，只打印卷积层的特征图，
import torch
from torchvision.models import resnet18
import torch.nn as nn
from torchvision import transformsimport matplotlib.pyplot as pltdef viz(module, input):x = input[0][0]#最多显示4张图min_num = np.minimum(4, x.size()[0])for i in range(min_num):plt.subplot(1, 4, i+1)plt.imshow(x[i].cpu())plt.show()import cv2
import numpy as np
def main():t = transforms.Compose([transforms.ToPILImage(),transforms.Resize((224, 224)),transforms.ToTensor(),transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])])device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")model = resnet18(pretrained=True).to(device)for name, m in model.named_modules():# if not isinstance(m, torch.nn.ModuleList) and # not isinstance(m, torch.nn.Sequential) and # type(m) in torch.nn.__dict__.values():# 这里只对卷积层的feature map进行显示if isinstance(m, torch.nn.Conv2d):m.register_forward_pre_hook(viz)img = cv2.imread('./cat.jpeg')img = t(img).unsqueeze(0).to(device)with torch.no_grad():model(img)if __name__ == '__main__':main()直接放几张中间层的图
2.2 模型大小，算力计算
同样的用法，可以直接参考pytorch-summary这个项目。
