如何通过 Networkx 图传播神经网络梯度？

问题描述

我正在尝试训练我的神经网络，使图中每条边的权重都为 10。我从生成随机点 (inp) 开始，并使每 2 个相邻点（使用 idx）具有 weight = 1 的边缘。然后如果 2 个相邻的点已经有一条边，边的权重就会被发送到神经网络，神经网络输出要添加的额外权重。

import warnings
warnings.filterwarnings("ignore",category=UserWarning)

import torch
import torch.nn as nn
import networkx as nx
import matplotlib.pyplot as plt
import torch.optim as optim

class Net(nn.Module):
    def __init__(self):
        super(Net,self).__init__()
        self.fc1 = nn.Linear(1,3)
        self.fc2 = nn.Linear(3,1)

    def forward(self,x):
        x = self.fc1(x)
        x = self.fc2(x)
        return x

g = nx.DiGraph() 
model = Net()
optimizer = optim.Adam(model.parameters(),lr = 1e-3)

def training(n_iter):
    for epoch in range(n_iter):
        print(epoch)
        inp = torch.randint(0,10,(20,))
        idx = 0
        while idx < len (inp) - 1:
            if  g.has_edge(inp[idx].item(),inp[idx+1].item()): #edge exist

                edge_weight = g[inp[idx].item()][inp[idx+1].item()]["weight"]
                edge_weight_tensor = torch.tensor([edge_weight]).float() #to tensor
                
                added_edge_weight = model(edge_weight_tensor) #value from network
                g[inp[idx].item()][inp[idx+1].item()]["weight"] += added_edge_weight
                idx +=1
            else:
                g.add_edge(inp[idx].item(),inp[idx+1].item(),weight = 1)
                idx +=1

        edges = g.edges()
        weights = [g[u][v]['weight'] for u,v in edges]
        optimizer.zero_grad()

        loss_list = [w for w in weights if not isinstance(w,int)] #only take tensors
        try:
            loss_tensors = (torch.stack(loss_list,dim=0)-10) 
            loss_square = torch.square(loss_tensors)
            loss = torch.sum(loss_square)
            print(loss)
        except RuntimeError: #no tensors - hence create a 0 loss
            loss = torch.tensor(0.0,requires_grad = True)

        loss.backward(retain_graph=True)
        optimizer.step()
    return weights

weights = training(5)

#plot
plt.figure(figsize=(6,6))
pos = nx.spring_layout(g,k = 0.5) 

nx.draw(g,with_labels=True,node_color='skyblue',font_weight='bold',width=weights,pos=pos)

我的问题是，我不确定梯度是否可以通过这种方式传播，而且我似乎无法将 NN 权重添加到边缘的权重中——出现以下错误：

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [3,1]],which is output 0 of TBackward,is at version 2; expected version 1 instead. Hint: enable anomaly detection to find the operation that Failed to compute its gradient,with torch.autograd.set_detect_anomaly(True).

解决方法

PyTorch Geometric 专门设计用于在图形对象上实现 PyTorch 方法。具体来说，它的 https://github.com/grails/grails-core/issues/11221 函数允许从 networkx 图转换为可以传播梯度的 PyTorch Geometric 对象。

networkx python pytorch