0

将 GNN 与 HetreoGraph 一起使用是否正确？

运行代码几次会报错，其他次都正常。

重现

多次运行代码会报错：DGLError: Expected data to have X rows, got Y。

import numpy as np

import torch
import torch.optim as optim
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.data as Data

import dgl
from dgl import DGLGraph
import dgl.nn as dglnn

class RGNN(nn.Module):
    def __init__(self, in_size, hid_size, out_size, rel_names):
        super().__init__()
        self.conv1 =  dglnn.HeteroGraphConv({rel : dglnn.GraphConv(in_size, hid_size) \
                                             for rel in rel_names}, aggregate="sum")
        self.conv2 =  dglnn.HeteroGraphConv({rel : dglnn.GraphConv(hid_size, out_size) \
                                             for rel in rel_names}, aggregate="sum")
        self.conv3 =  dglnn.HeteroGraphConv({rel : dglnn.GraphConv(out_size, out_size) \
                                             for rel in rel_names}, aggregate="sum")
        self.dropout=nn.Dropout(p=0.2)
        
    def forward(self, blocks, in_feats):
        print("before conv1", in_feats["user"].shape, in_feats["game"].shape)
        h = self.conv1(blocks[0], in_feats)
        print("after conv1", h["user"].shape, h["game"].shape)
        h = self.conv2(blocks[1], h)
        print("after conv2", h["user"].shape, h["game"].shape)
        h = self.conv3(blocks[2], h)
        print("conv3", h["user"].shape, h["game"].shape)
        return h

    
hetero_frontier = dgl.heterograph({
    ('user', 'follow', 'user'): ([1, 3, 7,1,2,3,4,5,6,7,8,9], [3, 6, 8,4,5,6,7,3,2,5,8,5]),
    ('user', 'play', 'game'): ([5, 5, 4] + list(range(10)), [6, 6, 2] + list(np.random.randint(0, 10,(1,10)).squeeze()))},
    num_nodes_dict={'user': 10, 'game': 10})

embed_size = 128
hetero_frontier.nodes['user'].data['embedding'] = torch.FloatTensor(np.random.normal(0, 0.01, (10, embed_size)))
hetero_frontier.nodes['game'].data['embedding'] = torch.FloatTensor(np.random.normal(0, 0.01, (10, embed_size)))

model = RGNN(embed_size, embed_size, embed_size, hetero_frontier.etypes)  

train_nid_dict = {"user":list(np.random.randint(0,10,(1,5)).squeeze()),"game":np.random.randint(10)}
sampler = dgl.dataloading.MultiLayerFullNeighborSampler(3)
dataloader = dgl.dataloading.NodeDataLoader(hetero_frontier, train_nid_dict, sampler,
    batch_size=128,
    shuffle=False,
    drop_last=False,
    num_workers=1)
for input_nodes, output_nodes, blocks in dataloader:
    print("input nodes", input_nodes["user"].shape,input_nodes["game"].shape)
    for idx, block in enumerate(blocks):
        print("block ",idx,block.srcdata["embedding"]["user"].shape,block.srcdata["embedding"]["game"].shape)
    print("ouput nodes", output_nodes["user"].shape,output_nodes["game"].shape)

    input_features = blocks[0].srcdata["embedding"]     # returns a dict     
    output_features = model(blocks, input_features)

DGLError: Expected data to have 5 rows, got 4. 但是我遵循指南

预期行为

无尺寸错误

环境

  • DGL 版本 0.6.1
  • PyTorch 1.8.1
  • 苹果系统：
  • 你是如何安装 DGL 的：conda
  • Python版本：3.7
  • CUDA/cuDNN 版本（如果适用）：仅 CPU
4

0 回答 0