Problems about train_sampling_multi_gpu.py

CapsulE07 · August 3, 2020, 1:08pm

after the for step, (input_nodes, seeds, blocks) in enumerate(dataloader):, we get the blocks which representaing the graph structure of every layer.
And then through batch_inputs, batch_labels = load_subtensor(train_g, train_g.ndata[‘labels’], seeds, input_nodes, dev_id) , we could get the node features and labels from the original graph.
But how could we get the edge data of the blocks?
I read through the source code of it, and only find that the block.edata is only a list of indexs.

github.com

dmlc/dgl/blob/562871e76b2d5adbcd907bcf62ca28c0611e50e1/examples/pytorch/graphsage/train_sampling.py

import dgl
import numpy as np
import torch as th
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.multiprocessing as mp
from torch.utils.data import DataLoader
import dgl.function as fn
import dgl.nn.pytorch as dglnn
import time
import argparse
from _thread import start_new_thread
from functools import wraps
from dgl.data import RedditDataset
import tqdm
import traceback

from load_graph import load_reddit, load_ogb, inductive_split

This file has been truncated. show original

BarclayII · August 4, 2020, 1:54am

The indices are the edge IDs in the original graph. For now you need to manually read it:

block.edata['feature'] = g.edata['feature'][block.edata[dgl.EID]]

CapsulE07 · August 4, 2020, 4:48pm

Thanks for your quick response!