Asking for a problem related to data transform which I discovered in dgl example of graphsage

[yichuan520030910320]

In dgl/node_classification.py at master · dmlc/dgl · GitHub
I have some problem about the code, if we use

train_idx = dataset.train_idx.to(device)
val_idx = dataset.val_idx.to(device)

before trainig (device is GPU) . If that means the GPU can accommodate all of the dataset. so I think it is unnecessary for us to use minibatch later

for it, (input_nodes, output_nodes, blocks) in enumerate(train_dataloader):

becuase our GPU is large enough

I have this problem because I read the passage in
https://docs.dgl.ai/guide_cn/minibatch-node.html#guide-cn-minibatch-node-classification-sampler

and find there is a line of code

blocks = [b.to(torch.device(‘cuda’)) for b in blocks]

in the loop body of minibatch

so I wonder which one is correct

because I want to find an example that cater to the process of

1. take a mini-batch of training nodes and sample a subgraph from these nodes
2. gather node / edge features used by the subgraph from CPU memory and
   copy into GPU
3. GNN model forward & backward & weight updates

also If anyone gan give me an example of a totally large dataset that A100 can not accommodate, thanks a lot

[dyru]

1. For the first question, train_idx and val_idx are just node indices. Full-graph training requires the storage of features and intermediate hidden states, which consume much larger memory. Besides, even if your GPU capacity is enough for full-graph training, mini-batch training is still worthy of attempts. The stochastic training may bring faster convergence and better performance.

2. Please find some large-scale datasets here.

[yichuan520030910320]

Thanks for your reply!