1.6 在 GPU 上使用 DGLGraph

(中文版)

可以在构建时传入两个 GPU 张量来在 GPU 上创建一个 DGLGraph。另一种方法是使用 to() API 将 DGLGraph 复制到 GPU，这会将图结构和特征数据复制到指定的设备。

>>> import dgl
>>> import torch as th
>>> u, v = th.tensor([0, 1, 2]), th.tensor([2, 3, 4])
>>> g = dgl.graph((u, v))
>>> g.ndata['x'] = th.randn(5, 3)  # original feature is on CPU
>>> g.device
device(type='cpu')
>>> cuda_g = g.to('cuda:0')  # accepts any device objects from backend framework
>>> cuda_g.device
device(type='cuda', index=0)
>>> cuda_g.ndata['x'].device       # feature data is copied to GPU too
device(type='cuda', index=0)

>>> # A graph constructed from GPU tensors is also on GPU
>>> u, v = u.to('cuda:0'), v.to('cuda:0')
>>> g = dgl.graph((u, v))
>>> g.device
device(type='cuda', index=0)

任何涉及 GPU 图的操作都在 GPU 上执行。因此，它们要求所有张量参数都已放置在 GPU 上，并且结果（图或张量）也将位于 GPU 上。此外，GPU 图只接受位于 GPU 上的特征数据。

>>> cuda_g.in_degrees()
tensor([0, 0, 1, 1, 1], device='cuda:0')
>>> cuda_g.in_edges([2, 3, 4])   # ok for non-tensor type arguments
(tensor([0, 1, 2], device='cuda:0'), tensor([2, 3, 4], device='cuda:0'))
>>> cuda_g.in_edges(th.tensor([2, 3, 4]).to('cuda:0'))  # tensor type must be on GPU
(tensor([0, 1, 2], device='cuda:0'), tensor([2, 3, 4], device='cuda:0'))
>>> cuda_g.ndata['h'] = th.randn(5, 4)  # ERROR! feature must be on GPU too!
DGLError: Cannot assign node feature "h" on device cpu to a graph on device
cuda:0. Call DGLGraph.to() to copy the graph to the same device.