Hi.I have a problem when I execute autograd in a GNN model build with DGL.
import dgl
import dgl.function as fn
import torch
from torch import nn
import numpy as np
src = [0, 0, 0, 0, 0, 1, 2, 3, 3, 4, 5, 6, 7, 7, 8, 9]
dst = [0, 1, 2, 3, 9, 6, 6, 3, 5, 7, 4, 9, 0, 1, 1, 2]
u = np.concatenate([src, dst])
v = np.concatenate([dst, src])
G = dgl.DGLGraph((u, v))
G.ndata['x'] = torch.randn(10, 4)
G.ndata['d'] = torch.randn(10, 4)
class net(nn.Module):
def __init__(self):
super(net, self).__init__()
self.lin = nn.Linear(4, 1)
def message_func(self, edges):
feat = edges.dst['x'].to(device)
message = self.lin(feat)
return {'m': message}
def forward(self, G, feat):
graph = G.local_var()
graph.srcdata['h'] = feat
graph.update_all(self.message_func, fn.sum(msg='m', out='h'))
rst = graph.dstdata['h']
out = torch.sum(rst)
d = - torch.autograd.grad(out.sum(), feat, create_graph=True)[0]
return out, d
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = net().to(device)
feat = G.ndata['x'].to(device)
feat.requires_grad = True
out, d = model(G, feat)
I want to calculate the derivative of out to the input,i.e.G.ndata['x'].If torch.device('cpu'),the code could run.However I get the error when torch.device('cuda')
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-4-e56cb4569601> in <module>
4 feat = G.ndata['x'].to(device)
5 feat.requires_grad = True
----> 6 model(G, feat)
~/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
530 result = self._slow_forward(*input, **kwargs)
531 else:
--> 532 result = self.forward(*input, **kwargs)
533 for hook in self._forward_hooks.values():
534 hook_result = hook(self, input, result)
<ipython-input-2-6274bd0ee471> in forward(self, G, feat)
23 rst = graph.dstdata['h']
24 out = torch.sum(rst)
---> 25 d = - torch.autograd.grad(out.sum(), feat, create_graph=True)[0]
26 return out, d
~/anaconda3/lib/python3.6/site-packages/torch/autograd/__init__.py in grad(outputs, inputs, grad_outputs, retain_graph, create_graph, only_inputs, allow_unused)
155 return Variable._execution_engine.run_backward(
156 outputs, grad_outputs, retain_graph, create_graph,
--> 157 inputs, allow_unused)
158
159
RuntimeError: One of the differentiated Tensors appears to not have been used in the graph. Set allow_unused=True if this is the desired behavior.
If I set allow_unused=True,the derivative will be None.It usually happen when the feat isn’t a leaf,However if I print feat.is_leaf,it returns True.
I want to reproduce this error without DGL but only pytorch,but I can’t.
import torch
from torch import nn
class net(nn.Module):
def __init__(self):
super(net, self).__init__()
self.lin = nn.Linear(4, 1)
def forward(self, x):
out = self.lin(x)
d = torch.autograd.grad(out.sum(), x, create_graph=True)
return out, d
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = net().to(device)
x = torch.randn(10, 4).to(device)
x.requires_grad = True
out, d = model(x)
which could run without any error.
So what’s wrong with the first code?Thank you.