Closed sunshineatnoon closed 5 years ago
Tong-ZHAO
commented
5 years ago Hi, thanks for the feedback. Did you use evaluate.py to load the model? What's more if you load the model on CPU, please use: network.load_state_dict(torch.load(opt.modelPath, map_location='cpu')). Please tell me if it does not solve your problem.
vbcpascal
commented
5 years ago I have the same problem here. I seems that each key needed have been saved as key_1 and key_2.
xiayizhan2017
commented
5 years ago @Tong-ZHAO hi, I printed the key of log/2019-01-18T02:32:24.320546/network_4.pth， GCN_0.blocks.o.conv1.weight_1 GCN_0.blocks.o.conv1.weight_2 GCN_0.blocks.o.conv1.bias and I printed the key of model, GCN_0.blocks.o.conv1.weight GCN_0.blocks.o.conv1.bias It is different.
Tong-ZHAO
commented
5 years ago Hi all, thanks for your comments. It seems that I forgot to update the file gcn_layers.py in the last commit. I feel really sorry >.< I'm on holiday and I'll update the repo once I go back. For a quick hint, you can simply replace the function GraphConvolution by the following one and see what happens. Normally it should work.
class GraphConvolution(Module):
"""Simple GCN layer
Similar to https://arxiv.org/abs/1609.02907
"""
def __init__(self, in_features, out_features, adjs, bias=True, use_cuda = True):
super(GraphConvolution, self).__init__()
self.in_features = in_features
self.out_features = out_features
adj0 = torch_sparse_tensor(*adjs[0], use_cuda)
adj1 = torch_sparse_tensor(*adjs[1], use_cuda)
self.adjs = [adj0, adj1]
self.weight_1 = Parameter(torch.FloatTensor(in_features, out_features))
self.weight_2 = Parameter(torch.FloatTensor(in_features, out_features))
if bias:
self.bias = Parameter(torch.FloatTensor(out_features))
else:
self.register_parameter('bias', None)
self.reset_parameters()
def reset_parameters(self):
stdv = 1. / math.sqrt(self.weight.size(1))
self.weight_1.data.uniform_(-stdv, stdv)
self.weight_2.data.uniform_(-stdv, stdv)
if self.bias is not None:
self.bias.data.uniform_(-stdv, stdv)
def forward(self, input):
support_1 = torch.matmul(input, self.weight_1)
support_2 = torch.matmul(input, self.weight_2)
#output = torch.spmm(adj, support)
output1 = dot(self.adjs[0], support_1, True)
output2 = dot(self.adjs[1], support_2, True)
output = output1 + output2
if self.bias is not None:
return output + self.bias
else:
return output
def __repr__(self):
return self.__class__.__name__ + ' (' \
+ str(self.in_features) + ' -> ' \
+ str(self.out_features) + ')'
@Tong-ZHAO hi, thanks for your reply. when I changed the code 'stdv = 1. / math.sqrt(self.weight.size(1))' to 'stdv = 1. / math.sqrt(self.weight1.size(1))' , it solved the problem of load the trained model.
I generated my own data to test, but the results were not good. Hope that the author can provide the code to generate the data. Have a good holiday! >.<
Hi, Thanks for open-sourcing this awesome project. However, when I try to load the
log/2019-01-18T02:32:24.320546/network_4.pth, I got the following error: