Tensorboardx: add_graph() show error 'torch._C.Value' object has no attribute 'uniqueName' with torch 1.2.0

[GinSoda]

Describe the bug
add_graph() with torch 1.2.0 will show error 'torch._C.Value' object has no attribute 'uniqueName'
but work well with torch 1.1.0

Minimal runnable code to reproduce the behavior

from tensorboardX import SummaryWriter
class LeNet(nn.Module):
    def __init__(self):
        super(LeNet, self).__init__()
        self.conv1 = nn.Sequential(     #input_size=(1*28*28)
            nn.Conv2d(1, 6, 5, 1, 2),
            nn.ReLU(),      #(6*28*28)
            nn.MaxPool2d(kernel_size=2, stride=2),  #output_size=(6*14*14)
        )
        self.conv2 = nn.Sequential(
            nn.Conv2d(6, 16, 5),
            nn.ReLU(),      #(16*10*10)
            nn.MaxPool2d(2, 2)  #output_size=(16*5*5)
        )
        self.fc1 = nn.Sequential(
            nn.Linear(16 * 5 * 5, 120),
            nn.ReLU()
        )
        self.fc2 = nn.Sequential(
            nn.Linear(120, 84),
            nn.ReLU()
        )
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = x.view(x.size()[0], -1)
        x = self.fc1(x)
        x = self.fc2(x)
        x = self.fc3(x)
        return x

dummy_input = torch.rand(13, 1, 28, 28)
model = LeNet()
with SummaryWriter(comment='Net', log_dir='/output') as w:
    w.add_graph(model, (dummy_input, ))

Expected behavior

Screenshots

Environment

Python environment
Python 3.6

Additional context
Add any other context about the problem here.

[YYRancho]

Same error for me with torch 1.2.0 and Python 3.5. BTW, the same code does not produce errors with the official torch.utils.tensorboard module but tensorboard does not show any graphs.

[Arideno]

@YYRancho The same problem.

[barryflower]

Also the example code provided in the repo fails in the same way.

Relevant packages are:

• tb-nightly 1.15.0a20190806
• tensorboardX 1.8+366bc8f
• torch 1.2.0
• torchvision 0.4.0

~/tmp/tensorboardX$ python3 examples/demo_graph.py
graph(%self : ClassType,
%input : Float(1, 3)):
%1 : ClassType = prim::GetAttrname="l"
%weight : Tensor = prim::GetAttrname="weight"
%bias : Tensor = prim::GetAttrname="bias"
%5 : Float(3!, 5!) = aten::t(%weight), scope: LinearInLinear/Linear[l] # /home/barryf/.local/lib/python3.6/site-packages/torch/nn/functional.py:1369:0
%6 : int = prim::Constantvalue=1, scope: LinearInLinear/Linear[l] # /home/barryf/.local/lib/python3.6/site-packages/torch/nn/functional.py:1369:0
%7 : int = prim::Constantvalue=1, scope: LinearInLinear/Linear[l] # /home/barryf/.local/lib/python3.6/site-packages/torch/nn/functional.py:1369:0
%8 : Float(1, 5) = aten::addmm(%bias, %input, %5, %6, %7), scope: LinearInLinear/Linear[l] # /home/barryf/.local/lib/python3.6/site-packages/torch/nn/functional.py:1369:0
return (%8)

Traceback (most recent call last):
File "examples/demo_graph.py", line 20, in
w.add_graph(LinearInLinear(), dummy_input, True)
File "/home/barryf/tmp/tensorboardX/examples/tensorboardX/writer.py", line 774, in add_graph
self._get_file_writer().add_graph(graph(model, input_to_model, verbose, kwargs))
File "/home/barryf/tmp/tensorboardX/examples/tensorboardX/pytorch_graph.py", line 261, in graph
list_of_nodes, node_stats = parse(graph, args)
File "/home/barryf/tmp/tensorboardX/examples/tensorboardX/pytorch_graph.py", line 218, in parse
nodes_py.append(NodePyIO(node, 'input'))
File "/home/barryf/tmp/tensorboardX/examples/tensorboardX/pytorch_graph.py", line 68, in __init__
super(NodePyIO, self).__init__(node_cpp, methods_IO)
File "/home/barryf/tmp/tensorboardX/examples/tensorboardX/pytorch_graph.py", line 63, in __init__
setattr(self, m, getattr(node_cpp, m)())
**AttributeError: 'torch._C.Value' object has no attribute 'uniqueName'

[ShirAmir]

@YYRancho same here with python=3.6.8, torch=1.2.0

[brickmaker]

same with python=3.7.4, torch=1.2.0

[kice]

I assume this commit break this
https://github.com/pytorch/pytorch/commit/f7b2778cb1db148f1c901ea08abcbfa0e0a78f44#diff-284de3c993add230888834d04a67d666

Now is debugName instead of uniqueName
https://github.com/pytorch/pytorch/blob/master/torch/utils/tensorboard/_pytorch_graph.py#L19

[lanpa]

@YYRancho Does the resulting graph looks OK with the official torch.utils.tensorboard? I fixed the corresponding code in tensorboardX and found the result looks bad.

[YYRancho]

@lanpa No, tensorboard shows no graphs but two black-edged boxes in the lower right corner. I have tried simple ANN and LSTM and got the same results. I would be glad to provide more information if necessary.

[lxb6476]

the same problem with python=3.7.4, torch=1.2.0.

[lanpa]

@YYRancho @a-pythoner Thanks for the report. Reopen this.

[lanpa]

@YYRancho btw, does the error torch._C.Value still happen?
@a-pythoner What is your result of pip list?
Thanks

[YYRancho]

@lanpa tensorboardX will not show any error if I follow your fix.

[amineHY]

I am encountering the same issue here:
Any solution ?

My installations:
torch 1.2.0+cu92
tensorboardX 1.8
python 3.6

Thank you in advance.

[MaorEi]

I am encountering the same issue when calling hiddenlayer.build_graph method:
In order to fix it in hiddenlayer\pytorch_builder.py, change get_shape method to:

def pytorch_id(node):
    """Returns a unique ID for a node."""
    # After ONNX simplification, the scopeName is not unique anymore
    # so append node outputs to guarantee uniqueness
    return node.scopeName() + "/outputs/" + "/".join([o.debugName() for o in node.outputs()])

Packages information:
torch 1.2.0+cu10
matplotlib 3.1.1
tensorflow 1.14.0
graphviz 0.12
hiddenlayer 0.2
python 3.7.4

[vinthony]

try to build from source with :

git clone https://github.com/lanpa/tensorboardX && cd tensorboardX && python setup.py install

[chengfeng0714]

try to build from source with :

git clone https://github.com/lanpa/tensorboardX && cd tensorboardX && python setup.py install

It's great. It's perfect for solving problems.

[lanpa]

Since pytorch 1.3 is officially released and works with tensorboardX 1.9, closing this.

[cswwp]

Since pytorch 1.3 is officially released and works with tensorboardX 1.9, closing this.
add_graph() still error with pytorch 1.1.0 tensorboardx 1.9, error info:
TypeError: __init__() got an unexpected keyword argument 'record_shapes' @lanpa