Can't add graph with pytorch v0.2

[miguelvr]

I tried to run this snippet:

import torch
import torch.nn as nn
from torch.autograd import Variable
from datetime import datetime
from tensorboard import SummaryWriter


x = Variable(torch.rand(10, 10), requires_grad=True)

model = nn.Linear(10, 10)

h = model(x)

writer = SummaryWriter('runs/'+datetime.now().strftime('%B%d  %H:%M:%S'))

writer.add_graph(model, h)
writer.close()

but I get this error message:

Traceback (most recent call last):
  File "/Users/Miguel/Documents/Unbabel/pytorch-tools/tests/test_tensorboard.py", line 16, in <module>
    writer.add_graph(model, h)
  File "/Users/Miguel/anaconda/envs/venv/lib/python2.7/site-packages/tensorboard/writer.py", line 259, in add_graph
    self.file_writer.add_graph(graph(model, lastVar))
  File "/Users/Miguel/anaconda/envs/venv/lib/python2.7/site-packages/tensorboard/graph.py", line 39, in graph
    make_list_of_nodes(lastVar.grad_fn)
  File "/Users/Miguel/anaconda/envs/venv/lib/python2.7/site-packages/tensorboard/graph.py", line 23, in make_list_of_nodes
    inputs.append(make_name(next_fn))
  File "/Users/Miguel/anaconda/envs/venv/lib/python2.7/site-packages/tensorboard/graph.py", line 12, in make_name
    return id2name[id(obj.variable)]+'_'+str(id(obj.variable))
KeyError: 4654832816

[lanpa]

Passing the variable directly to the model solved the problem. But I don't know why. XD
Does this bug happen only in v0.2?

import torch
import torch.nn as nn
from torch.autograd import Variable
from datetime import datetime
from tensorboard import SummaryWriter


model = nn.Linear(10, 10)

h = model(Variable(torch.rand(10, 10), requires_grad=True))

writer = SummaryWriter('runs/'+datetime.now().strftime('%B%d  %H:%M:%S'))

writer.add_graph(model, h)
writer.close()

[miguelvr]

in theory it shouldn't be possible to add a graph in v0.1.12, right?

I tried with a nn.Sequential model and the result was the same.

[lanpa]

If your 0.1.12 is installed via pip then this function wont work. But if you build by yourself (0.1.12+xxxx) then it should work. Have you tried my solution?

[miguelvr]

your solution works, but that shouldn't make any difference lol

[lucabergamini]

I found something..
if you write

x = Variable(torch.rand(10, 10), requires_grad=True)

model = nn.Linear(10, 10)

h = model(x)

writer = SummaryWriter('runs/'+datetime.now().strftime('%B%d  %H:%M:%S'))

writer.add_graph(model,h)
writer.close()

and check using the debug the lastVar.grad_fn.next_functions you will find the weight,the bias (as expected) but also the input, filled with the variable field, holding the matrix.
If you write

#x = Variable(torch.rand(10, 10), requires_grad=True)

model = nn.Linear(10, 10)

#h = model(x)

writer = SummaryWriter('runs/'+datetime.now().strftime('%B%d  %H:%M:%S'))

writer.add_graph(model,model(Variable(torch.rand(10, 10), requires_grad=True)))
writer.close()

The input is still an element of the tuple, but it has a variable field empty.Since we create a binding dict id2name with the id of the Parameters of the network, the input is not included. As such, in the first case the input is still evaluated in the make_name function as it hold the variable field, but there is no association with its id, producing the above error.
An easy way to fix it would be to check before insert using the id from id2name in the function make_name

[zhangmozhe]

I update my pytorch to version 0.3, but I got the same error after my update. Previously there is no problem. Here is my code:

output_sample = model(Variable(torch.Tensor(1, 2, patch_size, patch_size).type(dtype), requires_grad=True), Variable(torch.Tensor(1, 1, patch_size, patch_size).type(dtype), requires_grad=True))
writer = SummaryWriter()
writer.add_graph(model, output_sample)

Even using the simple code I still get the error:

model = nn.Linear(10, 10)
h = model(Variable(torch.rand(10, 10), requires_grad=True))
writer = SummaryWriter()
writer.add_graph(model, h)
writer.close()

[lanpa]

That's expected behavior on v0.3. Please have a try on the onnx version: https://github.com/lanpa/tensorboard-pytorch/blob/master/onnx_graph.py
(you may need to install onnx as well)

[lanpa]

Since v0.2 is legacy, closing this.