[Tensorboard] Result of the graphs are different across pytorch versions

[as754770178]

import torch
from torchvision.models import resnet50
from torch.utils.tensorboard import SummaryWriter


model = resnet50()
writer = SummaryWriter("/home/tensorboard_test")
board_input = torch.rand((3, 224, 224)).unsqueeze(0)
writer.add_graph(model, board_input)
writer.close()

I found that SummaryWriter in different version pytorch, I result is different.
In torch 1.1.0, the result is OK.
In torch 1.2.0, the program not report error, but no graph display in tensorboard.
In torch 1.3.0, the graph display in tensorboard, but the graph different between 1.1.0 and 1.3.0.
The graph, in torch 1.3.0, is very difficult to use.

I suggest revert SummaryWriter to torch 1.1.0.

[as754770178]

I once submmit Issue #24904, but don't deal with.

[orionr]

cc @sanekmelnikov, @J0Nreynolds, @lanpa

[orionr]

@as754770178, can you do us a favor and try the version in nightly and see if it improves things? We had a regression in 1.3, but should be greatly improved for 1.4.

[as754770178]

I try 1.4, the graph style is same with 1.3. The network show in horizontal, it's very hard to use.

[orionr]

@as754770178, sorry to hear that. @sanekmelnikov, @J0Nreynolds and @lanpa can you debug for the ResNet-50 architecture? Might not make it for 1.4, but we should try and improve for 1.5 at a minimum.

[as754770178]

I can't upload the tensorboard graph as a result of network, I describe it. In torch 1.1.0, I open the graph resnet50, the node is common operation such as conv, bn, etc. Now the node is detailed weight, and the graph is very long in horizontal.
I think the format in torch 1.1.0 is good, because I only want to know the network architecture. If user want to comprehend in detail, they open the node to view, like the way in tensorflow.

[lanpa]

@as754770178 @orionr This issue should already have been solved before 1.4.0a0+65f6e44 (a month ago)

And here is the result of this version. Would you like to try to build from master or use pytorch-nightly again?

[as754770178]

@lanpa Thanks. The effect is very good.

[J0Nreynolds]

Hi @as754770178. Did the suggestion from @lanpa fix your issue?

[as754770178]

my torch version is 1.4.0a0+2c51e06, the result still like before that is different with @lanpa

[as754770178]

tensorboard version is 1.14.0

[lanpa]

Just tested with torch 1.4.0 from PyPI and the result is good too. Closing this.