The add_graph does not work about tensorboardx HOT 9 CLOSED

What is DenseNet3? Any error message?

from tensorboardx.

DenseNet3 is a class used to create the DenseNet, listed as follows; There are no error message, it just stops here, and does not go to train(cause train the net will print the process), i also opened the tensorboard, found nothing in it. sad....
class DenseNet3(nn.Module):
def init(self, depth=10, num_classes=10, growth_rate=12,
reduction=0.5, bottleneck=False, dropRate=0.2, inputdim=24):

from tensorboardx.

Even the example in https://github.com/lanpa/tensorboard-pytorch/blob/880fbaf6edd85b9e8ed1ab884daf1ed0b836c16c/demo_graph.py doesn't work. The error message was KeyError: 4646265864.
I am using tensorflow 1.4.1 and pytorch 0.3. Are they not supported yet?

from tensorboardx.

@akurniawan The current add_graph function only supports pytorch v0.2. For 0.3 please try the onnx version: https://github.com/lanpa/tensorboard-pytorch/blob/master/onnx_graph.py
(you may need to install onnx as well)

from tensorboardx.

The onnx_graph.py script linked above does not exist for me.

Could you clarify how to get a working graph visualization using pytorch 0.3?

from tensorboardx.

@adamtwig I believe this is the correct one https://github.com/lanpa/tensorboard-pytorch/blob/master/demo_graph.py

from tensorboardx.

@adamtwig @akurniawan
https://github.com/lanpa/tensorboard-pytorch/blob/9a3f9cca7057af515df64fdbfa523dbf16c1211d/onnx_graph.py

The onnx version don't have scope feature. So the graph is not pretty.
I you know how to build pytorch0.4, I strongly suggest you to use the new graph plotting function.

from tensorboardx.

@lanpa are you referring to this one? https://github.com/szagoruyko/functional-zoo/blob/master/visualize.py, so far that's the only thing that I can found and it's not official from pytorch. so I don't think pytorch 0.4 is necessary(?)

from tensorboardx.

@akurniawan

The internal graph representation are slightly different between pytorch 0.2, 0.3, 0.4
So tensorboard-pytorch uses different code to dump the graph.

To be more clear:

In 0.3, the easiest way is first dump the graph from pytorch to onnx, then dump from onnx.
(need to install onnx)
https://github.com/lanpa/tensorboard-pytorch/blob/9a3f9cca7057af515df64fdbfa523dbf16c1211d/onnx_graph.py

In 0.4(master branch), the network can be dumped to tensorboard directly without temporarily exporting to onnx. And I think the best thing is that scope is supported in this version :)
https://github.com/lanpa/tensorboard-pytorch/blob/master/demo_graph.py

Btw, the visualize.py should work for v0.2

You can see the result from the newest version:
http://35.197.26.245:6006/

from tensorboardx.