unable to visualize the network with pydot

[twmht]

Hi,

I want to visualize FPN network with pydot. I add the rendering process after creating the network.

def create_model():
    """Build the model and look for saved model checkpoints in case we can
    resume from one.
    """
    # skip many lines
    logger.info('Building model: {}'.format(cfg.MODEL.TYPE))
    model = model_builder.create(cfg.MODEL.TYPE, train=True)

    # rendering the graph
    graph = net_drawer.GetPydotGraph(model.net)
    graph.write_png('network.png')

    return model, weights_file, start_iter, checkpoints, output_dir

but pydot get stuck at write_png without throwing errors.

any idea?

thank you.

[ir413]

Hi @twmht, net_drawer.GetPydotGraph may not be getting stuck but just taking a very long time.
You could try using net_drawer.GetPydotGraphMinimal which is faster but only draws the op nodes.

[twmht]

@ir413

even though switching to net_drawer.GetPydotGraphMinimal, it still took very long time to complete. I have been waiting for 30 minutes and it still does not finish.

[twmht]

after ctrl-C, it shows up where it stuck with

File "tools/train_net.py", line 129, in <module>
    main()
  File "tools/train_net.py", line 111, in main
    checkpoints = utils.train.train_model()
  File "/home/tumh/detectron/lib/utils/train.py", line 54, in train_model
    model, weights_file, start_iter, checkpoints, output_dir = create_model()
  File "/home/tumh/detectron/lib/utils/train.py", line 135, in create_model
    graph.write_png('network.png')
  File "/home/tumh/.pyenv/versions/faster-rcnn/lib/python2.7/site-packages/pydot.py", line 1691, in <lambda>
    self.write(path, format=f, prog=prog))
  File "/home/tumh/.pyenv/versions/faster-rcnn/lib/python2.7/site-packages/pydot.py", line 1774, in write
    s = self.create(prog, format)
  File "/home/tumh/.pyenv/versions/faster-rcnn/lib/python2.7/site-packages/pydot.py", line 1886, in create
    stdout_data, stderr_data = p.communicate()
  File "/home/tumh/.pyenv/versions/2.7.12/lib/python2.7/subprocess.py", line 800, in communicate
    return self._communicate(input)
  File "/home/tumh/.pyenv/versions/2.7.12/lib/python2.7/subprocess.py", line 1417, in _communicate
    stdout, stderr = self._communicate_with_poll(input)
  File "/home/tumh/.pyenv/versions/2.7.12/lib/python2.7/subprocess.py", line 1471, in _communicate_with_poll
    ready = poller.poll()

[twmht]

I write the raw data by graph.write("network.txt"), it contains 1550 nodes. I am not sure if this is too big for pydot.

However, even though all of these nodes can be visualized with png, it's still hard to visualize with human eyes.

It would be great if it only provides layer-level virtualization, with some visualization tools like netscope (http://ethereon.github.io/netscope/quickstart.html).

And I think this should be a caffe2 issue, not detectron.

[liujing1995]

Have you solved the question? I have the same problem.

[twmht]

@liujing1995

no, I have not solved this problem.

it seems that it's not an issue of caffe2. it's an issue of pydot.

[liujing1995]

@twmht
I found someone could visualize the network with pydot during test time.But I can't use it to visualize the training network #320

[twmht]

ok. reopen this issue to discuss more.

[drcege]

#508 works well for inference model. The problems are all about training model.

---

Even the GetPydotGraphMinimal is too large for graphviz.

After many tries, I still cannot draw a graph. I could only save the PydotGraph to a raw .dot file. When converting the dot file to pdf / png / svg by graphviz(dot) , it hangs just like this question.

neato and twopi may give results, but the nodes are overlapped so unreadable.