Does PyTorch have any tool,something like TensorBoard in TensorFlow,to do graph visualization to help users understand and debug network?
Not that I am aware of. However, due to its dynamic nature, it is much easier to debug a network in pytorch than tensorflow.
As one commenter on Reddit opines:
“Debugging is easier because a specific line in your specific code (not something deep under your sess.run() that worked with a large/generated Graph object) fails. Your stack traces don’t fill up three screens and make you play the “spot the actual error!” scrolling game.”
I’ve found it fairly simple to just instrument the code as needed when things don’t go as planned.
look at this post Graph visualization
got it.thank you:+1:
Try VisualDL: https://github.com/PaddlePaddle/VisualDL
VisualDL could produce graph with the ONNX file.
What I understand is that I need to export my model in a onnx file, as follows:
import torch.onnx dummy_input = Variable(torch.randn(4, 3, 32, 32)) torch.onnx.export(net, dummy_input, "model.onnx")
Then, VisualDL’s documentation suggests this. However, in such a case I don’t what logdir to use.
Netron can open ONNX files and supports reading basic sequential
Also check out this thread: Print Autograd Graph.
I have used Netron to visualize ONNX graphs. If you can export your model to ONNX successfully I would recommend giving it a try.
Check out HiddenLayer. I wrote this tool to visualize network graphs, and more specifically to visualize them in a way that is easier to understand. It merges related nodes together (e.g. Conv/Relu/MaxPool) and folds repeating blocks into one box and adds a x3 to imply that the block repeats 3 times rather than drawing it three times. This helps when you try to draw big networks, such as ResNet and Inception.
Great visualization tool! I just tried to visualize my slightly more complex models, and it works seamlessly.
I’ve started a new project for experiment documentation (Spyrkle).
It should be able to generate an interactive graph in an HTML page that you can style using CSS.