Easiest way to check for a "connected" graph

1

Hi,
Is there an easy way to check if my entire sequence of operations is connected? Apart from manually debugging through the entire flow and checking if each intermediate variable has a grad_fn? I haven’t used Tensorflow much either, but what I am looking for is probably a way to print out the entire “graph” in PyTorch (though I understand in PyTorch it’s dynamically created)?

Sorry if that didn’t make much sense. I mean if I had something like:
x = tensor which requires grad
y=x^2+3
z=y.mean()

Now I want something which tells me the entire sequence of grad_functions. The purpose is to troubleshoot a project where I have quite a long such chain of functions and hopefully I am not messing up somewhere.

2

Sorry this isn’t much (it’s kinda late here), but this post has a snippet that lets you examine the gradient flow, and @apaszke has a GitHub Gist that should let you see the network’s graph. Hope that helps!

1 Like