Closed lucasosouza closed 3 years ago
@vkakerbeck added a simple loop that goes through every object allocated in memory - my idea is we could expand it to know more about the tensors if that seemed to be the case. But the memory of the tensors itself is not increasing, although the overall memory consumption is. That reinforces my belief that is the computational graph being stored (at that point is probably being compressed into a different format and is no longer a tensor, hence my code is not catching it).
A good next step is to actually visualize this graph to understand where it is going wrong.
@vkakerbeck added a mixin that allows visualization of computational graph with pytorchviz
Please review the PR whenever you can, I think we should merge these tools, they can be helpful moving forward
Profiler code (don't merge it yet, if necessary fetch from branch).