autograd functions implemented in C++ do not expose previous_functions

[Maratyszcza]

According to PyTorch autograd documentation, autograd Functions provide previous_functions instance variables to track which autograd Functions produced their inputs. previous_functions is necessary to reconstruct the DAG in Python code and potentially export it to visualizers/analyzers/other frameworks. I found that autograd Function objects implemented in C++ code do not implement previous_functions instance variables, which makes reconstructing the DAG impossible (I believe this information is preserved somewhere in the C++ classes, but didn't find a way to access it from Python).

Repro:

import torch.autograd
import torch.nn

def qualified_type(var):
    if var is None:
        return "None"
    else:
        module = var.__class__.__module__
        type = var.__class__.__name__
        if module is None or module == "__builtin__":
            return type
        else:
            return module + "." + type

import torchvision.models as models
model = models.alexnet(pretrained=True)

input = torch.FloatTensor(1, 3, 224, 224)
input_var = torch.autograd.Variable(input)
output_var = model(input_var)

seen = set()
def add_nodes(var):
    if var not in seen:
        if isinstance(var, torch.autograd.Variable):
            print("Variable: " + qualified_type(var))
        else:
            print("Function: " + qualified_type(var))
        seen.add(var)
        if hasattr(var, 'previous_functions'):
            for u in var.previous_functions:
                add_nodes(u[0])
        elif hasattr(var, 'creator'):
            if var.creator is not None:
                add_nodes(var.creator)

add_nodes(output_var.creator)

[apaszke]

I think we're going to fix that only after we finish autograd refactor (previous_functions will likely be removed anyway).