Support dictionary outputs in TorchScript tracer

[f0k]

🚀 Feature

torch.jit.trace() currently requires traced functions to output plain tensors or tuples. It would be helpful if it also allowed functions that return dictionaries.

Minimal failing example:

import torch
torch.jit.trace(lambda x: {'out': x}, torch.arange(3))

Motivation

Having modules take and return dictionaries makes it easier to define architectures with multiple inputs and outputs that are not always guaranteed to be present (e.g., for multi-task learning). Currently, such modules cannot be traced, e.g., for visualizing their structure.

Pitch

Functions returning dictionaries should be handled similarly to (named) tuples. For an OrderedDict, values can be taken in their stored order. For a general dict, the order should probably be made deterministic by sorting the keys, so it stays consistent across re-running the interpreter.
(Disclaimer: I wanted to use this to add a graph to tensorboard; I don't know what the implications of this suggestion would be for TorchScript traces.)

Alternatives

The alternative would be for me to wrap the module/function in a function that converts the dictionary to a (named?) tuple before returning it, just for the torch.jit.trace() call. This would not give me a trace that creates a dictionary, but at least a trace useful for visualization.

Additional context

This was part of issue #16453, which also asked for dictionary inputs, which are possible by now:

import torch
torch.jit.trace(lambda x: x['input'], {'input': torch.arange(3)})

cc @suo

[eellison]

@wanchaol you were working on supporting named tuple outputs right ? That would cover this use case.

[thuyen]

@eellison: will this be part of 1.4 release?

[wanchaol]

@thuyen you can already pass named tuple as output of the tracer, it's just that the result of the traced function will be just plain tuple.

#29751 this will enable tracer to take named tuple as output and preserve the names. But there's on-going discussions/issues on this PR and will not go into the 1.4 release candidate that we are having soon. This will go into the next release after the 1.4 rc.

[thuyen]

@wanchaol: thanks, I care more about tracing dictionary output (as mentioned in the issues title). Does #29751 support that as well?

[wanchaol]

@wanchaol: thanks, I care more about tracing dictionary output (as mentioned in the issues title). Does #29751 support that as well?

No, supporting Dictionary as output will be a separate PR. I can put sometime on investigating how to do it after I merge #29751

[thuyen]

Got it. Thanks!

[wanchaol]

@thuyen this feature is now supported in master under a strict flag in tracer. Closing

[thuyen]

Awesome, thanks @wanchaol

[mohammedayub44]

@wanchaol I'm installed the stable 1.5.0 version on Ubuntu a week ago. I'm running into this same issue. I don't see the strict flag. Should I be using a different version ?

[wanchaol]

@mohammedayub44 this is a feature after 1.5, you have to install master or nightly builds to get this feature, or wait for 1.6 :)

[mohammedayub44]

@wanchaol i ran with master , works fine. thanks !

[lamhoangtung]

@suo If I exported a model with dictionary as output. How can I catch it in C++ Pytorch frontend ?

Question asked in https://stackoverflow.com/questions/63917001/get-value-from-c10dictc10ivalue-c10ivalue-in-pytorch-c and https://discuss.pytorch.org/t/get-value-from-c10-dict-c10-ivalue-c10-ivalue/96466