[Feature request] Tensor attributes about glow

[qcolombet]

Hi @jsubag,

1. Yes, but that is not mutually exclusive. When we build for training, we end up with variables being used both by save nodes and elsewhere.
2. At the instruction level, the variables visible from the outside are represented with the WeightVar class. Those would be the input/output of the program. Now to know what is an output, you would need to check which one are written into. That should be doable by walking through the users of the variables and check if they are ever used as output (@out in the textual representation).

Now, for figuring out what and when should these be copied around, we are reworking the graph IR to have variables accessed exclusively through save and a new load nodes. We would potentially lower then into some new IR instructions that would make it easy to determine when such copies should occur (basically, just where you see such instructions).

[jsubag]

Hi @qcolombet ,

Thanks for clarifying - that was pretty much what I understood.
I think all of these corner cases will be neatly resolved with Load/Save instructions.
Also, this makes it easier for backends to optimize data transfer (e.g. copy output tensors as soon as they're ready instead of at the end of the forward pass).

Btw, will this change be applied to all variables or just public ones? and will constants be treated differently (weights/biases)?

[nadavrot]

@jsubag Jacob, thanks for working on this. Input/output annotation is a solution to some problem. I am not sure if it's the only solution and if it's the best solution. We need to make sure that any new attributes match the design of the rest of the system. For example, how do new attributes interact with public/private, mutability, etc.

Let's talk about the problem. What usecase is the current system not handling well?

[jsubag]

@nadavrot Thanks for pointing out WeightVar::MutabilityKind, I missed that in the code :)
In that case, are the following categories correct?

1. Private+Constant: constants known during compilation e.g. weights/biases/initial hidden states
2. Public+Constant: constant during inference (not known during compilation) e.g. input tensors
3. Private+Mutable: internal tensors
4. Public+Mutable: tensors mutable during inference and exposed to the application e.g. output tensors

This however doesn't map the case where a tensor is used as both input and output, e.g. weights during training.
Unless this condition is illegal in Glow and the resulting function will create separate tensors for input and output.
Can you clarify this point please?

Thanks.

[jfix71]

@jsubag We actually do not have anything that can be Public+Constant currently; Public implies Mutable. So both input and output tensors are considered to be Public+Mutable.

Perhaps it makes more sense to allow input tensors to be Public+Constant, though, as they are still constant during execution even if modified/accessed before or after. If we made such a change, then if a tensor was used as both input and output then it would be Public+Mutable, as it currently is.

[jsubag]

@jfix71 I think the thing I'm missing is a clear way to figure out if a tensor is considered an input and/or output.
The motivation is figuring out what should be copied to/from memory resources of the backend before and after the forward pass.
Even though the common cases are easy to identify by traversing the graph, it probably makes more sense to either have an explicit notation on them or some shared utility function that all backends can use.

[nadavrot]

@jsubag @jfix71 I think that the description of the problem in the last post is great. The next step would be to post proposal for a design that solves the problem. @qcolombet can help in brainstorming ideas for a design. What do you think?

[qcolombet]

Hi @jsubag,

I am not sure I get the problem. Outputs should be easy to identify, they are used by Save node. Having a backend utility function sounds reasonable though.

Do you see cases where we would need to explicitly tag a variable?

[jsubag]

Hi @qcolombet ,

1. Is it fair to say then that all public tensors targeted by Save nodes are outputs and all other public tensors are inputs?
   This implies inputs and outputs are mutually exclusive, does that fit with weights usage during training?

2. Is there something propagated from Save nodes to the instruction/variable level?
   I.e. if a backend wants to execute a Function by going over its instructions how can it determine if an instruction destination variable is a Function output?

[qcolombet]

Btw, will this change be applied to all variables or just public ones? and will constants be treated differently (weights/biases)?

All variables public or not, constant or not.

[nadavrot]

@jsubag This issue is related to #1334