[onert/train] support expand_dims for training

[zetwhite]

Let's support expand_dims for the training feature.

[mbencer]

@zetwhite Can I take this issue?

[zetwhite]

@zetwhite Can I take this issue?

Sure.

BTW, May I ask why you are interested in this task?

[mbencer]

My team and I are already considering contribution in some parts of ONE project.
So far I had occasion to contribute in runtime part while working on fine tuning feature and I would like to continue contribution in this area ;-)
Do we have some deadline about this task (I have also some unfinished activities in other repositories) ?
BTW.: Do you know any other training operators that need to be implemented?

[zetwhite]

Do we have some deadline about this task (I have also some unfinished activities in other repositories) ?

About the background, I made this issue to support an inner model.
For Expand_dims there is no deadline.

BTW.: Do you know any other training operators that need to be implemented?

We might have other operators and other issues you might have some interest.
ping to @jyoungyun @ragmani other @Samsung/one_onert members

[ragmani]

Do we have some deadline about this task (I have also some unfinished activities in other repositories) ?

Currently, this op is removed from the inner model to be supported for training. So, there is no deadline and this issue might not be necessary.

My team and I are already considering contribution in some parts of ONE project.
So far I had occasion to contribute in runtime part while working on fine tuning feature and I would like to continue contribution in this area ;-)

I recommend #12836, The in-place feature can also be utilized to improve latency and memory usage when back-propagating in ops such as reshape.
If you would like to continue contribution in runtime part, you can find tasks in https://github.com/orgs/Samsung/projects/16/views/1.

[zetwhite]

Currently, this op is removed from the inner model to be supported for training.

Adding some details, I didn't know when I made this issue, but the target model's final version did not contain expand_dims.

So, about expand_dims, It might be good to be supported (future, there might be a model that include expand_dims), but there is no exact necessary right now.

[mbencer]

I recommend #12836, The in-place feature can also be utilized to improve latency and memory usage when back-propagating in ops such as reshape

Ok, so I'll start with and let's see what reasonable to do next ;)

If you would like to continue contribution in runtime part, you can find tasks in https://github.com/orgs/Samsung/projects
/16/views/1.

Thank you very much! Very useful board.