Support configurable input size

[eunwoosh]

Summary

This PR includes

• adds configurable input size and adaptive input size to OTX.
• refactor input size in model code.

Explanation

adds configurable input size feature to OTX

User can configure desired input size after this PR. input_size value in 'data' part of recipe is automatically passed to model unless model is already initialized before it's passed to the engine in API use case.
Also, user can use adaptive input size which finds appropriate input size based on dataset statistics. Also, user can force to use downscale only which doesn't change input size if bigger input size is considered as proper input size.
All interfaces for input size are added to OTXDataModule.
So CLI user can uses it by adding --data.input_size 1000 for configuring input size and --data.adaptive_input_size "auto" for adaptive input size.
API user can use it by setting arguments of OTXDataModule. If model is initialized in engine, input_size will be automatically set but if not, then user should set input_size argument of the model by himself or herself.

refactor input size in model code

Previously, how input_size is manged in model side is different depending on each model.
Now, it's unified. All models get input_size argument and keep it as attribute except zvp model which uses fixed input_size.
Model also initializes own module differently based on input_size if necessary.

How to test

• otx train --config ... --data_root ... --data.input_size 1024
• otx train --config ... --data_root ... --data.adaptive_input_size "auto | downscale"

Checklist

• I have added unit tests to cover my changes.​
• I have added integration tests to cover my changes.​
• I have ran e2e tests and there is no issues.
• I have added the description of my changes into CHANGELOG in my target branch (e.g., CHANGELOG in develop).​
• I have updated the documentation in my target branch accordingly (e.g., documentation in develop).
• I have linked related issues.

License

• I submit my code changes under the same Apache License that covers the project.
  Feel free to contact the maintainers if that's a concern.
• I have updated the license header for each file (see an example below).

# Copyright (C) 2024 Intel Corporation
# SPDX-License-Identifier: Apache-2.0

[goodsong81]

Nice work. A few points to mention are as follows:

• hasatrr() based branch logics might be quickly rusted and tangled. Hope it would be based on formal interfaces in the future.
• input_size term might be better to be more consistent in terms of usage

[sooahleex]

Looks good to me for h-label head!

[goodsong81]

LGTM w/ one last comment: we may want to touch the mem cache max image size to make this feature more effective in training
(Maybe in another PR? :p)

[eunwoosh]

LGTM w/ one last comment: we may want to touch the mem cache max image size to make this feature more effective in training (Maybe in another PR? :p)

yes, we can consider including that feature in the next release.

[goodsong81]

LGTM w/ one last comment: we may want to touch the mem cache max image size to make this feature more effective in training (Maybe in another PR? :p)

yes, we can consider including that feature in the next release.

FYI, the configured input size could be set to the max memcache image size.

[eunwoosh]

@goodsong81 I understand your intention. If it doesn't need huge code change, maybe it can be included in this release. Thanks for comment :)