Kate/splitter cli

[jihyeonyi]

Summary

This PR includes

• supporting CLI for task-specific split
• Revise re-identification split
• Update documentation regarding the task-specific split

How to test

Unittest

$ python -m unittest -v tests/test_splitter.py

Testing classification split with imagenet dataset.

Notes: Imagenet doesn't support subsets but, checking subsets at the project level is enough here.

$ pip install .
$ datum project create -o imagenet
$ datum source add path <path-to-source> -f imagenet -p imagenet/
$ datum project transform -t classification_split -p imagenet/ -- --subset train:.5 --subset val:.2 --subset test:.3
$ datum project info -p imagenet-classification_split

Testing detection split with voc dataset

$ pip install .
$ datum project import -i <path-to-voc> -f voc
$ cd voc/
$ datum project transform -t detection_split -- --subset train:.5 --subset val:.2 --subset test:.3
$ datum project info -p voc-detection_split

Testing re-identification split with imagenet dataset.

Notes: Datumaro doesn't support re-id dataset now, so the classification dataset is used instead.

$ pip install .
$ datum project create -o imagenet
$ datum source add path <path-to-imagenet> -f imagenet -p imagenet/
$ datum project transform -t reidentification_split -p imagenet/ -- --subset train:.5 --subset val:.2 --subset test:.3 --query .5
$ datum project info -p imagenet-reidentification_split

Checklist

• I submit my changes into the develop branch
• I have added description of my changes into CHANGELOG
• I have updated the documentation accordingly
• I have added tests to cover my changes
• I have linked related issues)

License

• I submit my code changes under the same MIT 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) 2020 Intel Corporation
#
# SPDX-License-Identifier: MIT

[zhiltsov-max]

Please check the updated class descriptions for correctness.

Future updates could include:

• ignoring attributes in classification split (for captions, descriptions and other technical attributes)
• splitting using an attribute as label in classification split
• using polygons and masks in detection split

[jihyeonyi]

Here, we avoid having the same person id or object id. It could be label or attribute if attr_for_id is specified.

[jihyeonyi]

One more thing is, actually train and val set share person id or object id. (Most person re-identification data doesn't have val set though). But they do not share IDs with test set.
I'm not sure how accurate the explanation should be.
If you feel the current explanation is sufficient, please leave it as it is.

[jihyeonyi]

Please check the updated class descriptions for correctness.

Future updates could include:

• ignoring attributes in classification split (for captions, descriptions and other technical attributes)
• splitting using an attribute as label in classification split
• using polygons and masks in detection split

Thank you for revising the descriptions.
And for future updates,

1. Would you like to remove the attribute-based splitting or just make it optional?
   I think the latter is better.
2. When you say 'splitting using an attribute as label', do you mean splitting using only attributes, regardless of labels?
3. Does the detection task have polygons or masks? I thought it is for the segmentation task. Maybe I'm wrong.
   For your information, I'll add a splitter for the segmentation task. So why don't you add polygons or masks later?

[zhiltsov-max]

1. Would you like to remove the attribute-based splitting or just make it optional?

Optional, enabled by default.

2. When you say 'splitting using an attribute as label', do you mean splitting using only attributes, regardless of labels?

I mean using a single attribute, like in re-id. Maybe, using some subset of them / ignoring some attributes.

3. Does the detection task have polygons or masks?

In Mask R-CNN they are intermixed with segmentation task. I, personally, consider these types of annotations more or less interchangeable, because all these types can be used for training a segmentation and a detection algorithm.