[Feature] Add Cutout transform #1022

lkm2835 · 2021-11-08T17:18:10Z

CutOut transform was verified in MMDetection.
Can I bring it as it is?

However, I'm worried about whether there is a copyright infringement in this PR.

paper : https://arxiv.org/abs/1708.04552
original code : CutOut, test_cutout in mmdet

codecov · 2021-11-08T17:27:43Z

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.62%. Comparing base (d665f6b) to head (554f1ea).
Report is 301 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1022      +/-   ##
==========================================
+ Coverage   89.55%   89.62%   +0.07%     
==========================================
  Files         119      119              
  Lines        6626     6673      +47     
  Branches     1034     1043       +9     
==========================================
+ Hits         5934     5981      +47     
  Misses        488      488              
  Partials      204      204

Flag	Coverage Δ
unittests	`89.62% <100.00%> (+0.07%)`	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

Junjun2016 · 2021-11-08T22:27:26Z

Hi @lkm2835
Thanks for your contribution.
Could you please add more info to the PR message?
E.g. paper link, original code link, and code or info link in mmdet.
Both mmseg and mmdet belong to openmmlab, so feel free to move code or feature from mmdet to mmseg.

Junjun2016 · 2021-11-08T22:37:46Z

Could you please add more unittests since some lines are not covered by unittests?
Could you please also provide some results with or without Cutout in mmseg?

mmseg/datasets/pipelines/transforms.py

MengzhangLI · 2021-11-09T05:50:10Z

Hi, @lkm2835

Nice work indeed.

Please resolve comments @Junjun2016 left above. We would review your pr quickly.

lkm2835 · 2021-11-09T06:51:11Z

I checked the comment.

As soon as my personal work is over, I will resolve it.

lkm2835 · 2021-11-09T14:20:06Z

@Junjun2016

Could you please add more unittests since some lines are not covered by unittests?

--> I added unittests.

.

Could you please also provide some results with or without Cutout in mmseg?

--> 'Some results' means some my experimental results?

.

Maybe fill with ignoring label?
So we need a fill in argument for annotation? or ignore label.

--> When I experiment in mmseg, I don't fill with ignoring label.
Because, mmdet don't use ignoring label in bbox.
But, if you think Cutout need ignoring label in segmentation, I'll try to implement it.

Junjun2016 · 2021-11-09T17:12:00Z

'Some results' means some of my experimental results?

Yes.

When I experiment in mmseg, I don't fill with ignoring label.
Because, mmdet don't use ignoring label in bbox.
But, if you think Cutout need ignoring label in segmentation, I'll try to implement it.

We can make a comparison of these two cases.
Could you please also provide some results? (ablation)

lkm2835 · 2021-11-10T01:21:35Z

In my experiment(trash segmentation),

I got a 0.7236 mIoU in my valid set. upernet-swin-large was used.

I got a 0.7316 mIoU with Cutout without ignoring label.

In another experiment, result is 0.6970 --> 0.7177.

Are you sure you asked for these experimental results?
I don't have some results like ade20k, coco and cityscapes.

Junjun2016 · 2021-11-10T01:44:52Z

In my experiment(trash segmentation),

I got a 0.7236 mIoU in my valid set. upernet-swin-large was used.

I got a 0.7316 mIoU with Cutout without ignoring label.

In another experiment, result is 0.6970 --> 0.7177.

Are you sure you asked for these experimental results? I don't have some results like ade20k, coco and cityscapes.

Could also provide the results with ignoring label.

Junjun2016 · 2021-11-10T01:54:55Z

I don't have some results like ade20k, coco, and cityscapes.

Besides, could provide configs and results (if available) on ade20k, cityscapes both with and without ignoring label.

mmseg/datasets/pipelines/transforms.py

lkm2835 · 2021-11-10T10:45:43Z

modified ignore_index with seg_fill_in.
seg_fill_in default is None.
if seg_fill_in is not None, modify the seg_fields in results.

Junjun2016 · 2021-11-10T12:07:28Z

modified ignore_index with seg_fill_in. seg_fill_in default is None. if seg_fill_in is not None, modify the seg_fields in results.

Great, thanks for your hard work.
Next, could do some ablation study.

lkm2835 · 2021-11-10T12:25:13Z

Great, thanks for your hard work.
Next, could do some ablation study.

Sorry, I'm asking because I'm not good at English.
Do I have to do some ablation study?

Junjun2016 · 2021-11-10T12:54:31Z

modified ignore_index with seg_fill_in. seg_fill_in default is None. if seg_fill_in is not None, modify the seg_fields in results.

Great, thanks for your hard work. Next, could do some ablation study.

We will do some ablation study with seg_fill_in to be None and ignore label, and record the results here.
You could also provide some experimental results with different hyperparameters in CutOut.
Besides, could you please also move other nice features in mmdet to mmseg?

RockeyCoss · 2021-11-15T08:25:59Z

Please merge the master branch into your branch, thank you.

lkm2835 · 2021-11-15T09:54:37Z

If possible, Could you advise me how to use Codecov for test code in my terminal? (before PR)

Junjun2016 · 2021-11-15T10:06:32Z

If possible, Could you advise me how to use Codecov for test code in my terminal? (before PR)

@MengzhangLI, please.

Junjun2016 · 2021-11-15T10:51:30Z

If possible, Could you advise me how to use Codecov for test code in my terminal? (before PR)

@MengzhangLI, please.

coverage run --source={package, e.g., mmdet} --branch -m pytest
coverage report -m

MengzhangLI · 2021-11-15T11:09:03Z

Using pytest to check unit test

pytest tests/${test_file}.py::${test_function}
pytest tests/${test_file}.py

Install coverage

pip install coverage

Using code below under your repo, thus you can get unittest and return its coverage rate.

coverage run --source={package, e.g., mmseg} --branch -m pytest
coverage report -m

lkm2835 · 2021-11-15T11:21:36Z

Thank you :)

Junjun2016 · 2021-11-15T11:46:48Z

Using pytest to check unit test

pytest tests/${test_file}.py::${test_function}
pytest tests/${test_file}.py

Install coverage

pip install coverage

Using code below under your repo, thus you can get unittest and return its coverage rate.

coverage run --source={package, e.g., mmseg} --branch -m pytest
coverage report -m

pytest tests # test all files in tests folder

Junjun2016 · 2021-11-15T13:57:59Z

Hi @lkm2835
We are doing some ablations now and we will merge this PR soon.
Could you please move the Mosaic to mmseg? and should adjust it to the semantic segmentation task.

lkm2835 · 2021-11-15T14:35:07Z

Hi @lkm2835 We are doing some ablations now and we will merge this PR soon. Could you please move the Mosaic to mmseg? and should adjust it to the semantic segmentation task.

I have to read code(Mosaic).
After that, I'll try PR.

Junjun2016 · 2021-11-15T14:59:24Z

Hi @lkm2835 We are doing some ablations now and we will merge this PR soon. Could you please move the Mosaic to mmseg? and should adjust it to the semantic segmentation task.

I have to read code(Mosaic). After that, I'll try PR.

Thanks a lot.

mmseg/datasets/pipelines/transforms.py

MengzhangLI · 2021-11-15T18:16:46Z

Hi, @lkm2835

Thank you so much for your warm-hearted work.

In the future, if you have fixed the comment by reviewer, the comment would be shown as Outdated and you could click Resolve conversation bottom like the figure below:

The pr would be more simple ;)

Best,

Junjun2016 · 2021-11-16T18:02:08Z

Hi @RockeyCoss
Please review it and give some comments if necessary.

MengzhangLI · 2021-11-19T10:26:52Z

Hi, @lkm2835
Thank you for your warm-hearted help on MMSegmentation.
The table above is numerical results of us, hoping it is useful for your related project.

lkm2835 · 2021-11-19T10:51:14Z

Thank you for providing the experimental results.
What does best index mean?

RockeyCoss · 2021-11-30T08:46:35Z

@lkm2835 Hi, sorry for the late reply. During training, we evaluate the model and save the model for every certain number of training iterations. Best index means the index of the saved model that performs best.

lkm2835 · 2021-11-30T12:18:20Z

@lkm2835 Hi, sorry for the late reply. During training, we evaluate the model and save the model for every certain number of training iterations. Best index means the index of the saved model that performs best.

@RockeyCoss, Thanks for your replay.
Is there anything to need more ablation study?
Recently, I have an 1 extra GPU.

Junjun2016 · 2021-11-30T12:22:09Z

@lkm2835 Hi, sorry for the late reply. During training, we evaluate the model and save the model for every certain number of training iterations. Best index means the index of the saved model that performs best.

@RockeyCoss, Thanks for your replay. Is there anything to need more ablation study? Recently, I have an 1 extra GPU.

@RockeyCoss Could report more results here.

Junjun2016 · 2021-11-30T12:23:14Z

@lkm2835 Hi, sorry for the late reply. During training, we evaluate the model and save the model for every certain number of training iterations. Best index means the index of the saved model that performs best.

@RockeyCoss, Thanks for your replay. Is there anything to need more ablation study? Recently, I have an 1 extra GPU.

Not needed yet since we have already run lots of ablation studies.

Junjun2016 · 2021-11-30T12:36:42Z

Hi @lkm2835
We will merge this augmentation in this version.
Thanks for your contribution again and looking forward to your PR again.

* Fix typo in usage example * [Feature] Add CutOut transform * CutOut repr covered by unittests * Cutout ignore index, test * ignore_index -> seg_fill_in, defualt is None * seg_fill_in is added to repr * test is modified for seg_fill_in is None * seg_fill_in (int), 0-255 * add seg_fill_in test * doc string for seg_fill_in * rename CutOut to RandomCutOut, add prob * Add unittest when cutout is False

…lab#1022) Deprecate `init_git_repo` and `push_to_hub`, refactor `train_unconditional.py`

lkm2835 and others added 5 commits October 29, 2021 00:12

Fix typo in usage example

525bb6b

Merge branch 'open-mmlab:master' into master

e34d8b4

Merge branch 'open-mmlab:master' into master

0328021

Merge branch 'open-mmlab:master' into master

13a3b00

[Feature] Add CutOut transform

fec2c0f

lkm2835 changed the title ~~Cutout~~ [Feature] Add Cutout transform Nov 8, 2021

Junjun2016 reviewed Nov 8, 2021

View reviewed changes

mmseg/datasets/pipelines/transforms.py Outdated Show resolved Hide resolved

CutOut repr covered by unittests

4aaebbe

Cutout ignore index, test

606bf12

Junjun2016 reviewed Nov 10, 2021

View reviewed changes

mmseg/datasets/pipelines/transforms.py Outdated Show resolved Hide resolved

Junjun2016 reviewed Nov 10, 2021

View reviewed changes

mmseg/datasets/pipelines/transforms.py Outdated Show resolved Hide resolved

lkm2835 added 3 commits November 10, 2021 18:34

ignore_index -> seg_fill_in, defualt is None

8cde1c9

seg_fill_in is added to repr

e23da40

test is modified for seg_fill_in is None

714093e

lkm2835 added 2 commits November 10, 2021 20:27

seg_fill_in (int), 0-255

82599e0

add seg_fill_in test

57779b5

lkm2835 and others added 2 commits November 15, 2021 18:18

Merge branch 'open-mmlab:master' into master

4e1f1e3

Merge master into cutout

c7c4329

lkm2835 commented Nov 15, 2021

View reviewed changes

mmseg/datasets/pipelines/transforms.py Show resolved Hide resolved

Add unittest when cutout is False

554f1ea

Junjun2016 approved these changes Nov 16, 2021

View reviewed changes

Junjun2016 requested a review from xvjiarui November 16, 2021 18:01

Junjun2016 merged commit 78a6ff6 into open-mmlab:master Nov 30, 2021

MengzhangLI mentioned this pull request Dec 16, 2021

Cannot add more augmentations by train options #1113

Closed

lkm2835 mentioned this pull request Dec 21, 2021

Changing CutOut to RandomCutOut open-mmlab/mmdetection#6850

Closed

aravind-h-v pushed a commit to aravind-h-v/mmsegmentation that referenced this pull request Mar 27, 2023

Deprecate init_git_repo, refactor train_unconditional.py (open-mm…

fbcc383

…lab#1022) Deprecate `init_git_repo` and `push_to_hub`, refactor `train_unconditional.py`

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Feature] Add Cutout transform #1022

[Feature] Add Cutout transform #1022

lkm2835 commented Nov 8, 2021 •

edited

Loading

codecov bot commented Nov 8, 2021 •

edited

Loading

Junjun2016 commented Nov 8, 2021

Junjun2016 commented Nov 8, 2021

MengzhangLI commented Nov 9, 2021

lkm2835 commented Nov 9, 2021

lkm2835 commented Nov 9, 2021 •

edited

Loading

Junjun2016 commented Nov 9, 2021

lkm2835 commented Nov 10, 2021

Junjun2016 commented Nov 10, 2021

Junjun2016 commented Nov 10, 2021

lkm2835 commented Nov 10, 2021 •

edited

Loading

Junjun2016 commented Nov 10, 2021

lkm2835 commented Nov 10, 2021

Junjun2016 commented Nov 10, 2021

RockeyCoss commented Nov 15, 2021

lkm2835 commented Nov 15, 2021

Junjun2016 commented Nov 15, 2021

Junjun2016 commented Nov 15, 2021

MengzhangLI commented Nov 15, 2021

lkm2835 commented Nov 15, 2021

Junjun2016 commented Nov 15, 2021

Junjun2016 commented Nov 15, 2021

lkm2835 commented Nov 15, 2021

Junjun2016 commented Nov 15, 2021

MengzhangLI commented Nov 15, 2021 •

edited

Loading

Junjun2016 commented Nov 16, 2021

MengzhangLI commented Nov 19, 2021

lkm2835 commented Nov 19, 2021

RockeyCoss commented Nov 30, 2021 •

edited

Loading

lkm2835 commented Nov 30, 2021 •

edited

Loading

Junjun2016 commented Nov 30, 2021

Junjun2016 commented Nov 30, 2021

Junjun2016 commented Nov 30, 2021

[Feature] Add Cutout transform #1022

[Feature] Add Cutout transform #1022

Conversation

lkm2835 commented Nov 8, 2021 • edited Loading

codecov bot commented Nov 8, 2021 • edited Loading

Codecov Report

Junjun2016 commented Nov 8, 2021

Junjun2016 commented Nov 8, 2021

MengzhangLI commented Nov 9, 2021

lkm2835 commented Nov 9, 2021

lkm2835 commented Nov 9, 2021 • edited Loading

Junjun2016 commented Nov 9, 2021

lkm2835 commented Nov 10, 2021

Junjun2016 commented Nov 10, 2021

Junjun2016 commented Nov 10, 2021

lkm2835 commented Nov 10, 2021 • edited Loading

Junjun2016 commented Nov 10, 2021

lkm2835 commented Nov 10, 2021

Junjun2016 commented Nov 10, 2021

RockeyCoss commented Nov 15, 2021

lkm2835 commented Nov 15, 2021

Junjun2016 commented Nov 15, 2021

Junjun2016 commented Nov 15, 2021

MengzhangLI commented Nov 15, 2021

lkm2835 commented Nov 15, 2021

Junjun2016 commented Nov 15, 2021

Junjun2016 commented Nov 15, 2021

lkm2835 commented Nov 15, 2021

Junjun2016 commented Nov 15, 2021

MengzhangLI commented Nov 15, 2021 • edited Loading

Junjun2016 commented Nov 16, 2021

MengzhangLI commented Nov 19, 2021

lkm2835 commented Nov 19, 2021

RockeyCoss commented Nov 30, 2021 • edited Loading

lkm2835 commented Nov 30, 2021 • edited Loading

Junjun2016 commented Nov 30, 2021

Junjun2016 commented Nov 30, 2021

Junjun2016 commented Nov 30, 2021

lkm2835 commented Nov 8, 2021 •

edited

Loading

codecov bot commented Nov 8, 2021 •

edited

Loading

lkm2835 commented Nov 9, 2021 •

edited

Loading

lkm2835 commented Nov 10, 2021 •

edited

Loading

MengzhangLI commented Nov 15, 2021 •

edited

Loading

RockeyCoss commented Nov 30, 2021 •

edited

Loading

lkm2835 commented Nov 30, 2021 •

edited

Loading