CodeCamp #151[Feature] Support HieraSeg on cityscapes #2444

AI-Tianlong · 2022-12-30T07:36:17Z

Support `HieraSeg` interface on `cityscapes`

Motivation

Support HieraSeg interface on cityscapes dataset
Paper link : https://ieeexplore.ieee.org/document/9878466/

@article{li2022deep,
  title={Deep Hierarchical Semantic Segmentation},
  author={Li, Liulei and Zhou, Tianfei and Wang, Wenguan and Li, Jianwu and Yang, Yi},
  journal={CVPR},
  year={2022}
}

Modification

Add HieraSeg_Projects on projects/
Add sep_aspp_contrast_head decoder head.
Add HieraSeg config.
Add hiera_loss, hiera_triplet_loss_cityscape, tree_triplet_loss

Checklist

Pre-commit or other linting tools are used to fix the potential lint issues.
The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
If the modification has potential influence on downstream projects, this PR should be tested with downstream projects, like MMDet or MMDet3D.
The documentation has been modified accordingly, like docstring or example tutorials.

…ig.md

CodeCamp #144 [Doc] Chinese version of config tutorial

Co-authored-by: 谢昕辰 <xiexinch@outlook.com>

CodeCamp #147 [Doc] Add Chinese version of train & test tutorial

Co-authored-by: 谢昕辰 <xiexinch@outlook.com>

CodeCamp #1562 [Doc] Add Chinese version of `overview.md`

[CI] Add torch1.13 checking in CI

[Doc] Add ZN datasets.md in dev-1.x

…/mmsegmentation into AI-Tianlong/Support_HieraSeg

projects/HieraSeg_project/README.md

projects/HieraSeg_project/configs/_base_/models/deeplabv3plus_r50-d8_vd_contrast.py

...aSeg_project/configs/hieraseg/deeplabv3plus_r101-d8_512x1024_80k_cityscapes_hiera_triplet.py

projects/HieraSeg_project/losses/hiera_loss.py

xiexinch · 2023-01-05T06:48:17Z

projects/HieraSeg_project/losses/tree_triplet_loss.py

+import torch.nn.functional as F
+
+
+class TreeTripletLoss(nn.Module):


Might also add to register and support customizing parameters

Might add docstring.

xiexinch · 2023-01-05T06:49:19Z

projects/HieraSeg_project/losses/hiera_triplet_loss_cityscape.py

+        super().__init__()
+        self.num_classes = num_classes
+        self.loss_weight = loss_weight
+        self.treetripletloss = TreeTripletLoss(19, hiera_map, hiera_index)


Might support customizing.

xiexinch · 2023-01-05T06:51:02Z

projects/HieraSeg_project/losses/hiera_loss.py

+    MCMA = predictions[:, :num_classes, :, :]
+    MCMB_back = torch.max(
+        torch.cat([
+            predictions[:, 0:1, :, :],
+            predictions[:, num_classes:num_classes + 1, :, :]
+        ],
+                  dim=1), 1, True)[0]
+    MCMB1 = torch.max(


Might add some comments to explain what MCMA, MCMB, and others mean.

projects/HieraSeg_project/decode_head/sep_aspp_contrast_head.py

projects/HieraSeg_project/losses/tree_triplet_loss.py

…ort_HieraSeg

MeowZheng · 2023-01-12T06:33:45Z

projects/HieraSeg_project/losses/tree_triplet_loss.py

+            feats_pos = feats[index_pos][:min_size]
+            feats_neg = feats[index_neg][:min_size]
+
+            distance = torch.zeros(min_size, 2).cuda()


the device of model training might be not cuda.

Suggested change

distance = torch.zeros(min_size, 2).cuda()

distance = torch.zeros(min_size, 2).to(feats)

## Motivation Supplementary PR #2444 Fix tiny bug and add loss_by_feat() to compute loss to train. The inference process have verified to be accurate. ## Modification - modify `sep_aspp_contrast_head.py` , add `loss_by_feat()` function to train(training still has bug, will fix in future😫) - fix testing commands path error `bash tools/dist_test.sh projects/HieraSeg_project/` to `bash tools/dist_test.sh projects/HieraSeg/` at README.md

…-mmlab#2444) ## Support `HieraSeg` interface on `cityscapes` ## Motivation Support `HieraSeg` interface on cityscapes dataset Paper link : https://ieeexplore.ieee.org/document/9878466/ ``` @Article{li2022deep, title={Deep Hierarchical Semantic Segmentation}, author={Li, Liulei and Zhou, Tianfei and Wang, Wenguan and Li, Jianwu and Yang, Yi}, journal={CVPR}, year={2022} } ``` ## Modification Add `HieraSeg_Projects` on `projects/` Add `sep_aspp_contrast_head` decoder head. Add `HieraSeg` config. Add `hiera_loss`, `hiera_triplet_loss_cityscape`, `tree_triplet_loss`

open-mmlab#2565) ## Motivation Supplementary PR open-mmlab#2444 Fix tiny bug and add loss_by_feat() to compute loss to train. The inference process have verified to be accurate. ## Modification - modify `sep_aspp_contrast_head.py` , add `loss_by_feat()` function to train(training still has bug, will fix in future😫) - fix testing commands path error `bash tools/dist_test.sh projects/HieraSeg_project/` to `bash tools/dist_test.sh projects/HieraSeg/` at README.md

BLUE-coconut and others added 30 commits November 30, 2022 20:18

doc

f2b8d2f

[Doc]Translate the 1_config.md and modify a wrong statement in 1_conf…

67b51b7

…ig.md

Translate the 1_config.md and modify a wrong statement in 1_config.md

4287fd2

modify part of content

5668182

Modify some expressions

ad470a3

changed parts of content

104a24d

Apply suggestions from code review

b0e57c6

Merge pull request #2371 from pofengdenihong/dev-1.x

9750c97

CodeCamp #144 [Doc] Chinese version of config tutorial

[Doc] Add ZN datasets.md in dev-1.x

8a3d6cd

modified

3d1f8b7

fix typo

addd38d

Update docs/zh_cn/user_guides/4_train_test.md

372a6ce

Co-authored-by: 谢昕辰 <xiexinch@outlook.com>

CodeCamp #1562 [Doc] update overview.md

dc8aa35

Update overview.md

dddf093

Merge pull request #2355 from BLUE-coconut/master

7edb141

CodeCamp #147 [Doc] Add Chinese version of train & test tutorial

Add torch1.13 in CI

b770884

use mim install mm packages

02030b1

Update docs/zh_cn/overview.md

d755707

Co-authored-by: 谢昕辰 <xiexinch@outlook.com>

Update docs/zh_cn/overview.md

c2042f3

Co-authored-by: 谢昕辰 <xiexinch@outlook.com>

install all requirements

fdd1b95

Merge pull request #2397 from tianleiSHI/dev-1.x

7537987

CodeCamp #1562 [Doc] Add Chinese version of `overview.md`

install wheel

492313b

add ref

815b24e

Merge pull request #2402 from xiexinch/add-torch1.13-in-ci-1.x

d7b85aa

[CI] Add torch1.13 checking in CI

fix

164fd3f

fix

e88489e

Merge pull request #2387 from MengzhangLI/zn_datasets_1.x

eeee12e

[Doc] Add ZN datasets.md in dev-1.x

add example project

f5b4c12

add ci ignore

81da7aa

DOC

532ff5b

AI-Tianlong and others added 7 commits January 4, 2023 00:08

Update README.md

b01ea49

update HieraSeg_projects

55d47d7

fix lint

97b76c0

update losses

d94caeb

Merge branch 'AI-Tianlong/Support_HieraSeg' of github.com:AI-Tianlong…

53ade89

…/mmsegmentation into AI-Tianlong/Support_HieraSeg

Update README.md

bfcfb69

Update README.md

978f94f

xiexinch mentioned this pull request Jan 4, 2023

Will you release the weights? lingorX/HieraSeg#5

Open

AI-Tianlong and others added 3 commits January 4, 2023 13:26

Update README.md

4a6a7fb

Update README.md

02a7017

update README.md

a202243

xiexinch reviewed Jan 5, 2023

View reviewed changes

update HieraSeg_Projects

622492a

MengzhangLI reviewed Jan 11, 2023

View reviewed changes

projects/HieraSeg_project/losses/tree_triplet_loss.py Outdated Show resolved Hide resolved

AI-Tianlong and others added 4 commits January 11, 2023 13:12

Create tree_triplet_loss.py

6595ed0

Merge remote-tracking branch 'upstream/dev-1.x' into AI-Tianlong/Supp…

67d7712

…ort_HieraSeg

update README and add some docstring

c328f41

minor change

c9005cc

xiexinch approved these changes Jan 12, 2023

View reviewed changes

MeowZheng reviewed Jan 12, 2023

View reviewed changes

xiexinch added 3 commits January 12, 2023 15:35

rename project folder and add docstring

2f775e7

minor change

ebfdcf1

add docstring to loss

815002a

MeowZheng approved these changes Jan 12, 2023

View reviewed changes

MeowZheng merged commit ce09639 into open-mmlab:dev-1.x Jan 12, 2023

This was referenced Feb 3, 2023

[Fix] Fix projects HieraSeg bug (fix Pull requests #2444) #2561

Closed

[Fix] Rename and Fix bug of projects HieraSeg (old PR #2444) #2565

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

CodeCamp #151[Feature] Support HieraSeg on cityscapes #2444

CodeCamp #151[Feature] Support HieraSeg on cityscapes #2444

AI-Tianlong commented Dec 30, 2022 •

edited

Loading

xiexinch Jan 5, 2023

xiexinch Jan 5, 2023

xiexinch Jan 5, 2023

xiexinch Jan 5, 2023

MeowZheng Jan 12, 2023

		import torch.nn.functional as F


		class TreeTripletLoss(nn.Module):

	distance = torch.zeros(min_size, 2).cuda()
	distance = torch.zeros(min_size, 2).to(feats)

CodeCamp #151[Feature] Support HieraSeg on cityscapes #2444

CodeCamp #151[Feature] Support HieraSeg on cityscapes #2444

Conversation

AI-Tianlong commented Dec 30, 2022 • edited Loading

Support HieraSeg interface on cityscapes

Motivation

Modification

Checklist

xiexinch Jan 5, 2023

Choose a reason for hiding this comment

xiexinch Jan 5, 2023

Choose a reason for hiding this comment

xiexinch Jan 5, 2023

Choose a reason for hiding this comment

xiexinch Jan 5, 2023

Choose a reason for hiding this comment

MeowZheng Jan 12, 2023

Choose a reason for hiding this comment

AI-Tianlong commented Dec 30, 2022 •

edited

Loading

Support `HieraSeg` interface on `cityscapes`