YOLO8 support #50
YOLO8 support #50
Conversation
|
Important Review skippedAuto incremental reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the You can disable this status message by setting the WalkthroughThe recent updates significantly enhance the YOLO format support within the Datumaro framework. New converter and extractor classes have been introduced, allowing for improved handling of complex annotation types, including segmentation and pose estimation. The updates also streamline the dataset management process through YAML configuration files, enhancing flexibility and clarity. Overall, these changes bolster the framework's capabilities, making it easier to work with various YOLO formats and ensuring robust data handling for machine learning tasks. Changes
Sequence Diagram(s)sequenceDiagram
participant User
participant Yolo8Importer
participant Yolo8Extractor
participant Dataset
User->>Yolo8Importer: Import YOLOv8 dataset
Yolo8Importer->>Yolo8Extractor: Detect and validate dataset files
Yolo8Extractor->>Dataset: Load annotations and media
Dataset->>User: Provide imported dataset
Thank you for using CodeRabbit. We offer it for free to the OSS community and would appreciate your support in helping us grow. If you find it useful, would you consider giving us a shout-out on your favorite social media? TipsChatThere are 3 ways to chat with CodeRabbit:
Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments. CodeRabbit Commands (invoked as PR comments)
Additionally, you can add CodeRabbit Configuration File (
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 0
Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Files ignored due to path filters (9)
tests/assets/yolo_dataset/yolo/obj_train_data/1.jpgis excluded by!**/*.jpgtests/assets/yolo_dataset/yolo8/images/train/1.jpgis excluded by!**/*.jpgtests/assets/yolo_dataset/yolo8_obb/images/train/1.jpgis excluded by!**/*.jpgtests/assets/yolo_dataset/yolo8_pose/images/train/1.jpgis excluded by!**/*.jpgtests/assets/yolo_dataset/yolo8_pose_two_values_per_point/images/train/1.jpgis excluded by!**/*.jpgtests/assets/yolo_dataset/yolo8_segmentation/images/train/1.jpgis excluded by!**/*.jpgtests/assets/yolo_dataset/yolo8_with_list_of_imgs/images/train/1.jpgis excluded by!**/*.jpgtests/assets/yolo_dataset/yolo8_with_list_of_names/images/train/1.jpgis excluded by!**/*.jpgtests/assets/yolo_dataset/yolo8_with_subset_txt/images/train/1.jpgis excluded by!**/*.jpg
Files selected for processing (27)
- datumaro/plugins/yolo_format/converter.py (6 hunks)
- datumaro/plugins/yolo_format/extractor.py (9 hunks)
- datumaro/plugins/yolo_format/format.py (1 hunks)
- datumaro/plugins/yolo_format/importer.py (1 hunks)
- datumaro/util/meta_file_util.py (1 hunks)
- tests/assets/yolo_dataset/yolo8/data.yaml (1 hunks)
- tests/assets/yolo_dataset/yolo8/labels/train/1.txt (1 hunks)
- tests/assets/yolo_dataset/yolo8_obb/data.yaml (1 hunks)
- tests/assets/yolo_dataset/yolo8_obb/labels/train/1.txt (1 hunks)
- tests/assets/yolo_dataset/yolo8_pose/data.yaml (1 hunks)
- tests/assets/yolo_dataset/yolo8_pose/labels/train/1.txt (1 hunks)
- tests/assets/yolo_dataset/yolo8_pose_two_values_per_point/data.yaml (1 hunks)
- tests/assets/yolo_dataset/yolo8_pose_two_values_per_point/labels/train/1.txt (1 hunks)
- tests/assets/yolo_dataset/yolo8_segmentation/data.yaml (1 hunks)
- tests/assets/yolo_dataset/yolo8_segmentation/labels/train/1.txt (1 hunks)
- tests/assets/yolo_dataset/yolo8_with_list_of_imgs/data.yaml (1 hunks)
- tests/assets/yolo_dataset/yolo8_with_list_of_imgs/labels/train/1.txt (1 hunks)
- tests/assets/yolo_dataset/yolo8_with_list_of_names/data.yaml (1 hunks)
- tests/assets/yolo_dataset/yolo8_with_list_of_names/labels/train/1.txt (1 hunks)
- tests/assets/yolo_dataset/yolo8_with_subset_txt/data.yaml (1 hunks)
- tests/assets/yolo_dataset/yolo8_with_subset_txt/labels/train/1.txt (1 hunks)
- tests/assets/yolo_dataset/yolo8_with_subset_txt/train.txt (1 hunks)
- tests/conftest.py (2 hunks)
- tests/unit/data_formats/init.py (1 hunks)
- tests/unit/data_formats/test_yolo_format.py (1 hunks)
- tests/utils/assets.py (1 hunks)
- tests/utils/test_utils.py (1 hunks)
Files skipped from review due to trivial changes (16)
- tests/assets/yolo_dataset/yolo8/data.yaml
- tests/assets/yolo_dataset/yolo8/labels/train/1.txt
- tests/assets/yolo_dataset/yolo8_obb/data.yaml
- tests/assets/yolo_dataset/yolo8_obb/labels/train/1.txt
- tests/assets/yolo_dataset/yolo8_pose/data.yaml
- tests/assets/yolo_dataset/yolo8_pose/labels/train/1.txt
- tests/assets/yolo_dataset/yolo8_pose_two_values_per_point/data.yaml
- tests/assets/yolo_dataset/yolo8_pose_two_values_per_point/labels/train/1.txt
- tests/assets/yolo_dataset/yolo8_segmentation/data.yaml
- tests/assets/yolo_dataset/yolo8_with_list_of_imgs/data.yaml
- tests/assets/yolo_dataset/yolo8_with_list_of_imgs/labels/train/1.txt
- tests/assets/yolo_dataset/yolo8_with_list_of_names/data.yaml
- tests/assets/yolo_dataset/yolo8_with_subset_txt/data.yaml
- tests/assets/yolo_dataset/yolo8_with_subset_txt/train.txt
- tests/unit/data_formats/init.py
- tests/utils/assets.py
Additional comments not posted (62)
tests/assets/yolo_dataset/yolo8_with_list_of_names/labels/train/1.txt (2)
1-1: Class ID format is correct.The class ID is correctly formatted as an integer.
2-2: Bounding box coordinates format is correct.The bounding box coordinates are correctly formatted and normalized.
tests/assets/yolo_dataset/yolo8_with_subset_txt/labels/train/1.txt (2)
1-1: Class ID format is correct.The class ID is correctly formatted as an integer.
2-2: Bounding box coordinates format is correct.The bounding box coordinates are correctly formatted and normalized.
tests/assets/yolo_dataset/yolo8_segmentation/labels/train/1.txt (2)
1-1: Class ID format is correct.The class ID is correctly formatted as an integer.
2-2: Polygon coordinates format is correct.The polygon coordinates are correctly formatted and normalized.
datumaro/plugins/yolo_format/format.py (3)
10-11: LGTM! The added attributes enhance file management.The addition of
LABELS_EXTandSUBSET_LIST_EXTattributes standardizes the file extensions for labels and subset lists.
14-21: LGTM! The extended configurability and organization are beneficial.The added configuration keys (
path,kpt_shape,flip_idx) and folder names (IMAGES_FOLDER_NAME,LABELS_FOLDER_NAME) enhance the configurability and organization of YOLOv8-specific data.
24-25: LGTM! The attribute enhances pose estimation capabilities.The addition of
KPT_SHAPE_FIELD_NAMEfor keypoint shape representation is beneficial for pose estimation tasks.tests/utils/test_utils.py (6)
9-12: LGTM! The class helps transition fromunittesttopytest.The
TestCaseHelperclass is a practical approach to minimize migration work fromunittesttopytest.
14-15: LGTM! The method mimicsassertTruefromunittest.TestCase.The
assertTruemethod correctly replicates the behavior ofunittest.TestCase.assertTrue.
17-18: LGTM! The method mimicsassertFalsefromunittest.TestCase.The
assertFalsemethod correctly replicates the behavior ofunittest.TestCase.assertFalse.
20-21: LGTM! The method mimicsassertEqualfromunittest.TestCase.The
assertEqualmethod correctly replicates the behavior ofunittest.TestCase.assertEqual.
23-27: LGTM! The method mimicsassertListEqualfromunittest.TestCase.The
assertListEqualmethod correctly replicates the behavior ofunittest.TestCase.assertListEqual.
29-30: LGTM! The method mimicsfailfromunittest.TestCase.The
failmethod correctly replicates the behavior ofunittest.TestCase.fail.tests/conftest.py (2)
26-29: LGTM! The fixture enhances test setup.The
test_dirfixture uses theTestDircontext manager to create a temporary directory for tests, improving test setup and teardown.
32-34: LGTM! The fixture improves test management.The
helper_tcfixture provides access toTestCaseHelperfor all tests within a class, improving test management and execution.datumaro/util/meta_file_util.py (2)
53-55: Ensure consistency in label categories.The addition of
label_categoriesis correct. However, ensure that all labels have a corresponding parent category. If not, consider handling cases wherelabel.parentmight beNone.
56-60: Handle potential missing attributes in point categories.The handling of
point_categoriesis correct. However, ensure that allcat.labelsandcat.jointsare present and valid. Consider adding checks or handling cases where these attributes might be missing or malformed.datumaro/plugins/yolo_format/importer.py (8)
24-25: Ensure the presence ofobj.datafile.The
detectmethod correctly requires the presence of theobj.datafile. Ensure that this file is always present in the expected format.
28-29: Recursive search for sources.The
find_sourcesmethod correctly performs a recursive search for sources with the.dataextension. Ensure that this extension is consistently used for YOLO datasets.
36-47: Ensure correct YAML parsing and field validation.The
detectmethod correctly requires the presence of thedata.yamlfile and checks for the absence of theKPT_SHAPE_FIELD_NAMEfield. Ensure that the YAML file is always in the expected format and handle potential parsing errors gracefully.
49-51: Recursive search for sources.The
find_sourcesmethod correctly performs a recursive search for sources with the.yamlextension. Ensure that this extension is consistently used for YOLOv8 datasets.
54-55: Correctly set the extractor.The
EXTRACTORis correctly set toYolo8SegmentationExtractor. Ensure that this extractor is implemented correctly and handles segmentation data as expected.
58-59: Correctly set the extractor.The
EXTRACTORis correctly set toYolo8ObbExtractor. Ensure that this extractor is implemented correctly and handles oriented bounding box data as expected.
62-63: Correctly set the extractor.The
EXTRACTORis correctly set toYolo8PoseExtractor. Ensure that this extractor is implemented correctly and handles pose estimation data as expected.
66-77: Ensure correct YAML parsing and field validation for pose estimation.The overridden
detectmethod correctly requires the presence of thedata.yamlfile and checks for the presence of theKPT_SHAPE_FIELD_NAMEfield. Ensure that the YAML file is always in the expected format and handle potential parsing errors gracefully.datumaro/plugins/yolo_format/converter.py (15)
113-115: Ensure subset naming consistency.The
applymethod correctly handles subset naming. Ensure that subset names are consistent and do not clash with reserved keywords.
119-122: Create subset folders.The
applymethod correctly creates subset folders for images and annotations. Ensure that these directories are always created successfully and handle potential errors gracefully.
127-134: Export media and annotations.The
applymethod correctly exports media and annotations. Ensure that the exported files are in the expected format and handle potential errors gracefully.
143-149: Save configuration files.The
_save_config_filesmethod correctly saves configuration files. Ensure that the configuration files are always saved successfully and handle potential errors gracefully.
163-173: Create subset list file.The
_make_subset_list_filemethod correctly creates a subset list file. Ensure that the file is always created successfully and handle potential errors gracefully.
175-192: Export media.The
_export_mediamethod correctly exports media. Ensure that the exported media files are in the expected format and handle potential errors gracefully.
196-212: Export item annotations.The
_export_item_annotationmethod correctly exports item annotations. Ensure that the exported annotations are in the expected format and handle potential errors gracefully.
216-225: Create annotation line.The
_make_annotation_linemethod correctly creates annotation lines for bounding boxes. Ensure that the annotation lines are always created successfully and handle potential errors gracefully.
267-278: Save configuration files for YOLOv8.The
_save_config_filesmethod correctly saves configuration files for YOLOv8. Ensure that the configuration files are always saved successfully and handle potential errors gracefully.
279-281: Create image subset folder for YOLOv8.The
_make_image_subset_foldermethod correctly creates image subset folders for YOLOv8. Ensure that these directories are always created successfully and handle potential errors gracefully.
283-285: Create annotation subset folder for YOLOv8.The
_make_annotation_subset_foldermethod correctly creates annotation subset folders for YOLOv8. Ensure that these directories are always created successfully and handle potential errors gracefully.
290-302: Create annotation line for segmentation.The
_make_annotation_linemethod correctly creates annotation lines for segmentation annotations. Ensure that the annotation lines are always created successfully and handle potential errors gracefully.
307-313: Create annotation line for oriented bounding boxes.The
_make_annotation_linemethod correctly creates annotation lines for oriented bounding box annotations. Ensure that the annotation lines are always created successfully and handle potential errors gracefully.
317-353: Save configuration files for pose estimation.The
_save_config_filesmethod correctly saves configuration files for pose estimation. Ensure that the configuration files are always saved successfully and handle potential errors gracefully.
354-380: Create annotation line for pose estimation.The
_make_annotation_linemethod correctly creates annotation lines for pose estimation annotations. Ensure that the annotation lines are always created successfully and handle potential errors gracefully.datumaro/plugins/yolo_format/extractor.py (10)
8-18: LGTM! New imports are necessary.The new imports for
math,yaml, andcached_propertyare necessary for the new functionality and refactoring.
Line range hint
52-117:
LGTM! Refactored configuration management improves flexibility.The
_configproperty now uses a YAML configuration file, enhancing flexibility in configuration parsing and error handling.
122-129: LGTM! New method_iterate_over_image_pathsimproves modularity.The method helps in iterating over image paths, improving code modularity and readability.
246-252: LGTM! Refactored_parse_annotationsmethod enhances error reporting.The refactoring modularizes annotation loading logic and improves error reporting.
253-279: LGTM! New method_load_one_annotationimproves clarity.The method modularizes the logic for loading individual annotations, improving code clarity and maintainability.
Line range hint
280-302:
LGTM! Refactored_load_categoriesmethod enhances category management.The refactoring accommodates different formats of category definitions, enhancing the ability to parse and manage category data.
318-389: LGTM! New classYolo8Extractorenhances framework capabilities.The class introduces support for YOLOv8 formats, enhancing the framework's capabilities.
391-419: LGTM! New classYolo8SegmentationExtractorenhances segmentation support.The class introduces support for segmentation annotations, enhancing the framework's capabilities.
421-459: LGTM! New classYolo8ObbExtractorenhances oriented bounding box support.The class introduces support for oriented bounding boxes, enhancing the framework's capabilities.
461-601: LGTM! New classYolo8PoseExtractorenhances pose estimation support.The class introduces support for pose estimation annotations, enhancing the framework's capabilities.
tests/unit/data_formats/test_yolo_format.py (10)
1-61: LGTM! New imports are necessary.The new imports for
yaml,Yolo8Converter,Yolo8ObbConverter,Yolo8PoseConverter,Yolo8SegmentationConverter,Yolo8Extractor,Yolo8ObbExtractor,Yolo8PoseExtractor,Yolo8SegmentationExtractor,Yolo8Importer,Yolo8ObbImporter,Yolo8PoseImporter,Yolo8SegmentationImporterare necessary for the new functionality and tests.
359-433: LGTM! New classYolo8ConverterTestenhances test coverage.The class introduces tests for the YOLOv8 converter, enhancing test coverage.
435-480: LGTM! New classYolo8SegmentationConverterTestenhances test coverage.The class introduces tests for the YOLOv8 segmentation converter, enhancing test coverage.
524-550: LGTM! New classYolo8ObbConverterTestenhances test coverage.The class introduces tests for the YOLOv8 oriented bounding box converter, enhancing test coverage.
552-719: LGTM! New classYolo8PoseConverterTestenhances test coverage.The class introduces tests for the YOLOv8 pose converter, enhancing test coverage.
803-821: LGTM! New classYolo8ImporterTestenhances test coverage.The class introduces tests for the YOLOv8 importer, enhancing test coverage.
823-845: LGTM! New classYolo8SegmentationImporterTestenhances test coverage.The class introduces tests for the YOLOv8 segmentation importer, enhancing test coverage.
847-867: LGTM! New classYolo8ObbImporterTestenhances test coverage.The class introduces tests for the YOLOv8 oriented bounding box importer, enhancing test coverage.
869-926: LGTM! New classYolo8PoseImporterTestenhances test coverage.The class introduces tests for the YOLOv8 pose importer, enhancing test coverage.
1046-1066: LGTM! New classYolo8ExtractorTestenhances test coverage.The class introduces tests for the YOLOv8 extractor, enhancing test coverage.
| │ │ ├── image2.jpg | ||
| │ │ ├── image3.jpg | ||
| │ │ └── ... | ||
| │ ├── valid/ # directory with images for validation subset |
There was a problem hiding this comment.
No, data.yaml says where are the images of a subset, and the folder does not have to have the same name as a subset. Tried to write it in docs in a more understandable way.
There was a problem hiding this comment.
The standard name for validation subset is "val". Nobody calls it "valid" as far as I know. @zhiltsov-max , do you have any opinion regarding that?
| │ │ ├── image2.txt | ||
| │ │ ├── image3.txt | ||
| │ │ └── ... | ||
| │ ├── valid/ # directory with annotations for validation subset |
| ``` | ||
|
|
||
| Files in directories `labels/train/` and `labels/valid/` should | ||
| contain information about labels for images in `images/train` and `images/valid` respectively. |
| ... | ||
| ``` | ||
|
|
||
| For **Oriented Bounding Box** and **Segmentation** it contains coordinates |
There was a problem hiding this comment.
I would separate OBB and Segmentation because a polygon can have more than 4 points. Does the original documentation also join them?
| @@ -0,0 +1,30 @@ | |||
| # Copyright (C) 2019-2024 Intel Corporation | |||
There was a problem hiding this comment.
@zhiltsov-max , what is the policy regarding Copyright inside Datumaro?
There was a problem hiding this comment.
I copypasted the code, as well as the header, from the upstream: https://github.com/openvinotoolkit/datumaro/blob/develop/tests/utils/test_utils.py#L467
I am not sure it will be right to change the header.
| @@ -0,0 +1,1223 @@ | |||
| import copy | |||
| @@ -7,3 +7,21 @@ class YoloPath: | |||
| DEFAULT_SUBSET_NAME = "train" | |||
| SUBSET_NAMES = ["train", "valid"] | |||
There was a problem hiding this comment.
I see also valid here. It looks strange. Is it a typo?
| SUBSET_LIST_EXT = ".txt" | ||
|
|
||
|
|
||
| class Yolo8Path(YoloPath): |
There was a problem hiding this comment.
Official name is YOLOv8 (Yolo version 8).
It looks it isn't important, but I would to have users here when they a looking for YOLOv8 support. It is critical for search engine optimization.
Let's try to add v8 instead of 8 anywhere.
There was a problem hiding this comment.
Would it be enough too have YOLOv8 in the docs, or would it be better to also rename all the formats, like yolo8_segmentation -> yolo_v8_segmentation ?
There was a problem hiding this comment.
I would rename to yolov8_segmentation.
Also, I would recommend to refactor code as well, to be consistent.
There was a problem hiding this comment.
renamed YOLO8 to YOLOv8 everywhere
| @@ -0,0 +1,324 @@ | |||
| --- | |||
| title: 'YOLO8' | |||
| class Yolo8OrientedBoxesExtractor(Yolo8Extractor): | ||
| @staticmethod | ||
| def _check_is_rectangle(p1, p2, p3, p4): | ||
| p12_angle = math.atan2(p2[0] - p1[0], p2[1] - p1[1]) |
There was a problem hiding this comment.
probably it is possible to do that without math library.
- My way is to check that opposite sides are equal and after that check a^2 + b^2 == c^2
- On the internet I have found another way. Find the center of mass and calculate distance till all points from the center mass. They should be equal.
Because the check will be run multiple times, I would recommend to choose the optimal implementation. Can you compare them?
There was a problem hiding this comment.
Compared my implementation
def _check_is_rectangle_atan2(p1, p2, p3, p4):
p12_angle = math.atan2(p2[0] - p1[0], p2[1] - p1[1])
p23_angle = math.atan2(p3[0] - p2[0], p3[1] - p2[1])
p43_angle = math.atan2(p3[0] - p4[0], p3[1] - p4[1])
p14_angle = math.atan2(p4[0] - p1[0], p4[1] - p1[1])
if abs(p12_angle - p43_angle) > 0.001 or abs(p23_angle - p14_angle) > 0.001:
return False
return Truewith your second option. I have not tried your option 1, because it will have more pows and more sqrts, therefore should be slower than option 2.
def _check_is_rectangle_2(p1, p2, p3, p4):
center_x = (p1[0] + p2[0] + p3[0] + p4[0]) / 4
center_y = (p1[1] + p2[1] + p3[1] + p4[1]) / 4
distances_squared = [
(x - center_x) ** 2 + (y - center_y) ** 2
for x, y in (p1, p2, p3, p4)
]
if math.sqrt(max(distances_squared)) - math.sqrt(min(distances_squared)) > 0.1:
return False
return TrueResults:
%%timeit
_check_is_rectangle_atan2(f_p1, f_p2, f_p3, f_p4)475 ns ± 6.64 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
%%timeit
_check_is_rectangle_2(f_p1, f_p2, f_p3, f_p4)1.52 μs ± 9.1 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
I tried to do it without a list, did not hrlp much.
I have no idea why atan2 implementation is faster
There was a problem hiding this comment.
The best I managed is that:
def _check_is_rectangle_vector(p1, p2, p3, p4):
# check vectors p1->p2 and p3->p4 sum to 0
if abs(p2[0] - p1[0] + p4[0] - p3[0]) > 0.001 or abs(p2[1] - p1[1] + p4[1] - p3[1]) > 0.001:
return False
# check vectors p2->p3 and p4->p1 sum to 0
if abs(p3[0] - p2[0] + p1[0] - p4[0]) > 0.001 or abs(p3[1] - p2[1] + p1[1] - p4[1]) > 0.001:
return False
return True338 ns ± 5.1 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
In case if points do not form a rectangle it works even faster. But this code also is less understandable, so I am not sure it would be justifiable to use it instead of atan2 version
| config_file=None, | ||
| **kwargs, | ||
| ) -> None: | ||
| super().__init__(*args, **kwargs) |
There was a problem hiding this comment.
@Eldies , do we have a test for custom config name for the extractor?
There was a problem hiding this comment.
|
|
||
| ```bash | ||
| datum create | ||
| datum import --format yolo8 <path/to/dataset> # for Detection dataset |
There was a problem hiding this comment.
| datum import --format yolo8 <path/to/dataset> # for Detection dataset | |
| datum import --format yolov8 <path/to/dataset> # for Detection dataset |
see yolov8.org (it is now ultralitics spells the name)
|
Summary
adds support for yolo8 formats - detection, segmentation, pose, oriented bounding boxes
How to test
Checklist
developbranchLicense
Feel free to contact the maintainers if that's a concern.
Summary by CodeRabbit
New Features
Bug Fixes
Documentation
Tests