Problems with yolo loose format

[nik123]

Hello, I'm trying to get my hands on datumaro library and have some troubles with yolo format. According to the documentation it's possible to import yolo dataset in a "loose format'. I have a directory with the following structure:

yolo-ds
├── images
│   ├── 2024-01-09_13.52.57.png
│   ├── 2024-01-09_22.21.52.png
│   ├── 2024-01-10_00.44.59.png
│   ├── 2024-01-10_06.48.19.png
│   ├── 2024-01-12_01.56.08.png
│   └── 2024-01-12_09.10.30.png
├── labels
│   ├── 2024-01-09_13.52.57.txt
│   ├── 2024-01-09_22.21.52.txt
│   ├── 2024-01-10_00.44.59.txt
│   ├── 2024-01-10_06.48.19.txt
│   ├── 2024-01-12_01.56.08.txt
│   └── 2024-01-12_09.10.30.txt
└── obj.names

I'm importing the dataset with the following code:

ds = dm.Dataset.import_from("yolo-ds", "yolo_loose")
print(type(ds))
print(ds)

Dataset imports successfully and line print(type(ds)) outputs datumaro.components.dataset.Dataset. However, the third line fails with "TypeError: 'NoneType' object is not iterable" error.

The full traceback can be found here

I use Python 3.10.13 and datumaro version is 1.5.2

[vinnamkim]

Hi @nik123, Sorry for your inconvenience. I tried to reproduce it with a dummy dataset in our codebase (you can see here) such as

import datumaro as dm

print(f"Datumaro version: {dm.__version__}")

dataset = dm.Dataset.import_from("tests/assets/yolo_dataset/labels")
print(type(dataset))

However, it shows no problem as follows.

Datumaro version: 1.5.2
<class 'datumaro.components.dataset.Dataset'>

If possible, would you attach a small reproducible dataset as a zip file here? It will help us to find the problem.

[nik123]

@vinnamkim here it is https://github.com/nik123/datumaro-yolo-loose-issue/

I experience problems executing print(ds). I suppose the dataset should be printable. However, I receive errors.

[vinnamkim]

Hi, I've forgot what I implemented in the mean time, haha. You need to give format=yolo rather than format=yolo_loose even if it's yolo loose format (documentation here)

I can import the dataset you provided as follows and it looks successful.

git clone https://github.com/nik123/datumaro-yolo-loose-issue/
python test.py

# test.py
import datumaro as dm

print(f"Datumaro version: {dm.__version__}")

dataset = dm.Dataset.import_from("datumaro-yolo-loose-issue/yolo-loose-ds", format="yolo")
print(dataset)

Datumaro version: 1.5.2
Dataset
        size=6
        source_path=datumaro-yolo-loose-issue/yolo-loose-ds
        media_type=<class 'datumaro.components.media.Image'>
        annotated_items_count=6
        annotations_count=22
subsets
        train: # of items=3, # of annotated items=3, # of annotations=7, annotation types=['bbox']
        val: # of items=3, # of annotated items=3, # of annotations=15, annotation types=['bbox']
infos
        categories
        label: ['label1', 'label2', 'label3', 'label4', 'label5', 'label6', 'label7', 'label8', 'label9']

[nik123]

Thanks @vinnamkim ! This indeed helped

[nik123]

There still seems to be a slight issue regarding the behavior. Despite using format=yolo_loose, the dataset is imported successfully. Perhaps it would be more appropriate to raise an UnknownFormatError instead.

[vinnamkim]

There still seems to be a slight issue regarding the behavior. Despite using format=yolo_loose, the dataset is imported successfully. Perhaps it would be more appropriate to raise an UnknownFormatError instead.

Thanks for your good suggestion. We will put it in our backlog.

[wonjuleee]

Hi @nik123, thank you for your interests in Datumaro! yolo_loose is not the format defined in importers, but it is declared as an extractor. And Datumaro is able to import a dataset with extractor too as described in https://github.com/openvinotoolkit/datumaro/blob/f9366173a0a5ba6fe479703800a2f3a0caf15530/src/datumaro/components/dataset.py#L822

But, as you observed, importing a dataset with yolo_loose was failed and this is because of lack of arguments. That means, we need to specify more arguments to import a dataset through extractor. Please see the example below.

import os
import os.path as osp
import datumaro as dm

path = "yolo-loose-ds"
sources = []
for subset in ["train", "val"]:
        ann_path = osp.join(path, 'labels', subset)
        urls = [osp.join(ann_path, ann_file) for ann_file in os.listdir(ann_path)]
        sources.append(dm.Dataset.import_from(
                        path=path,
                        format="yolo_loose",
                        subset=subset,
                        urls=urls,
                )
        )
dataset = dm.Dataset.from_extractors(*sources)
print(dataset)

Yes, it is indeed much complex compared to the case of importer. Therefore, we recommend users to use importer because it automates parameterization! But, I want to show that the use of exatractor.

Thanks again and hope to see you again!