Corrupt JPEG data

[ashaheedq]

Search before asking

• [X] I have searched the Autodistill issues and found no similar bug report.

Bug

Hi

I am following this tutorial: https://medium.com/@corpy.ai.lab/autodistill-automating-dataset-labeling-for-efficient-model-training-6aed5f63bea

and I am facing an error with the converted jpg images. This is from helpers.py's split_data function. When I am done running this snippet:

        self.base_model.label(
            input_folder=config.image_dir,
            output_folder=config.dataset_dir,
            extension=".png",
        )

I run the following snippet:

        sv.DetectionDataset.from_yolo(
            images_directory_path=config.image_path,
            annotations_directory_path=config.annotation_path,
            data_yaml_path=config.data_yaml_path,
        )

I get this error:

Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: 1 extraneous bytes before marker 0xd9
Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: premature end of data segment
Corrupt JPEG data: 3 extraneous bytes before marker 0xd9

If I use the original png images it works fine. I assume the conversion happens to compress the images? as far as I know yolo supports png images.

Environment

• OS: macos 14.1
• autodistil: 0.1.15
• supervision: 0.16.0
• Python: 3.8.13

Minimal Reproducible Example

import supervision as sv
from tqdm import tqdm
from pathlib import Path
from typing import Union
from autodistill_yolov8 import YOLOv8
from omegaconf import OmegaConf, DictConfig
from autodistill.detection import CaptionOntology
from autodistill_grounded_sam import GroundedSAM
from autodistill_grounding_dino import GroundingDINO

class AUTODISTILLATION:
    def __init__(self, config: Union[str, DictConfig]) -> None:
        if isinstance(config, DictConfig):
            self.config = config
        else:
            self.config = OmegaConf.load(config)
        self.onotology = CaptionOntology(self.config.labels)
        self.base_model = GroundingDINO(ontology=self.onotology)
        self.target_model = YOLOv8(self.config.target_model)

    @staticmethod
    def extract_frame(config: DictConfig):
        train_video_path = [config.train_video_path]
        image_dir = config.image_dir
        frame_stride = config.frame_stride
        for video_path in tqdm(train_video_path):
            if isinstance(video_path, str):
                video_path = Path(video_path)
            video_name = video_path.stem
            image_name_pattern = video_name + "-{:05d}.png"
            with sv.ImageSink(
                target_dir_path=image_dir, image_name_pattern=image_name_pattern
            ) as sink:
                for image in sv.get_video_frames_generator(
                    source_path=str(video_path), stride=frame_stride
                ):
                    sink.save_image(image=image)

    def label_dataset(self, config: DictConfig):
        print("[INFO] Start labelling")
        self.base_model.label(
            input_folder=config.image_dir,
            output_folder=config.dataset_dir,
            extension=".png",
        )

    def dataset_to_yolo(self, config: DictConfig):
        print("[INFO] Converting dataset to yolo format")
        sv.DetectionDataset.from_yolo(
            images_directory_path=config.image_path,
            annotations_directory_path=config.annotation_path,
            data_yaml_path=config.data_yaml_path,
        )

    def make_dataset(self, config: DictConfig):
        print("[INFO] Making dataset start ...")
        image_dir = Path(config.image_dir)
        if not any(image_dir.iterdir()):
            self.extract_frame(config)
        self.label_dataset(config)
        self.dataset_to_yolo(config)

    def train(self, config: DictConfig):
        print("[INFO] Training ....")
        self.target_model.train(
            config.data_yaml_path, epochs=config.epochs
        )

    def inference(self, source, conf):
        self.target_model.predict(source, conf)

if __name__ == "__main__":
    config_path = "config.yaml"
    config = OmegaConf.load(config_path)
    distillation = AUTODISTILLATION(config=config)
    distillation.make_dataset(config=distillation.config)
    # distillation.train(config=distillation.config)
    # distillation.inference(source=distillation.config.test_video_path, conf=distillation.config.conf)

Additional

No response

Are you willing to submit a PR?

• [X] Yes I'd like to help by submitting a PR!

[capjamesg]

Hello! Thank you for creating this Issue. Is it possible for you to share an example image from your dataset that raised the Corrupt JPEG data warning?

[capjamesg]

@ashaheedq Do you still require assistance? If not, can we close this issue?