coco2yolo-kaggle 0.2.1

Convert COCO dataset to YOLO format in Kaggle environment

COCO2YOLO Kaggle

A tool for converting COCO dataset to YOLO format in Kaggle environment. This tool is based on ultralytics/JSON2YOLO and optimized for Kaggle environment.

Features

• Automatically convert COCO JSON annotations to YOLO format
• Intelligently handle Kaggle storage space limitations
• Support parallel file operations for faster processing
• Support segment annotations
• Flexible commands for different operations
• Customizable directory structure

Installation

pip install coco2yolo-kaggle

Usage

Python API

Label Conversion Only

from coco2yolo_kaggle import convert_coco_labels

# Convert labels only
labels_dir = convert_coco_labels(
    json_dir="/kaggle/input/coco-2017-dataset/coco2017/annotations",
    output_dir="/kaggle/working/coco_yolo",
    use_segments=True,
    cls91to80=True  # Map 91 COCO classes to 80 classes (default: True)
)

File Copy Only (with Custom Directory Names)

from coco2yolo_kaggle import copy_dataset

# Copy dataset files to final destination
dataset_path = copy_dataset(
    json_dir="/kaggle/input/coco-2017-dataset/coco2017/annotations",
    output_dir="/kaggle/working/coco_yolo",
    final_dest="/kaggle/tmp/COCO2017",
    max_workers=8,
    train_dir_name="train",  # Custom train directory name (default: "train2017")
    val_dir_name="val"       # Custom validation directory name (default: "val2017")
)

Complete Conversion Process

from coco2yolo_kaggle import convert_coco_dataset

# Complete process with custom directory names
dataset_path = convert_coco_dataset(
    json_dir="/kaggle/input/coco-2017-dataset/coco2017/annotations",
    output_dir="/kaggle/working/coco_yolo",
    final_dest="/kaggle/tmp/COCO2017",
    use_segments=True,
    cls91to80=True,  # Map 91 COCO classes to 80 classes (default: True)
    max_workers=8,
    copy_files=True,  # Set to False to skip copying files
    train_dir_name="train",  # Custom destination directory (default: "train2017")
    val_dir_name="val",      # Custom destination directory (default: "val2017")
    src_train_dir_name="train2017",  # Source train directory name
    src_val_dir_name="val2017"       # Source validation directory name
)

Command Line Usage

Convert Labels Only (Default Mode)

coco2yolo-kaggle --json-dir=/kaggle/input/coco-2017-dataset/coco2017/annotations --output-dir=/kaggle/working/coco_yolo

By default, this will map the 91 COCO classes to 80 classes. If you want to keep the original 91 classes, use:

coco2yolo-kaggle --json-dir=/kaggle/input/coco-2017-dataset/coco2017/annotations --output-dir=/kaggle/working/coco_yolo --no-cls91to80

Copy Files Only (with Custom Directory Names)

coco2yolo-kaggle --mode=copy --json-dir=/kaggle/input/coco-2017-dataset/coco2017/annotations --output-dir=/kaggle/working/coco_yolo --final-dest=/kaggle/tmp/COCO2017 --train-dir-name=train --val-dir-name=val

Complete Process with Custom Directory Names

coco2yolo-kaggle --mode=all --json-dir=/kaggle/input/coco-2017-dataset/coco2017/annotations --output-dir=/kaggle/working/coco_yolo --final-dest=/kaggle/tmp/COCO2017 --train-dir-name=train --val-dir-name=val

Class Mapping and Customizing Directory Structure

Class Mapping

• By default, the tool maps the original 91 COCO classes to the standard 80 classes (cls91to80=True)
• To preserve the original 91 classes, use the --no-cls91to80 flag

Directory Name Customization

You can customize the directory names using these parameters:

• --train-dir-name: Set destination train directory name (default: "train2017")
• --val-dir-name: Set destination validation directory name (default: "val2017")
• --src-train-dir: Source train directory name (default: "train2017")
• --src-val-dir: Source validation directory name (default: "val2017")

Kaggle Example Code

# 1. Install the package
!pip install coco2yolo-kaggle

# 2. Convert labels only (with class mapping)
!coco2yolo-kaggle --json-dir=/kaggle/input/coco-2017-dataset/coco2017/annotations --output-dir=/kaggle/working/coco_yolo

# 3. Copy dataset files with custom directory names
!coco2yolo-kaggle --mode=copy --json-dir=/kaggle/input/coco-2017-dataset/coco2017/annotations --output-dir=/kaggle/working/coco_yolo --final-dest=/kaggle/tmp/COCO2017 --train-dir-name=train --val-dir-name=val

# 4. Train a YOLOv8 model
!pip install ultralytics
from ultralytics import YOLO

# Create dataset configuration file with custom directory structure
%%writefile coco.yaml
path: /kaggle/tmp/COCO2017
train: images/train  # Using custom directory name
val: images/val      # Using custom directory name
nc: 80
names: ['person', 'bicycle', 'car', ... ] # Complete 80 class names

# Train the model
model = YOLO('yolov8n.pt')  # Use nano model
results = model.train(data='coco.yaml', epochs=3, imgsz=640)

Acknowledgements

This tool is based on ultralytics/JSON2YOLO, thanks to the original authors' contributions.