cvat-data-flow 0.1.0

A utility for working with CVAT data flow

CVAT DATA FLOW

New

• Add support for export in YOLOv8 segmentation and detection formats (for segmentation dataset_format = "yolo_seg")
• Add support for installation via pip
• Add support for downloading multiple tasks from different projects

Overview

This utility facilitates the download of datasets from CVAT, preparing them for training and testing purposes. Key features include:

• Downloading datasets from CVAT in any supported format.
• Merging multiple tasks to form a consolidated dataset.
• Randomly segmenting the dataset into training, validation, and testing subsets.

Local Installation

Clone the repository:

git clone https://github.com/ai-iaguzhinskii/cvat_data_flow.git

Install the dependencies:

pip3 install -r requirements.txt

Configuration

Edit the configuration as per your requirements in the config_file:

[CVAT]
URL = # URL of the CVAT server, e.g., http://localhost:8080
LOGIN = # CVAT user login, e.g., admin
PASS = # CVAT user password, e.g., admin

[DOWNLOAD]
TASKS_IDS = # Task IDs for downloading, e.g., [111, 222, 333]
PROJECTS_IDS = # Project IDs for downloading, e.g., [111, 222, 333]

[DATASET]
FORMAT = # Desired dataset format. Refer to CVAT documentation for options, e.g., coco
# For YOLOv8 segmentation and detection formats use "yolo_seg" and "yolo_det" respectively
SAVE_PATH = # Directory to save the downloaded datasets, e.g., /home/user/datasets
SPLIT = # Define random splits for the dataset, e.g., {"train": 0.8, "val": 0.1, "test": 0.1}

[OPTIONS]
ONLY_BUILD_DATASET = # Toggle to build dataset only if tasks are previously downloaded, e.g., True
LABELS_MAPPING = # Define label mappings, e.g., {"car": "vehicle", "person": "pedestrian"}
DEBUG = # Toggle debug mode, e.g., True

Usage

To run the program:

python3 cvat_data_flow/main.py

Note: Ensure to replace placeholders in the configuration with actual values before executing.

Installation via PIP

The utility can be installed via pip:

pip3 install cvat-data-flow

Usage as a Python Package

The utility can also be used as a Python package. Here's an example:

from cvat_data_flow import CVATDataFlow

cvat_data_flow = CVATDataFlow(
        url='http://cvat.example.com',
        login='username',
        password='password',
        save_path='/path/to/save/dataset',
        projects_ids=[1, 2, 3],
        only_build_dataset=False,
        dataset_format='coco',
        split=[('train', 0.7), ('val', 0.2), ('test', 0.1)],
        labels_mapping={'person': 'person', 'car': 'vehicle'}
    )

cvat_data_flow.download_data()
cvat_data_flow.build_dataset()

---

License

This project is licensed under the MIT License - see the LICENSE file for details.