Export a Dataset Version

Export data from Roboflow for training.

About

Exporting produces a downloadable snapshot of a dataset version in the annotation format you choose - YOLO, COCO, Pascal VOC, CreateML, and many others (see the full list in the formats directory). You can export from the Roboflow web app, the Python SDK, the REST API, or the CLI. Because dataset versions are built for training, exported images are compressed and class names are sanitized; see Export Behavior for details and for how to retrieve original-quality images.

Web App

You can export data from Roboflow at any time from the web interface.

To export data, first generate a dataset version in the Roboflow dashboard. You can do so on the "Versions" page associated with your project.

After you have generated a dataset, click "Export" next to your dataset version:

You can download your data in a wide variety of formats. You can see a full list of supported export formats in the "Export" tab of our formats directory.

After selecting an export format, you can choose to either download the data as a .zip file, or as a curl link to download from the command line.

Exporting to a .zip folder on your device.

The curl and Python code will contain a private key unique to your account. Do not share this key!

The window that appears for "show download code" after selecting "Continue."

HTTP API

/:workspace/:project/:version/:format is the route you should use to get the download link for an exported dataset in a specific format. You can use this in the Jupyter notebooks from our model library or your own custom training scripts.

The following endpoint returns an export value that contains a link key with a URL from which you can download a dataset:

curl "https://api.roboflow.com/roboflow/chess-sample-4ckfl/1/yolov5pytorch?api_key=$ROBOFLOW_API_KEY"

Here is an example payload returned by the endpoint:

{
    "workspace": {
        "name": "Roboflow",
        "url": "roboflow",
        "members": 7
    },
    "project": {
        "id": "roboflow/chess-sample-4ckfl",
        "type": "object-detection",
        "name": "Chess Sample",
        "created": 1630335544.592,
        "updated": 1630335741.988,
        "images": 12,
        "unannotated": 3,
        "annotation": "pieces",
        "public": false,
        "splits": {
            "test": 1,
            "train": 9,
            "valid": 2
        },
        "classes": {
            "white-bishop": 11,
            "black-king": 8,
            "black-knight": 11,
            "white-queen": 7,
            "black-bishop": 8,
            "white-rook": 10,
            "black-rook": 10,
            "white-king": 8,
            "black-queen": 4,
            "black-pawn": 37,
            "white-pawn": 34,
            "white-knight": 10
        }
    },
    "version": {
        "id": "roboflow/chess-sample-4ckfl/1",
        "name": "augmented",
        "created": 1630335698.746,
        "images": 30,
        "splits": {
            "train": 27,
            "test": 1,
            "valid": 2
        },
        "model": {
            "id": "chess-sample-4ckfl/1",
            "endpoint": "https://serverless.roboflow.com/infer/chess-sample-4ckfl/1",
            "start": 1630335799.682,
            "end": 1630337523.889,
            "fromScratch": false,
            "tfjs": true,
            "oak": true,
            "map": "62.87",
            "recall": "85.29",
            "precision": "23.44"
        },
        "preprocessing": {
            "grayscale": {
                "enabled": true
            },
            "resize": {
                "width": 416,
                "height": 416,
                "enabled": true,
                "format": "Stretch to"
            },
            "auto-orient": {
                "enabled": true
            }
        },
        "augmentation": {
            "rotate": {
                "enabled": true,
                "degrees": "5"
            },
            "exposure": {
                "enabled": true,
                "percent": "25"
            },
            "noise": {
                "enabled": true,
                "percent": "2"
            },
            "image": {
                "versions": "3",
                "enabled": true
            },
            "flip": {
                "horizontal": true,
                "enabled": true,
                "vertical": false
            },
            "brightness": {
                "enabled": true,
                "brighten": true,
                "percent": "25",
                "darken": true
            },
            "crop": {
                "percent": 30,
                "enabled": true,
                "min": 0
            }
        },
        "exports": [
            "yolov5pytorch"
        ]
    },
    "export": {
        "format": "yolov5pytorch",
        "link": "https://app.roboflow.com/ds/XXXXXXXXXX?key=XXXXXXXXXX"
    }
}

Export Format Options

Here are the settings options available for dataset export:

Object Detection	Single-Label Classification	Multi-Label Classification	Instance Segmentation	Semantic Segmentation	Keypoint Detection
clip	folder	multiclass	coco-segmentation	coco-segmentation	coco
coco	clip	folder	clip	png-mask-semantic	yolov5pytorch
createml		clip	coco
darknet			createml
multiclass			darknet
tensorflow			multiclass
tfrecord			tensorflow
voc			tfrecord
yolokeras			voc
yolov5pytorch			yolokeras
yolov7pytorch			yolov4pytorch
mt-yolov6			yolov4scaled
retinanet			yolov5-obb
benchmarker			yolov5pytorch
			yolov7pytorch
			mt-yolov6
			retinanet
			benchmarker

Python SDK

You can both generate versions and export datasets with the Python package.

To create a ZIP file of a dataset for export from the Python SDK, begin by retrieving a specific version from a project:

version = project.version(version_number)

Then download the dataset directly:

version.download(model_format="yolov5", location="./downloads")

The download method handles export creation automatically if needed, extracts the ZIP file to your specified location, and returns an object describing the dataset.

For generating a version before export, see Create a Dataset Version.

Download Original-Quality Images

Exported images are compressed for training (see Export Behavior). To download the original, full-resolution images for an entire dataset, use the Image Search API:

import os
import requests
from roboflow import Roboflow

rf = Roboflow(api_key="YOUR_API_KEY")

project = rf.project("my-dataset-id")

records = []

for page in project.search_all(
    offset = 0,
    limit = 100,
    in_dataset = True,
    batch = False,
    fields = ["id", "name", "owner"],
):
    records.extend(page)

print(f"{len(records)} images found")

for record in records:
        base_url = "https://source.roboflow.com"
        url = f"{base_url}/{record['owner']}/{record['id']}/original.jpg"

        try:
            response = requests.get(url)
            response.raise_for_status()

            # Save to temp directory
            save_path = os.path.join('temp_images', record['name'])
            with open(save_path, 'wb') as f:
                f.write(response.content)

            print(f"Downloaded: {record['name']}")

        except requests.exceptions.RequestException as e:
            print(f"Error downloading image: {e}")

CLI

You can export data through the Roboflow CLI using the following command:

roboflow download <datasetUrl>

Find more information about this CLI command in our docs.

Export Behavior

Dataset versions are designed to be used as training data for computer vision models. Therefore, we make some optimizations to improve the training experience and performance of the models.

Image Compression

To prevent training slowdowns, we compress images at a level that maintains a balance between training speed and resolution needed for sufficient model performance.

If you're looking to download the original quality image, you can do so by clicking on a image on your dataset and selecting "Download Image".

You can also access your images programmatically via the Image Details API. The image.urls.original property states the link to the original quality image.

To download every original-quality image in a dataset programmatically, see Download Original-Quality Images in the Python SDK section.

Accepted Characters

To prevent issues from arising during training, we sanitize class names both at upload/import and export. At export, we perform the following:

• Class names are converted to ASCII

  • Where possible, characters are anglicized (ex: ü to u)

  • Otherwise, they are replaced with a dash (-)

Class name sanitization also occurs during upload