Download Datasets From Highlighter

The Highlighter SDK allows you to download your Datasets from your Highlighter account and save it in some common formats.

When converting to common formats sucs as Coco or Yolo things like entity_id will not be preserved. Only the information nessessary for training will end up in the resulting saved dataset. If you want to save a dataset localled and not loose this information you must use the hdf or json format.

CLI

hl dataset read --help
Usage: hl dataset read [OPTIONS] COMMAND [ARGS]...

Options:
  -i, --dataset-ids TEXT  integet <id> or <id>:<split>
  --page-size INTEGER     [default: 200]
  --help                  Show this message and exit.

Commands:
  coco
  hdf
  yolo

Use --help to see the format specific cli options

For example, the following will:

download dataset 123 and 456
save the images to /my/image/cache/
save the annotations as a coco dataset to my_dataset/

hl dataset read -i 123:train -i 456:test coco --annotations-dir my_dataset/ --data-file-dir /my/image/cache/

ls my_dataset/
> test.json train.json

Use --help to see the format specific cli Options

hl dataset read -i 123 yolo --help

Python API

The following will do same download and conversion as in the CLI example

from pathlib import Path
from highlighter.datasets import Dataset 
from highlighter.datasets.formats.coco import CocoWriter
from highlighter import HLClient

client = HLClient.from_env()

train_ds = Dataset.read_highlighter_dataset_assessments(
        client, 123
        )
train_ds.data_files_df.loc[:, "split"] = "train"

test_ds = Dataset.read_highlighter_dataset_assessments(
        client, 456
        )
test_ds.data_files_df.loc[:, "split"] = "test"

combined_ds = Dataset.combine([train_ds, test_ds])

annotations_dir = Path("my_dataset/")
writer = CocoWriter(annotations_dir)
writer.write(combined_ds)

images_dir = Path("/my/image/cache/")
Dataset.download_dataset_files(
        client,
        images_dir,
        combined_ds.data_files_df,
        )