Skip to main content
Version: 0.6.12

Dataset Building

Starwhale provides a highly flexible method to build datasets, allowing you to build dataset from various file types including images, audio, video, CSV, JSON, and JSONL files. Python scripts and datasets from the Huggingface Hub can also be used for construction.

Building from Data Files​

Image​

Starwhale supports recursively traversing image files within directories to build a dataset without any coding:

  • Supported image formats: png, jpg, jpeg, webp, svg, apng.
  • Images are converted to Starwhale.Image type and can be viewed in the Starwhale Server Web page.
  • Supported by swcli dataset build --image command line and starwhale.Dataset.from_folder Python SDK.
  • Label mechanism: when SDK sets auto_label=True or command line sets --auto-label, the parent directory name will be used as the label.
  • Metadata mechanism: dataset columns can be expanded by setting metadata.csv or metadata.jsonl files in the root directory.
  • Caption mechanism: when {image-name}.txt files are found in the same directory, the content will be automatically imported and populated into the caption column.

Assuming there are the following four files in the folder directory:

folder/dog/1.png
folder/cat/2.png
folder/dog/3.png
folder/cat/4.png

Command line construction:

❯ swcli dataset build --image folder --name image-folder
🚧 start to build dataset bundle...
👷 uri local/project/self/dataset/image-folder/version/latest
🌊 creating dataset local/project/self/dataset/image-folder/version/uw6mdisnf7alg4t4fs2myfug4ie4636w3x4jqcu2...
đŸĻ‹ update 4 records into dataset
đŸŒē congratulation! you can run swcli dataset info image-folder/version/uw6mdisnf7al
❯ swcli dataset head image-folder -n 2
row ───────────────────────────────────────
đŸŒŗ id: cat/2.png
🌀 features:
🔅 file_name : cat/2.png
🔅 label : cat
🔅 file : ArtifactType.Image, display:2.png, mime_type:MIMEType.PNG, shape:[None, None, 3], encoding:
row ───────────────────────────────────────
đŸŒŗ id: cat/4.png
🌀 features:
🔅 file_name : cat/4.png
🔅 label : cat
🔅 file : ArtifactType.Image, display:4.png, mime_type:MIMEType.PNG, shape:[None, None, 3], encoding:

Python SDK construction:

from starwhale import Dataset
ds = Dataset.from_folder("folder", kind="image")
print(ds)
print(ds.fetch_one().features)
🌊 creating dataset local/project/self/dataset/folder/version/nyc2ay4gnyayv4zqalovdgakl3k2esvrne42cjna...
đŸĻ‹ update 4 records into dataset
Dataset: folder, stash version: d22hdiwyakdfh5xitcpn2s32gblfbhrerzczkb63, loading version: nyc2ay4gnyayv4zqalovdgakl3k2esvrne42cjna
{'file_name': 'cat/2.png', 'label': 'cat', 'file': ArtifactType.Image, display:2.png, mime_type:MIMEType.PNG, shape:[None, None, 3], encoding: }

Video​

Recursive traversal of video files in a directory to construct Starwhale datasets without any coding:

  • Supported video formats: mp4, webm and avi.
  • Videos are converted to Starwhale.Video types and can be viewed in the Starwhale Server Web page.
  • Supported by swcli dataset build --video command line and starwhale.Dataset.from_folder Python SDK.
  • Label, caption and metadata mechanisms are the same as for images.

Audio​

Recursive traversal of audio files in a directory to construct Starwhale datasets without any coding:

  • Supported audio formats: mp3 and wav.
  • Audio is converted to Starwhale.Audio types and can be viewed in the Starwhale Server Web page.
  • Supported by swcli dataset build --audio command line and starwhale.Dataset.from_folder Python SDK.
  • Label, caption and metadata mechanisms are the same as for images.

csv Files​

Command line or Python SDK can directly convert local or remote csv files into Starwhale datasets:

  • Support one or more local csv files.
  • Support recursive finding of csv files in a local directory.
  • Support one or more remote csv files specified by http urls.

Command line construction:

❯ swcli dataset build --name product-desc-modelscope --csv https://modelscope.cn/api/v1/datasets/lcl193798/product_description_generation/repo\?Revision\=master\&FilePath\=test.csv --encoding=utf-8-sig
🚧 start to build dataset bundle...
👷 uri local/project/self/dataset/product-desc-modelscope/version/latest
🌊 creating dataset local/project/self/dataset/product-desc-modelscope/version/wzaz4ccodpyj4jelgupljreyida2bleg5xp7viwe...
đŸĻ‹ update 3848 records into dataset
đŸŒē congratulation! dataset build from csv files(('https://modelscope.cn/api/v1/datasets/lcl193798/product_description_generation/repo?Revision=master&FilePath=test.csv',)) has been built. You can run swcli dataset info product-desc-modelscope/version/wzaz4ccodpyj

Python SDK construction:

from starwhale import Dataset
ds = Dataset.from_csv(path="http://example.com/data.csv", name="my-csv-dataset")

json/jsonl Files​

Command line or Python SDK can directly convert local or remote json/jsonl files into Starwhale datasets:

  • Support one or more local json/jsonl files.
  • Support recursive finding of json/jsonl files in a local directory.
  • Support one or more remote json/jsonl files specified by http urls.

For JSON files:

  • By default, the parsed json object is assumed to be a list, and each object in the list is a dict, which maps to one row in the Starwhale dataset.
  • The --field-selector or field_selector parameter can be used to locate a specific list.

For example, for the json file:

{
"p1": {
"p2":{
"p3": [
{"a": 1, "b": 2},
{"a": 10, "b": 20}
]
}
}
}

Set --field-selector=p1.p2.p3 to accurately add two rows of data to the dataset.

Command line construction:

❯ swcli dataset build --json https://modelscope.cn/api/v1/datasets/damo/100PoisonMpts/repo\?Revision\=master\&FilePath\=train.jsonl
🚧 start to build dataset bundle...
👷 uri local/project/self/dataset/json-b0o2zsvg/version/latest
🌊 creating dataset local/project/self/dataset/json-b0o2zsvg/version/q3uoziwqligxdggncqywpund75jz55h3bne6a5la...
đŸĻ‹ update 906 records into dataset
đŸŒē congratulation! dataset build from ('https://modelscope.cn/api/v1/datasets/damo/100PoisonMpts/repo?Revision=master&FilePath=train.jsonl',) has been built. You can run swcli dataset info json-b0o2zsvg/version/q3uoziwqligx

Python SDK construction:

from starwhale import Dataset
myds = Dataset.from_json(
name="translation",
text='{"content": {"child_content": [{"en":"hello","zh-cn":"äŊ åĨŊ"},{"en":"how are you","zh-cn":"最čŋ‘æ€Žäšˆæ ˇ"}]}}',
field_selector="content.child_content"
)
print(myds[0].features["zh-cn"])
🌊 creating dataset local/project/self/dataset/translation/version/kblfn5zh4cpoqxqbhgdfbvonulr2zefp6lojq44y...
đŸĻ‹ update 2 records into dataset
äŊ åĨŊ

Building from Huggingface Hub​

There are numerous datasets available on the Huggingface Hub, which can be converted into Starwhale Dataset with a single line of code or command.

tip

Huggingface Datasets conversion relies on the datasets library.

Command line:

swcli dataset build -hf lambdalabs/pokemon-blip-captions --name pokemon

Python SDK:

from starwhale import Dataset

# You only specify starwhale dataset expected name and huggingface repo name
# example: https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions
ds = Dataset.from_huggingface("pokemon", "lambdalabs/pokemon-blip-captions")
print(ds)
print(len(ds))
print(repr(ds.fetch_one()))

Building from Python SDK scripts​

The Starwhale Dataset SDK provides a way similar to Python dict to add or update data, enabling the creation and update of local or remote datasets.

Starwhale defines two attributes for each row of data: key and features.

  • key: int or str type. There is only one type of key in a dataset. key indicates the unique index of that row of data.
  • features: dict type. Starwhale Dataset adopts a schema-free design, so the features structure of each row can be different.
    • features data supports Python constant types like int, float, str, as well as Starwhale types like Image, Video, Audio, Text, and Binary. It also supports Python compound types like list, tuple, dict.

Dataset Initialization​

To create, update, or load a dataset, you need to get a Starwhale.Dataset object, usually in the following ways:

from starwhale import dataset

# Create a dataset named new-test in standalone instance. If it exists, raise an exception.
local_ds = dataset("new-test", create="empty")
print(local_ds)
print(len(local_ds))

# If the mnist64 dataset does not exist, create one; otherwise, load this existing dataset.
remote_ds = dataset("https://cloud.starwhale.cn/project/starwhale:helloworld/dataset/mnist64", create="auto")
print(remote_ds)
print(len(remote_ds))

# Load the existing dataset named mnist64, and if it does not exist, an error will be raised.
existed_ds = dataset("mnist64", create="forbid")
print(existed_ds)
print(len(existed_ds))
Dataset: new-test, stash version: y4touw3btifhkd4f2gg4x3qvydgnfmvoztqqm5cf, loading version: y4touw3btifhkd4f2gg4x3qvydgnfmvoztqqm5cf
0

Dataset: mnist64, stash version: 4z5wpbpozsxlelma3j6soeatekufymnyxdeihoqo, loading version: vs3gnaauakidjcc5effevaoh63vivu7dzodo5cmc
500

Dataset: mnist64, stash version: 3ahtfbizw63myxcz34ebd72lhgc25dualcmtznts, loading version: lwhvvixpimlsghfs2xqmtgrwti4yn2z5nevz7hth
500

Adding and Updating Dataset Elements​

After adding data, calling commit will generate a new version that can then be used to access the dataset.

The append Method​

The Dataset provides the append function, which automatically adds features to a new row in the dataset when called.

from starwhale import dataset
ds = dataset("new-test", create="empty")

# key is the auto increment index. The example key is zero.
ds.append({"a": 0, "b": 0})

# Keys in the dataset can also be explicitly declared, but they must maintain consistency with the key types of other rows.
# When data is added in the form of a list or tuple, the first element (at index 0) represents the key for that particular row, while the second element (at index 1) contains the corresponding features.
ds.append((1, {"a":1, "b":1}))

ds.commit()

__setitem__ Method​

The Dataset's __setitem__ method provides a dict-like way to add data by index.

ds[2] = {"a":2, "b":2}
ds.commit()

Building from Python Handler​

Supports reading functions in Python files through the swcli command line as input to build datasets. The return value of the function needs to be iterable.

Example python script dataset.py:

def iter_item():
for i in range(100):
# only return features. key is auto increment index.
yield {"a": i, "b": i}

def iter_item_with_key():
for i in range(100):
# key + features
yield i, {"a": i, "b": i}

Build datasets by triggering through the swcli command line:

swcli dataset build --handler dataset:iter_item --name test1
swcli dataset build --handler dataset:iter_item_with_key --name test2