Types

Deep Lake provides a comprehensive type system designed for efficient data storage and retrieval. The type system includes basic numeric types as well as specialized types optimized for common data formats like images, embeddings, and text.

Each type can be specified either using the full type class or a string shorthand:

# Using type class
ds.add_column("col1", deeplake.types.Float32())

# Using string shorthand
ds.add_column("col2", "float32")

Types determine:

• How data is stored and compressed
• What operations are available
• How the data can be queried and indexed
• Integration with external libraries and frameworks

Numeric Types

All basic numeric types:

import deeplake

# Integers
ds.add_column("int8", deeplake.types.Int8())      # -128 to 127
ds.add_column("int16", deeplake.types.Int16())    # -32,768 to 32,767
ds.add_column("int32", deeplake.types.Int32())    # -2^31 to 2^31-1
ds.add_column("int64", deeplake.types.Int64())    # -2^63 to 2^63-1

# Unsigned Integers
ds.add_column("uint8", deeplake.types.UInt8())    # 0 to 255
ds.add_column("uint16", deeplake.types.UInt16())  # 0 to 65,535
ds.add_column("uint32", deeplake.types.UInt32())  # 0 to 2^32-1
ds.add_column("uint64", deeplake.types.UInt64())  # 0 to 2^64-1

# Floating Point
ds.add_column("float32", deeplake.types.Float32())
ds.add_column("float64", deeplake.types.Float64())

deeplake.types.Image

Image(
    dtype: DataType | str = "uint8",
    sample_compression: str = "png",
) -> Type

An image of a given format. The value returned will be a multidimensional array of values rather than the raw image bytes.

Available formats:

• png (default)
• apng
• jpg / jpeg
• tiff / tif
• jpeg2000 / jp2
• bmp
• nii
• nii.gz
• dcm

Parameters:

Name	Type	Description	Default
dtype	DataType | str	The data type of the array elements to return	'uint8'
sample_compression	str	The on-disk compression/format of the image	'png'

Examples:

ds.add_column("col1", types.Image)
ds.add_column("col2", types.Image(sample_compression="jpg"))

# Basic image storage
ds.add_column("images", deeplake.types.Image())

# JPEG compression
ds.add_column("images", deeplake.types.Image(
    sample_compression="jpeg"
))

# With specific dtype
ds.add_column("images", deeplake.types.Image(
    dtype="uint8"  # 8-bit RGB
))

deeplake.types.Embedding

Embedding(
    size: int | None = None,
    dtype: DataType | str = "float32",
    quantization: QuantizationType | None = None,
) -> Type

Creates a single-dimensional embedding of a given length.

Parameters:

Name	Type	Description	Default
size	int | None	int | None The size of the embedding	None
dtype	DataType | str	DataType | str The datatype of the embedding. Defaults to float32	'float32'
quantization	QuantizationType | None	QuantizationType | None How to compress the embeddings in the index. Default uses no compression, but can be set to :class:deeplake.types.QuantizationType.Binary	None

Returns:

Name	Type	Description
Type	Type	A new embedding data type.

See Also

:func:deeplake.types.Array for a multidimensional array.

Examples:

Create embedding columns:

ds.add_column("col1", types.Embedding(768))
ds.add_column("col2", types.Embedding(768, quantization=types.QuantizationType.Binary))

# Basic embeddings
ds.add_column("embeddings", deeplake.types.Embedding(768))

# With binary quantization for faster search
ds.add_column("embeddings", deeplake.types.Embedding(
    size=768,
    quantization=deeplake.types.QuantizationType.Binary
))

# Custom dtype
ds.add_column("embeddings", deeplake.types.Embedding(
    size=768,
    dtype="float32"
))

deeplake.types.Text

Text(index_type: str | TextIndexType | None = None) -> Type

Creates a text data type of arbitrary length.

Parameters:

Name	Type	Description	Default
index_type	str | TextIndexType | None	str | TextIndexType | None How to index the data in the column for faster searching. Options are: :class:deeplake.types.Inverted :class:deeplake.types.BM25 Default is None meaning "do not index"	None

Returns:

Name	Type	Description
Type	Type	A new text data type.

Examples:

Create text columns with different configurations:

ds.add_column("col1", types.Text)
ds.add_column("col2", "text")
ds.add_column("col3", str)
ds.add_column("col4", types.Text(index_type=types.Inverted))
ds.add_column("col5", types.Text(index_type=types.BM25))

# Basic text
ds.add_column("text", deeplake.types.Text())

# Text with BM25 index for semantic search
ds.add_column("text2", deeplake.types.Text(
    index_type=deeplake.types.BM25
))

# Text with inverted index for keyword search
ds.add_column("text3", deeplake.types.Text(
    index_type=deeplake.types.Inverted
))

deeplake.types.Dict

Dict() -> Type

Creates a type that supports storing arbitrary key/value pairs in each row.

Returns:

Name	Type	Description
Type	Type	A new dictionary data type.

See Also

:func:deeplake.types.Struct for a type that supports defining allowed keys.

Examples:

Create and use a dictionary column:

ds.add_column("col1", types.Dict)
ds.append([{"col1": {"a": 1, "b": 2}}])
ds.append([{"col1": {"b": 3, "c": 4}}])

# Store arbitrary key/value pairs
ds.add_column("metadata", deeplake.types.Dict())

# Add data
ds.append([{
    "metadata": {
        "timestamp": "2024-01-01",
        "source": "camera_1",
        "settings": {"exposure": 1.5}
    }
}])

deeplake.types.Array

Array(dtype: DataType | str, dimensions: int) -> DataType

Array(dtype: DataType | str, shape: list[int]) -> DataType

Array(dtype: DataType | str) -> DataType

Array(
    dtype: DataType | str,
    dimensions: int | None,
    shape: list[int] | None,
) -> DataType

Creates a generic array of data.

Parameters:

Name	Type	Description	Default
dtype	DataType | str	DataType | str The datatype of values in the array	required
dimensions	int | None	int | None The number of dimensions/axes in the array. Unlike specifying shape, there is no constraint on the size of each dimension.	required
shape	list[int] | None	list[int] | None Constrain the size of each dimension in the array	required

Returns:

Name	Type	Description
DataType	DataType	A new array data type with the specified parameters.

Examples:

Create a three-dimensional array, where each dimension can have any number of elements:

ds.add_column("col1", types.Array("int32", dimensions=3))

Create a three-dimensional array, where each dimension has a known size:

ds.add_column("col2", types.Array(types.Float32(), shape=[50, 30, 768]))

# Fixed-size array
ds.add_column("features", deeplake.types.Array(
    "float32",
    shape=[512]  # Enforces size
))

# Variable-size array
ds.add_column("sequences", deeplake.types.Array(
    "int32",
    dimensions=1  # Allows any size
))

deeplake.types.BinaryMask

BinaryMask(
    sample_compression: str | None = None,
    chunk_compression: str | None = None,
) -> Type

In binary mask, pixel value is a boolean for whether there is/is-not an object of a class present.

NOTE: Since binary masks often contain large amounts of data, it is recommended to compress them using lz4.

Parameters:

Name	Type	Description	Default
sample_compression	str | None	How to compress each row's value. Possible values: lz4, null (default: null)	None
chunk_compression	str | None	How to compress all the values stored in a single file. Possible values: lz4, null (default: null)	None

Examples:

ds.add_column("col1", types.BinaryMask(sample_compression="lz4"))
ds.append([{"col1": np.zeros((512, 512, 5), dtype="bool")}])

# Basic binary mask
ds.add_column("masks", deeplake.types.BinaryMask())

# With compression
ds.add_column("masks", deeplake.types.BinaryMask(
    sample_compression="lz4"
))

deeplake.types.SegmentMask

SegmentMask(
    dtype: DataType | str = "uint8",
    sample_compression: str | None = None,
    chunk_compression: str | None = None,
) -> Type

Segmentation masks are 2D representations of class labels where a numerical class value is encoded in an array of same shape as the image.

NOTE: Since segmentation masks often contain large amounts of data, it is recommended to compress them using lz4.

Parameters:

Name	Type	Description	Default
sample_compression	str | None	How to compress each row's value. Possible values: lz4, null (default: null)	None
chunk_compression	str | None	How to compress all the values stored in a single file. Possible values: lz4, null (default: null)	None

Examples:

ds.add_column("col1", types.SegmentMask(sample_compression="lz4"))
ds.append([{"col1": np.zeros((512, 512, 3))}])

# Basic segmentation mask
ds.add_column("segmentation", deeplake.types.SegmentMask())

# With compression
ds.add_column("segmentation", deeplake.types.SegmentMask(
    dtype="uint8",
    sample_compression="lz4"
))

deeplake.types.BoundingBox

BoundingBox(
    dtype: DataType | str = "float32",
    format: str | None = None,
    bbox_type: str | None = None,
) -> Type

Stores an array of values specifying the bounding boxes of an image.

Parameters:

Name	Type	Description	Default
dtype	DataType | str	The datatype of values (default float32)	'float32'
format	str | None	The bounding box format. Possible values: ccwh, ltwh, ltrb, unknown	None
bbox_type	str | None	The pixel type. Possible values: pixel, fractional	None

Examples:

ds.add_column("col1", types.BoundingBox())
ds.add_column("col2", types.BoundingBox(format="ltwh"))

# Basic bounding boxes
ds.add_column("boxes", deeplake.types.BoundingBox())

# With specific format
ds.add_column("boxes", deeplake.types.BoundingBox(
    format="ltwh"  # left, top, width, height
))

deeplake.types.Point

Point() -> Type

ds.add_column("points", deeplake.types.Point())

deeplake.types.Struct

Struct(fields: dict[str, DataType | str]) -> DataType

Defines a custom datatype with specified keys.

See deeplake.types.Dict for a type that supports different key/value pairs per value.

Parameters:

Name	Type	Description	Default
fields	dict[str, DataType | str]	A dict where the key is the name of the field, and the value is the datatype definition for it	required

Examples:

ds.add_column("col1", types.Struct({
   "field1": types.Int16(),
   "field2": "text",
}))

ds.append([{"col1": {"field1": 3, "field2": "a"}}])
print(ds[0]["col1"]["field1"]) # Output: 3

# Define fixed structure with specific types
ds.add_column("info", deeplake.types.Struct({
    "id": deeplake.types.Int64(),
    "name": "text",
    "score": deeplake.types.Float32()
}))

# Add data
ds.append([{
    "info": {
        "id": 1,
        "name": "sample",
        "score": 0.95
    }
}])

deeplake.types.Sequence

Sequence(nested_type: DataType | str | Type) -> Type

Creates a sequence type that represents an ordered list of other data types.

A sequence maintains the order of its values, making it suitable for time-series data like videos (sequences of images).

Parameters:

Name	Type	Description	Default
nested_type	DataType | str | Type	DataType | str | Type The data type of the values in the sequence. Can be any data type, not just primitive types.	required

Returns:

Name	Type	Description
Type	Type	A new sequence data type.

Examples:

Create a sequence of images:

ds.add_column("col1", types.Sequence(types.Image(sample_compression="jpg")))

# Sequence of images (e.g., video frames)
ds.add_column("frames", deeplake.types.Sequence(
    deeplake.types.Image(sample_compression="jpeg")
))

# Sequence of embeddings
ds.add_column("token_embeddings", deeplake.types.Sequence(
    deeplake.types.Embedding(768)
))

# Add data
ds.append([{
    "frames": [frame1, frame2, frame3],  # List of images
    "token_embeddings": [emb1, emb2, emb3]  # List of embeddings
}])

deeplake.types.Link

Link(type: Type) -> Type

A link to an external resource. The value returned will be a reference to the external resource rather than the raw data.

Parameters:

Name	Type	Description	Default
type	Type	The type of the linked data	required

Examples:

ds.add_column("col1", types.Link(types.Image()))