ResourceSource
Represents a source of a resource, which can be a string (e.g., URL or file path), a number (e.g., resource ID), or an object (e.g., binary data).
type ResourceSource = string | number | object;
ScalarType
Enum representing the scalar types of tensors.
enum ScalarType {
BYTE = 0, // Byte type (8-bit unsigned integer)
CHAR = 1, // Character type (8-bit signed integer)
SHORT = 2, // Short integer type (16-bit signed integer)
INT = 3, // Integer type (32-bit signed integer)
LONG = 4, // Long integer type (64-bit signed integer)
HALF = 5, // Half-precision floating point type (16-bit)
FLOAT = 6, // Single-precision floating point type (32-bit)
DOUBLE = 7, // Double-precision floating point type (64-bit)
BOOL = 11, // Boolean type
QINT8 = 12, // Quantized 8-bit signed integer type
QUINT8 = 13, // Quantized 8-bit unsigned integer type
QINT32 = 14, // Quantized 32-bit signed integer type
QUINT4X2 = 16, // Packed Quantized Unsigned 4-bit Integers type (2 number in 1 byte)
QUINT2X4 = 17, // Packed Quantized Unsigned 2-bit Integer type (4 numbers in 1 byte)
BITS16 = 22, // Raw Bits type
FLOAT8E5M2 = 23, // Quantized 8-bit floating point type: Sign bit, 5 Exponent bits, 2 Mantissa bits
FLOAT8E4M3FN = 24, // Quantized 8-bit floating point type: Sign bit, 4 Exponent bits, 3 Mantissa bits
FLOAT8E5M2FNUZ = 25,// Quantized 8-bit floating point type with No Unsigned Zero (NUZ): Sign bit, 5 Exponent bits, 2 Mantissa bits
FLOAT8E4M3FNUZ = 26,// Quantized 8-bit floating point type with No Unsigned Zero (NUZ): Sign bit, 4 Exponent bits, 3 Mantissa bits
UINT16 = 27, // Unsigned 16-bit integer type
UINT32 = 28, // Unsigned 32-bit integer type
UINT64 = 29, // Unsigned 64-bit integer type
}
TensorBuffer
Represents the data buffer of a tensor, which can be one of several typed array formats.
type TensorBuffer =
| ArrayBuffer
| Float32Array
| Float64Array
| Int8Array
| Int16Array
| Int32Array
| Uint8Array
| Uint16Array
| Uint32Array
| BigInt64Array
| BigUint64Array;
TensorPtr
Represents a pointer to a tensor, including its data buffer, size dimensions, and scalar type.
The data buffer of the tensor.
An array representing the size of each dimension of the tensor.
The scalar type of the tensor, as defined in the ScalarType enum.
LabelEnum
A readonly record mapping string keys to numeric or string values. Used to represent enum-like label maps for models.
type LabelEnum = Readonly<Record<string, number | string>>;
Triple
A readonly triple of values, typically used for per-channel normalization parameters.
type Triple<T> = readonly [T, T, T];
PixelData
Represents raw pixel data in RGB format for vision models.
This type extends TensorPtr with constraints specific to image data:
- dataPtr must be Uint8Array (8-bit unsigned integers)
- scalarType is always BYTE (ScalarType.BYTE)
- sizes represents [height, width, channels] where channels must be 3 (RGB)
RGB pixel data as Uint8Array. Expected format: RGB (3 channels), not RGBA or BGRA. Size must equal: width * height * 3.
sizes
[number, number, 3]
required
Dimensions of the pixel data: [height, width, channels].
- sizes[0]: height (number of rows)
- sizes[1]: width (number of columns)
- sizes[2]: channels (must be 3 for RGB)
Scalar type is always BYTE for pixel data.
Example
const pixelData: PixelData = {
dataPtr: new Uint8Array(width * height * 3), // RGB pixel data
sizes: [height, width, 3], // [height, width, channels]
scalarType: ScalarType.BYTE
};
Frame
Frame data for vision model processing.
getNativeBuffer
() => { pointer: bigint; release(): void }
required
Pointer to native platform buffer (zero-copy, best performance).
- On iOS: CVPixelBufferRef pointer
- On Android: AHardwareBuffer* pointer
Obtain from Vision Camera v5: frame.getNativeBuffer().pointer 