enums

Classes

AirTensorDim

AIR tensor dimensions.

Attributes:

• N (int) –

  Batch dimension.

• H (int) –

  Height dimension.

• W (int) –

  Width dimension.

• C (int) –

  Channel dimension.

AirSpatialPad

AIR spatial padding specification.

Attributes:

• top (int) –

  Padding on the top.

• bottom (int) –

  Padding on the bottom.

• left (int) –

  Padding on the left.

• right (int) –

  Padding on the right.

AirPaddingType

AIR padding types for AIR operators.

Attributes:

• SAME (int) –

  Same padding.

• VALID (int) –

  Valid padding.

AirTensorType

AIR tensor data types.

AirActivationType

AIR activation function types.

Attributes:

• NONE (str) –

  No activation.

• RELU (str) –

  ReLU activation.

• RELU_N1_TO_1 (str) –

  ReLU activation with range [-1, 1].

• RELU_0_TO_1 (str) –

  ReLU activation with range [0, 1].

• RELU6 (str) –

  ReLU activation with range [0, 6].

• TANH (str) –

  Tanh activation.

• SIGN_BIT (str) –

  Sign bit activation.

• PRELU (str) –

  Parametric ReLU activation.

Functions

compute_range

compute_range(output_zero_point: int, output_scale: float, *, dtype: DTypeLike | None = None, qmin: int | None = None, qmax: int | None = None, num_bits: int | None = None, signed: bool | None = None) -> tuple[int, int]

Compute quantized activation clamp range.

Precedence for the quantized range

1) dtype (e.g., np.int8 / np.uint8 / np.int16) 2) (num_bits, signed) (e.g., 4-bit signed) 3) qmin/qmax (explicit) 4) default to int8: [-128, 127]

AirTensorKind

AIR tensor kinds.

Attributes:

• CONSTANT (str) –

  Constant tensors (read-only).

• PERSISTENT (str) –

  Persistent including resource variables (read-write).

• SCRATCH (str) –

  Scratch tensors for ops (read-write).

AirReluType

AIR RELU types.

Attributes:

• RELU (str) –

  ReLU activation.

• RELU6 (str) –

  ReLU6 activation.

• RELU_N1_TO_1 (str) –

  ReLU activation with range [-1, 1].

• RELU_0_TO_1 (str) –

  ReLU activation with range [0, 1].

Functions

compute_range

compute_range(output_zero_point: int, output_scale: float, *, dtype: DTypeLike | None = None, qmin: int | None = None, qmax: int | None = None, num_bits: int | None = None, signed: bool | None = None) -> tuple[int, int]

Compute quantized activation clamp range.

Precedence for the quantized range

1) dtype (e.g., np.int8 / np.uint8 / np.int16) 2) (num_bits, signed) (e.g., 4-bit signed) 3) qmin/qmax (explicit) 4) default to int8: [-128, 127]

AirComparisonType

AIR comparison operation types.

AirFullyConnectedWeightsFormat

AIR fully-connected weights format.

Attributes:

• DEFAULT (int) –

  Default weights format.

• SHUFFLED4x16INT8 (int) –

  Shuffled 4x16 INT8 weights format.

AirOpType

AIR supported operator types.