softmax

Classes

SoftmaxOperator

SoftmaxOperator(op: AirOperator, model: AirModel, platform: SocPlatform, prefix: str = 'aot', attributes: dict[str, str] = {})

SOFTMAX operator.

This operator computes the softmax activation function on the input tensor.

Parameters:

• op

  (AirOperator) –

  The AIR operator to wrap.

• model

  (AirModel) –

  The AIR model.

• platform

  (SocPlatform) –

  The target platform for code generation.

• prefix

  (str, default: 'aot' ) –

  Prefix for generated code files. Defaults to "aot".

• attributes

  (dict[str, str], default: {} ) –

  Attributes for template values. Defaults to {}.

Functions

compute_values

compute_values() -> dict[str, str]

Compute the values for the operator.

Returns:

• dict[str, str] –

  dict[str, str]: Computed values for the operator.

emit

emit(save_path: Path)

Generate the source code for the operator.

Parameters:

• save_path

  (Path) –

  Path to save the generated code.

Functions

preprocess_softmax_scaling

preprocess_softmax_scaling(beta: float, input_scale: float, scaled_diff_integer_bits: int = 5) -> AirFixedPointScale

Mimics litert::PreprocessSoftmaxScaling.

Parameters:

• beta

  (float) –

  The softmax beta parameter (often 1.0).

• input_scale

  (float) –

  The quantization scale for the input tensor.

• scaled_diff_integer_bits

  (int, default: 5 ) –

  Number of integer bits to reserve in the scaled difference (typically 5).

Returns:

• input_multiplier ( int ) –

  The fixed-point multiplier (in Q31 format).

• input_left_shift ( int ) –

  The left shift to apply; computed as (shift from quantization + scaled_diff_integer_bits).

make_softmax_s16_exp_luts

make_softmax_s16_exp_luts()

Build the exp LUT NS-CMSIS-NN expects, matching TFLM.

exp_lut: input_scale = 10.0 / 65535 input_zp = INT16_MAX output_scale = 2.0 / 65535 output_zp = 0 transform = exp(x)

Returns:

• –

  np.ndarray: The generated exp LUT (shape: (513,))

make_softmax_s16_luts

make_softmax_s16_luts()

Build the one_by_one_lut LUT NS-CMSIS-NN expects, matching TFLM.

one_by_one_lut: input_scale = 1.0 / 65535 input_zp = INT16_MIN output_scale = 2.0 / 65535 output_zp = 0 transform = 1 / (1 + x)

Returns:

• –

  np.ndarray: The generated one_by_one_lut (shape: (513,))