Skip to main content

afe.apis.defines

Source: afe/apis/defines.py

Back to index

This file contains definitions of the types exposed by the development API for AFE.

Imports​

  • afe.apis.error_handling_variables
  • afe.backends.mla.afe_to_n2a_compiler.defines.TensorDRAMLayout
  • afe.backends.mla.afe_to_n2a_compiler.defines.TensorTessellateParameters
  • afe.backends.mla.afe_to_n2a_compiler.defines.TessellateParameters
  • afe.ir.defines.BiasCorrectionType
  • afe.ir.defines.InputName
  • afe.ir.defines.NodeName
  • afe.ir.defines.RequantizationMode
  • afe.ir.tensor_type.ScalarType
  • afe.ir.tensor_type.TensorType
  • afe.ir.tensor_type.scalar_type_from_dtype
  • afe.ir.tensor_type.scalar_type_to_dtype
  • afe.ir.utils.transpose_tensor_according_to_layout_strings
  • dataclasses
  • dataclasses.dataclass
  • enum.Enum
  • enum.auto
  • numpy as np
  • sima_utils.common.CustomPlatformParams
  • sima_utils.common.Platform
  • sima_utils.common.load_custom_config
  • sima_utils.logging.sima_logger
  • typing.Any
  • typing.ContextManager
  • typing.Dict
  • typing.List
  • typing.Optional
  • typing.Set

Constants​

  • InputValues (line 35) [default/value Dict[InputName, np.ndarray]]
  • gen1_target (line 38) [default/value Platform.GEN1]
  • gen2_target (line 39) [default/value Platform.GEN2]
  • gen_custom_target (line 40) [default/value Platform.GEN_CUSTOM]
  • BT_COLOR_COEFF (line 131) [type Dict[ColorSpaceStandard, List[float]] ; default/value {ColorSpaceStandard.BT601: [0.299, 0.587, 0.114, 1.772, 1.402], ColorSpaceStandard.BT709: [0.2126, 0.7152, 0.0722, 1.8556, 1.5748], ColorSpaceStandard.BT2020: [0.2627, 0.678, 0.0593, 1.8814, 1.4747]}]
  • YUV2RGB_FULL_RANGE_CONSTANTS (line 138) [type Dict[str, List[float]] ; default/value {'offset': [16, 128, 128], 'scale': [255 / 219, 255 / 224, 255 / 224]}]
  • default_quantization (line 371) [type QuantizationParams ; default/value QuantizationParams(calibration_method=(default_calibration()), activation_quantization_scheme=(quantization_scheme(True, False)), weight_quantization_scheme=(quantization_scheme(False, True)), requantization_mode=(RequantizationMode.sima), node_names={''}, custom_quantization_configs=None)]

Classes​

  • ExceptionFuncType(Enum) (line 43) Enum Members:
    • LOADED_NET_LOAD (line 44) [default/value auto()]
    • LOADED_NET_EXECUTE (line 45) [default/value auto()]
    • LOADED_NET_QUANTIZE (line 46) [default/value auto()]
    • LOADED_NET_CONVERT (line 47) [default/value auto()]
    • MODEL_EXECUTE (line 48) [default/value auto()]
    • MODEL_SAVE (line 49) [default/value auto()]
    • MODEL_LOAD (line 50) [default/value auto()]
    • MODEL_COMPILE (line 51) [default/value auto()]
    • MODEL_CREATE_AUXILIARY (line 52) [default/value auto()]
    • MODEL_COMPOSE (line 53) [default/value auto()]
    • MODEL_EVALUATE (line 54) [default/value auto()]
    • MODEL_PERFORMANCE (line 55) [default/value auto()]
    • GENERATE_ELF_FILES (line 56) [default/value auto()]
    • QUANTIZATION_ERROR_ANALYSIS (line 57) [default/value auto()]
  • ColorSpaceStandard(str, Enum) (line 121): Color space standards for YUV and RGB conversion. BT601 is for SD video; BT709 is for HD video; BT2020 is for HDR. Enum Members:
    • BT601 (line 126) [default/value 'BT601']
    • BT709 (line 127) [default/value 'BT709']
    • BT2020 (line 128) [default/value 'BT2020']
  • ColorConversion(str, Enum) (line 144): Color conversion direction. Enum Members:
    • YUV2RGB (line 148) [default/value 'YUV2RGB']
    • RGB2YUV (line 149) [default/value 'RGB2YUV']
    • BGR2RGB (line 150) [default/value 'BGR2RGB']
    • RGB2BGR (line 151) [default/value 'RGB2BGR']
  • ChromaSampling(str, Enum) (line 154): Chroma sub-sampling representation. Enum Members:
    • NV12 (line 158) [default/value 'NV12']
    • YUV420 (line 159) [default/value 'YUV420']
    • YUV422 (line 160) [default/value 'YUV422']
  • ResizeMethod(str, Enum) (line 163): Interpolation method used in resize transform. Enum Members:
    • LINEAR (line 167) [default/value 'linear']
    • NEAREST (line 168) [default/value 'nearest']
    • AREA (line 169) [default/value 'area']
    • CUBIC (line 170) [default/value 'cubic']
  • ResizeDepositLocation(str, Enum) (line 173): Deposit location of resized image in padded frame. Enum Members:
    • TOPLEFT (line 177) [default/value 'topleft']
    • CENTER (line 178) [default/value 'center']
    • BOTTOMRIGHT (line 179) [default/value 'bottomright']
  • CalibrationMethod (line 182) Decorators: dataclass. Attributes:

    • name (line 187)

    • from_str(method: str) (line 190) Decorators: staticmethod.

      Parameters:

      • method: type str
  • MinMaxMethod(CalibrationMethod) (line 210) Decorators: dataclass.
  • HistogramMSEMethod(CalibrationMethod) (line 216) Decorators: dataclass. Attributes:
    • num_bins (line 221) [type int ; default/value num_bins]
  • MovingAverageMinMaxMethod(CalibrationMethod) (line 225) Decorators: dataclass.
  • HistogramEntropyMethod(CalibrationMethod) (line 231) Decorators: dataclass. Attributes:
    • num_bins (line 236) [type int ; default/value num_bins]
  • HistogramPercentileMethod(CalibrationMethod) (line 240) Decorators: dataclass. Attributes:
    • percentile_value (line 246) [type float ; default/value percentile_value]
    • num_bins (line 247) [type int ; default/value num_bins]
  • SkipCalibration(CalibrationMethod) (line 251): Directive to skip calibration. Decorators: dataclass.
  • QuantizationScheme (line 264): Quantization scheme. Decorators: dataclass. Attributes:
    • asymmetric (line 274) [type bool]
    • per_channel (line 275) [type bool]
    • bits (line 276) [type int ; default/value 8]
    • bf16 (line 277) [type bool ; default/value False]
  • QuantizationParams (line 295): Parameters controlling how to quantize a network. Decorators: dataclass. Attributes:

    • calibration_method (line 312) [type CalibrationMethod]

    • activation_quantization_scheme (line 313) [type QuantizationScheme]

    • weight_quantization_scheme (line 314) [type QuantizationScheme]

    • requantization_mode (line 315) [type RequantizationMode ; default/value RequantizationMode.sima]

    • node_names (line 316) [type Set[str] ; default/value dataclasses.field(default_factory=set)]

    • custom_quantization_configs (line 317) [type Optional[Dict[NodeName, Dict[str, Any]]] ; default/value None]

    • biascorr_type (line 318) [type BiasCorrectionType ; default/value BiasCorrectionType.NONE]

    • channel_equalization (line 319) [type bool ; default/value False]

    • smooth_quant (line 320) [type bool ; default/value False]

    • prefer_int8_udf (line 321) [type bool ; default/value True]

    • with_calibration(method: CalibrationMethod) -> QuantizationParams (line 323)

      Parameters:

      • method: type CalibrationMethod

      Returns: QuantizationParams

    • with_activation_quantization(scheme: QuantizationScheme) -> QuantizationParams (line 327)

      Parameters:

      • scheme: type QuantizationScheme

      Returns: QuantizationParams

    • with_weight_quantization(scheme: QuantizationScheme) -> QuantizationParams (line 331)

      Parameters:

      • scheme: type QuantizationScheme

      Returns: QuantizationParams

    • with_requantization_mode(requantization_mode: RequantizationMode) (line 335)

      Parameters:

      • requantization_mode: type RequantizationMode

    • with_unquantized_nodes(node_names: Set[str]) -> QuantizationParams (line 339)

      Parameters:

      • node_names: type Set[str]

      Returns: QuantizationParams

    • with_custom_quantization_configs(custom_quantization_configs: Dict[NodeName, Dict[str, Any]]) (line 343)

      Parameters:

      • custom_quantization_configs: type Dict[NodeName, Dict[str, Any]]

    • with_bias_correction(enable: bool | BiasCorrectionType = True) (line 347)

      Parameters:

      • enable: type bool | BiasCorrectionType, default True

    • with_channel_equalization(enable: bool = True) (line 358)

      Parameters:

      • enable: type bool, default True

    • with_smooth_quant(enable: bool = True) (line 362)

      Parameters:

      • enable: type bool, default True

    • with_prefer_int8_udf(mode: bool = True) (line 366)

      Parameters:

      • mode: type bool, default True

Functions​

  • default_calibration() -> CalibrationMethod (line 260)

    Returns: CalibrationMethod

  • quantization_scheme(asymmetric: bool, per_channel: bool, bits: int = 8) -> QuantizationScheme (line 280): Constructs quantization scheme.

    Parameters:

    • asymmetric: type bool
    • per_channel: type bool
    • bits: type int, default 8

    Returns: QuantizationScheme

  • bfloat16_scheme() -> QuantizationScheme (line 287): Constructs a bfloat16 quantization scheme. It directs the compiler to use bfloat16 instead of integer quantization.

    Returns: QuantizationScheme