Skip to content

Latest commit

 

History

History
165 lines (125 loc) · 5.47 KB

quantization_tool_api.md

File metadata and controls

165 lines (125 loc) · 5.47 KB

Quantization Toolkit API [中文]

Calibrator Class

Initialization

Calibrator(quantization_bit, granularity='per-tensor', calib_method='minmax')

Arguments

  • quantization_bit (string):
    • 'int8' for full int8 quantization.
    • 'int16' for full int16 quantization.
  • granularity (string):
    • If granularity = 'per-tensor'(default), there will be one exponent per entire tensor.
    • If granularity = 'per-channel', there will be one exponent for each channel of a convolution layer.
  • calib_method (string):
    • If calib_method = 'minmax'(default), the threshold is derived from the minimum and maximum values of the layer outputs from calibration dataset.
    • If calib_method = 'entropy', the threshold is derived from Kullback-Leibler divergence (KL divergence).

check_model method

Calibrator.check_model(model_proto)

Checks the compatibility of your model.

Argument

  • model_proto (ModelProto): An FP32 ONNX model.

Return

  • -1: The model is incompatible.

set_method method

Calibrator.set_method(granularity, calib_method)

Configures quantization.

Arguments

  • granularity (string):
    • If granularity = 'per-tensor', there will be one exponent per entire tensor.
    • If granularity = 'per-channel', there will be one exponent for each channel of a convolution layer.
  • calib_method (string):
    • If calib_method = 'minmax', the threshold is derived from the minimum and maximum values of the layer outputs from calibration dataset.
    • If calib_method = 'entropy', the threshold is derived from Kullback-Leibler divergence (KL divergence).

set_providers method

Calibrator.set_providers(providers)

Configures the execution provider of ONNX Runtime.

Argument

  • providers (list of strings): An execution provider in the list, for example 'CPUExecutionProvider', and 'CUDAExecutionProvider'.

generate_quantization_table method

Calibrator.generate_quantization_table(model_proto, calib_dataset, pickle_file_path)

Generates the quantization table.

Arguments

  • model_proto (ModelProto): An FP32 ONNX model.
  • calib_dataset (ndarray): The calibration dataset used to compute the threshold. The larger the dataset, the longer time it takes to generate the quantization table.
  • pickle_file_path (string): Path of the pickle file that stores the dictionary of quantization parameters.

export_coefficient_to_cpp method

Calibrator.export_coefficient_to_cpp(model_proto, pickle_file_path, target_chip, output_path, file_name, print_model_info=False)

Exports the quantized model coefficient such as weight to deploy on ESP SoCs.

Arguments

  • model_proto (ModelProto): An FP32 ONNX model.
  • pickle_file_path (string): Path of the pickle file that stores the dictionary of quantization parameters.
  • target_chip (string): Currently support 'esp32', 'esp32s2', 'esp32c3' and 'esp32s3'.
  • output_path (string): Path of output files.
  • file_name (string): Name of output files.
  • print_model_info(bool):
    • False (default): No log will be printed.
    • True: Information of the model will be printed.

Evaluator Class

Initialization

Evaluator(quantization_bit, granularity, target_chip)

Arguments

  • quantization_bit (string):
    • 'int8' for full int8 quantization.
    • 'int16' for full int16 quantization.
  • granularity (string):
    • If granularity = 'per-tensor', there will be one exponent per entire tensor.
    • If granularity = 'per-channel', there will be one exponent for each channel of a convolution layer.
  • target_chip (string): 'esp32s3' by default.

check_model method

Evaluator.check_model(model_proto)

Checks the compatibility of your model.

Argument

  • model_proto (ModelProto): An FP32 ONNX model.

Return

  • -1: The model is incompatible.

set_target_chip method

Evaluator.set_target_chip(target_chip)

Configures the chip environment to simulate.

Argument

  • target_chip (string): For now only 'esp32s3' is supported.

set_providers method

Evaluator.set_providers(providers)

Configures the execution provider of ONNX Runtime.

Argument

  • providers (list of strings): An execution provider in the list, for example 'CPUExecutionProvider', and 'CUDAExecutionProvider'.

generate_quantized_model method

Evaluator.generate_quantized_model(model_proto, pickle_file_path)

Generates the quantized model.

Arguments

  • model_proto (ModelProto): An FP32 ONNX model.
  • pickle_file_path (string): Path of the pickle file that stores all quantization parameters for the FP32 ONXX model. This pickle file must contain a dictionary of quantization parameters for all input and output nodes in the model graph.

evaluate_quantized_model method

Evaluator.evaluate_quantized_model(batch_fp_input, to_float=False)

Obtains outputs of the quantized model.

Arguments

  • batch_fp_input (ndarray): Batch of floating-point inputs.
  • to_float (bool):
    • False (default): Outputs will be returned directly.
    • True: Outputs will be converted to floating-point values.

Returns

A tuple of outputs and output_names:

  • outputs (list of ndarray): Outputs of the quantized model.
  • output_names (list of strings): Names of outputs.