class quantization.Quantization

__init__(model)

quantize(input_generator_func, quantizer=QuantizerType.ASYMMETRIC_AFFINE, qtype='uint8', iteration=1, minimize_layer_error=False)

Usually 500~1000 pictures are needed. Do not perform quantization on acuity networks converted from TensorFlow Lite models that have been quantized.

Parameters:

• input_generator_func – A inputs data generator function, requires function type.

• quantizer – The quantizer to quantize network tensors. Get QuantizerType class, support QuantizeType.ASYMMETRIC_AFFINE and QuantizeType.PERCHANNEL_SYMMETRIC_AFFINE now.

• qtype – A string, the qtype to quantizer. Support uint8 for ASYMMETRIC_AFFINE and int8 for PERCHANNEL_SYMMETRIC_AFFINE now.

• iteration – The number of sample image batches. requires int type.

• minimize_layer_error – Minimize the quantization error, requires bool type. If set it will improve quantization accuracy, but take more time.