class QAT(object):
"""
Quant Aware Training(QAT): Add the fake quant logic for given quantizable layers, namely add the quant_dequant computational logic both for activation inputs and weight inputs.
"""
def __init__(self,
config=None,
weight_preprocess=None,
act_preprocess=None,
weight_quantize=None,
act_quantize=None):
"""
Args:
model(nn.Layer)
config(dict, optional): configs for quantization. if None, will use default config.
Default: None.
weight_quantize(class, optional): Defines how to quantize weight. Using this
can quickly test if user's quantization method works or not. In this method, user should
both define quantization function and dequantization function, that is, the function's input
is non-quantized weight and function returns dequantized weight. If None, will use
quantization op defined by 'weight_quantize_type'.
Default is None.
act_quantize(class, optional): Defines how to quantize activation. Using this
can quickly test if user's quantization method works or not. In this function, user should
both define quantization and dequantization process, that is, the function's input
is non-quantized activation and function returns dequantized activation. If None, will use
quantization op defined by 'activation_quantize_type'.
Default is None.
weight_preprocess(class, optional): Defines how to preprocess weight before quantization. Using this
can quickly test if user's preprocess method works or not. The function's input
is non-quantized weight and function returns processed weight to be quantized. If None, will
use preprocess method defined by 'weight_preprocess_type'.
Default is None.
act_preprocess(class, optional): Defines how to preprocess activation before quantization. Using this
can quickly test if user's preprocess method works or not. The function's input
is non-quantized activation and function returns processed activation to be quantized. If None,
will use preprocess method defined by 'activation_preprocess_type'.
Default is None.
"""
yghstill
commented
2 years ago
如何自定义量化方法，例如dorefa 8bit量化，有没有这方面的案例参考?
我看\dygraph\quant\qat.py里面的QAT对象留了自定义量化方法的接口，有没有实际例子？