Process fake_quant_dequant_op to support all quantized model, use input var name to denote the scale value !3125
Created by: juncaipeng
- 处理fake_quant_dequant_op，支持所有无权重的op进行量化
- 使用输入tensor的名字来表示scale，避免同一个op多个输入tensor的scale重名
- 兼容所有以前训练的量化模型
训练端量化代码： PR
- Quantized ops with weight have attrs: is_quantized_with_weight, activation_bits, weight_bits, activation_quantize_type, weight_quantize_type
- For quantized ops without weight have attrs: is_quantized_without_weight, activation_bits
- All quantized ops have the input threshold (KL threshold or abs_max value), the name of input threshold is the intput var name of fake_quant/fake_quant_dequant.