QuantizeWrapperV2 实际上是做什么的？

Question

所以我正在训练这个

small CNN model which has few Conv2D layers and some MaxPool2D, Activations, Dense

，基本上是

Tensorflow

提供的基本层。

我想要它

run on an embedded system

，它没有很多空间，并且不能进行浮点计算。

因此，我尝试使用该模型进行 QAT 训练（量化感知训练），以便权重（最终）量化为 8 位，并且我正在使用

tfmot.quantization.keras.QuantizeWrapperV2

。

我无法理解参数计数和它对每种类型的层执行的操作（数学上），并且我希望获得一些帮助来理解此 API 给出的未记录的数学运算。

下面是同一模型的摘要，一次带 QAT，一次不带。

Param #

列上有一个我无法理解的差异。

感谢您的帮助。

这是模型摘要没有应用了QAT：

 Layer (type)                Output Shape              Param #   
=================================================================
 input_2 (InputLayer)        [(None, 56, 56, 1)]       0         
                                                                 
 conv2d_3 (Conv2D)           (None, 54, 54, 30)        300       
                                                                 
 activation_3 (Activation)   (None, 54, 54, 30)        0         
                                                                 
 max_pooling2d_2 (MaxPooling  (None, 27, 27, 30)       0         
 2D)                                                             
                                                                 
 conv2d_4 (Conv2D)           (None, 25, 25, 16)        4336      
                                                                 
 activation_4 (Activation)   (None, 25, 25, 16)        0         
                                                                 
 max_pooling2d_3 (MaxPooling  (None, 12, 12, 16)       0         
 2D)                                                             
                                                                 
 conv2d_5 (Conv2D)           (None, 10, 10, 16)        2320      
                                                                 
 activation_5 (Activation)   (None, 10, 10, 16)        0         
                                                                 
 global_average_pooling2d_1   (None, 16)               0         
 (GlobalAveragePooling2D)                                        
                                                                 
 dense (Dense)               (None, 8)                 136       
                                                                 
 activation_6 (Activation)   (None, 8)                 0         
                                                                 
 dense_1 (Dense)             (None, 1)                 9         
                                                                 
 activation_7 (Activation)   (None, 1)                 0         
                                                                 
=================================================================
Total params: 7,101
Trainable params: 7,101
Non-trainable params: 0
_________________________________________________________________

这里是模型摘要WITH QAT 应用：

 Layer (type)                Output Shape              Param #   
=================================================================
 input_2 (InputLayer)        [(None, 56, 56, 1)]       0         
                                                                 
 quantize_layer_1 (QuantizeL  (None, 56, 56, 1)        3         
 ayer)                                                           
                                                                 
 quant_conv2d_3 (QuantizeWra  (None, 54, 54, 30)       301       
 pperV2)                                                         
                                                                 
 quant_activation_3 (Quantiz  (None, 54, 54, 30)       3         
 eWrapperV2)                                                     
                                                                 
 quant_max_pooling2d_2 (Quan  (None, 27, 27, 30)       1         
 tizeWrapperV2)                                                  
                                                                 
 quant_conv2d_4 (QuantizeWra  (None, 25, 25, 16)       4337      
 pperV2)                                                         
                                                                 
 quant_activation_4 (Quantiz  (None, 25, 25, 16)       3         
 eWrapperV2)                                                     
                                                                 
 quant_max_pooling2d_3 (Quan  (None, 12, 12, 16)       1         
 tizeWrapperV2)                                                  
                                                                 
 quant_conv2d_5 (QuantizeWra  (None, 10, 10, 16)       2321      
 pperV2)                                                         
                                                                 
 quant_activation_5 (Quantiz  (None, 10, 10, 16)       3         
 eWrapperV2)                                                     
                                                                 
 quant_global_average_poolin  (None, 16)               3         
 g2d_1 (QuantizeWrapperV2)                                       
                                                                 
 quant_dense (QuantizeWrappe  (None, 8)                137       
 rV2)                                                            
                                                                 
 quant_activation_6 (Quantiz  (None, 8)                3         
 eWrapperV2)                                                     
                                                                 
 quant_dense_1 (QuantizeWrap  (None, 1)                14        
 perV2)                                                          
                                                                 
 quant_activation_7 (Quantiz  (None, 1)                1         
 eWrapperV2)                                                     
                                                                 
=================================================================
Total params: 7,131
Trainable params: 7,101
Non-trainable params: 30
_________________________________________________________________

Answer 1

QuantizeWrapperV2

量化其所包装的 keras 层的权重和激活。这涉及将高精度权重和激活转换为较低精度的格式。当您应用

QAT

时，层会被修改以支持训练期间的量化，这会导致管理量化过程的参数数量略有增加。要了解更多信息，请参阅此文档。