Char_Level_CNN_Model 模型介绍

模型介绍

模型参数总览

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
sent_input (InputLayer)      (None, 1014)              0         
_________________________________________________________________
embedding_14 (Embedding)     (None, 1014, 128)         8960      
_________________________________________________________________
conv1d_30 (Conv1D)           (None, 1008, 256)         229632    
_________________________________________________________________
activation_20 (Activation)   (None, 1008, 256)         0         
_________________________________________________________________
max_pooling1d_13 (MaxPooling (None, 336, 256)          0         
_________________________________________________________________
conv1d_31 (Conv1D)           (None, 330, 256)          459008    
_________________________________________________________________
activation_21 (Activation)   (None, 330, 256)          0         
_________________________________________________________________
max_pooling1d_14 (MaxPooling (None, 110, 256)          0         
_________________________________________________________________
conv1d_32 (Conv1D)           (None, 108, 256)          196864    
_________________________________________________________________
activation_22 (Activation)   (None, 108, 256)          0         
_________________________________________________________________
conv1d_33 (Conv1D)           (None, 106, 256)          196864    
_________________________________________________________________
activation_23 (Activation)   (None, 106, 256)          0         
_________________________________________________________________
conv1d_34 (Conv1D)           (None, 104, 256)          196864    
_________________________________________________________________
activation_24 (Activation)   (None, 104, 256)          0         
_________________________________________________________________
conv1d_35 (Conv1D)           (None, 102, 256)          196864    
_________________________________________________________________
activation_25 (Activation)   (None, 102, 256)          0         
_________________________________________________________________
max_pooling1d_15 (MaxPooling (None, 34, 256)           0         
_________________________________________________________________
flatten_5 (Flatten)          (None, 8704)              0         
_________________________________________________________________
dense_13 (Dense)             (None, 1024)              8913920   
_________________________________________________________________
dropout_9 (Dropout)          (None, 1024)              0         
_________________________________________________________________
dense_14 (Dense)             (None, 1024)              1049600   
_________________________________________________________________
dropout_10 (Dropout)         (None, 1024)              0         
_________________________________________________________________
dense_15 (Dense)             (None, 4)                 4100      
=================================================================
Total params: 11,452,676
Trainable params: 11,452,676
Non-trainable params: 0
_________________________________________________________________

可以看出该模型参数量较大，因此对数据集的要求也很庞大，CPU跑如此大的数据集应该很不方便，但如果数据集太小，那么就会很容易overfit这些参数。

逐层介绍

Input Layer

input_size = 1014，表示输入层每次输入的单词量为1014，输入层无参数， 多退少补，每行就是1014。

# Input 
inputs = Input(shape=(input_size,), name='sent_input', dtype='int64')  # shape=(?, 1014)

Embedding Layer

需要识别的单词类别是69个, 'abcdefghijklmnopqrstuvwxyz0123456789-,;.!?:\'"/\\|_@#$%^&*~+-=<>()[]{}'， 其他无法识别的得用一个<\UNK>去代替，所以识别单词数是70个， 而embedding_size = 128, 所以这一层的参数数是 $128\ast 70=8960$。

# Embedding layer
conv = Embedding(alphabet_size+1, embedding_size, input_length=input_size)(