update daily 20170904

main.py

@@ -8,8 +8,11 @@ import keras as kr
 import numpy as np
 
 from keras.models import Sequential
-from keras.layers import Dense, Dropout, Flatten # ,Input,LSTM,Convolution1D,MaxPooling1D,Merge
+from keras.layers import Dense, Dropout, Input # , Flatten,LSTM,Convolution1D,MaxPooling1D,Merge
-from keras.layers import Conv1D,LSTM,MaxPooling1D,Merge # Conv2D, MaxPooling2D,Conv1D
+from keras.layers import Conv1D,LSTM,MaxPooling1D, Lambda #, Merge, Conv2D, MaxPooling2D,Conv1D
+from keras import backend as K
+
+from readdata import DataSpeech
 
 class ModelSpeech(): # 语音模型类
 	def __init__(self,MS_EMBED_SIZE = 64,BATCH_SIZE = 32):
@@ -22,24 +25,52 @@ class ModelSpeech(): # 语音模型类
 
 	def CreateModel(self):
 		'''
-		定义LSTM/CNN模型，尚未完成
+		定义CNN/LSTM/CTC模型，使用函数式模型
+		输入层：39维的特征值序列，一条语音数据的最大长度设为1500（大约15s）
+		隐藏层一：1024个神经元的卷积层
+		隐藏层二：池化层，池化窗口大小为2
+		隐藏层三：Dropout层，需要断开的神经元的比例为0.3，防止过拟合
+		隐藏层四：循环层、LSTM层
+		隐藏层五：Dropout层，需要断开的神经元的比例为0.3，防止过拟合
+		输出层：全连接层，神经元数量为1279，使用softmax作为激活函数，使用CTC的loss作为损失函数
 		'''
-		_model = Sequential()
+		# 每一帧使用13维mfcc特征及其13维一阶差分和13维二阶差分表示，最大信号序列长度为1500
-		_model.add(LSTM(self.MS_EMBED_SIZE, return_sequences=True, input_shape = (200,400))) # input_shape需要修改
+		layer_input = Input((1500,39))
-		_model.add(Dropout(0.3))
-		_model.add(Conv1D(self.QA_EMBED_SIZE // 2, 5, border_mode="valid"))
-		_model.add(MaxPooling1D(pool_length=2, border_mode="valid"))
-		_model.add(Dropout(0.3))
-		_model.add(Flatten())
 		
+		layer_h1 = Conv1D(256, 5, use_bias=True, padding="valid")(layer_input) # 卷积层
+		layer_h2 = MaxPooling1D(pool_size=2, strides=None, padding="valid")(layer_h1) # 池化层
+		layer_h3 = Dropout(0.2)(layer_h2) # 随机中断部分神经网络连接，防止过拟合
+		layer_h4 = LSTM(256, activation='relu', use_bias=True)(layer_h3) # LSTM层
+		layer_h5 = Dropout(0.2)(layer_h4) # 随机中断部分神经网络连接，防止过拟合
+		layer_h6 = Dense(1279, activation="softmax")(layer_h5) # 全连接层
 		
+		#labels = Input(name='the_labels', shape=[60], dtype='float32')
+		layer_out = Lambda(ctc_lambda_func,output_shape=(1279,), name='ctc')(layer_h6) # CTC
+		_model = Model(inputs = layer_input, outputs = layer_out)
 		
 		#_model = Sequential()
-		#_model.add(Merge([m_lstm, aenc], mode="concat", concat_axis=-1))
+		
-		_model.add(Dense(1279, activation="softmax"))
+		#_model.add(Conv1D(256, 5,input_shape=(1500,39), use_bias=True, padding="valid"))
-		_model.compile(optimizer="adam", loss='categorical_crossentropy',metrics=["accuracy"])
+		#_model.add(MaxPooling1D(pool_size=2, strides=None, padding="valid"))
+		#_model.add(Dropout(0.3)) # 随机中断部分神经网络连接
+		
+		#_model.add(LSTM(256, activation='relu', use_bias=True))
+		#_model.add(Dropout(0.3)) # 随机中断部分神经网络连接
+		
+		#_model.add(Dense(1279, activation="softmax"))
+       ##_model.add(Lambda(ctc_lambda_func,output_shape=(1,),name='ctc'))
+       
+		#_model.compile(optimizer="sgd", loss='categorical_crossentropy',metrics=["accuracy"])
+		_model.compile(optimizer="sgd", loss='ctc',metrics=["accuracy"])
 		return _model
-
+		
+	def ctc_lambda_func(args):
+		#labels, y_pred, input_length, label_length = args
+		y_pred = args
+		#y_pred = y_pred[:, 2:, :]
+		return K.ctc_decode(y_pred,1279)
+		#return K.ctc_batch_cost(labels, y_pred, input_length, label_length)
+	
 	def TrainModel(self,datapath,epoch = 2,save_step=1000,filename='model_speech/LSTM_CNN_model'):
 		'''
 		训练模型
@@ -49,10 +80,26 @@ class ModelSpeech(): # 语音模型类
 			save_step: 每多少步保存一次模型
 			filename: 默认保存文件名，不含文件后缀名
 		'''
-		for epoch in range(epoch):
+		data=DataSpeech(datapath)
-			pass
+		data.LoadDataList('train')
-		pass
+		num_data=DataSpeech.GetDataNum() # 获取数据的数量
-
+		for epoch in range(epoch): # 迭代轮数
+			n_step = 0 # 迭代数据数
+			while True:
+				try:
+					data_input, data_label = data.GetData(n_step) # 读数据
+					
+					pass
+					# 需要写一个生成器函数
+					self._model.fit_generator(yielddatas, save_step)
+					n_step += 1
+				except StopIteration:
+					print('[error] generator error. please check data format.')
+					break
+				
+				self.SaveModel(comment='_e_'+str(epoch)+'_step_'+str(n_step))
+				
+				
 	def LoadModel(self,filename='model_speech/LSTM_CNN_model.model'):
 		'''
 		加载模型参数
@@ -71,13 +118,21 @@ class ModelSpeech(): # 语音模型类
 		'''
 		pass
 
+	def Predict(self,x):
+		'''
+		预测结果
+		'''
+		r = predict_on_batch(x)
+		return r
+		pass
+		
 	@property
 	def model(self):
 		'''
 		返回keras model
 		'''
 		return self._model
-	
+
 
 if(__name__=='__main__'):
 	pass

readdata.py

@@ -93,8 +93,18 @@ class DataSpeech():
 				v=self.NumToVector(n)
 				feat_out.append(v)
 		# 返回值分别是mfcc特征向量的矩阵及其一阶差分和二阶差分矩阵，以及对应的拼音符号矩阵
-		return feat_mfcc,feat_mfcc_d,feat_mfcc_dd,np.array(feat_out)
+		data_input = np.column_stack((feat_mfcc, feat_mfcc_d, feat_mfcc_dd))
+		data_label = np.array(feat_out)
+		return data_input, data_label
 	
+	def data_genetator(self, data_input, data_label):
+		'''
+		数据生成器函数，用于Keras的generator_fit训练
+		输入GetData函数产生的输出
+		'''
+		
+		pass
+		
 	def GetSymbolList(self):
 		'''
 		加载拼音符号列表，用于标记符号
@@ -109,7 +119,7 @@ class DataSpeech():
 				txt_l=i.split('\t')
 				list_symbol.append(txt_l[0])
 		txt_obj.close()
-		list_symbol.append(' ')
+		list_symbol.append('_')
 		return list_symbol
 
 	def SymbolToNum(self,symbol):