diff --git a/SpeechModel2.py b/SpeechModel2.py index 8cdf6af..3ed809f 100644 --- a/SpeechModel2.py +++ b/SpeechModel2.py @@ -116,7 +116,7 @@ class ModelSpeech(): # 语音模型类 - def TrainModel(self, datapath, epoch = 2, save_step = 1000, batch_size = 32, filename = 'model_speech/LSTM_CNN_model'): + def TrainModel(self, datapath, epoch = 2, save_step = 1000, batch_size = 32, filename = 'model_speech/speech_model2'): ''' 训练模型 参数： @@ -146,13 +146,13 @@ class ModelSpeech(): # 语音模型类 self.SaveModel(comment='_e_'+str(epoch)+'_step_'+str(n_step * save_step)) - def LoadModel(self,filename='model_speech/LSTM_CNN_model.model'): + def LoadModel(self,filename='model_speech/speech_model2.model'): ''' 加载模型参数 ''' self._model.load_weights(filename) - def SaveModel(self,filename='model_speech/LSTM_CNN_model',comment=''): + def SaveModel(self,filename='model_speech/speech_model2',comment=''): ''' 保存模型参数 ''' diff --git a/SpeechModel3.py b/SpeechModel3.py index 0788f7b..85dcfed 100644 --- a/SpeechModel3.py +++ b/SpeechModel3.py @@ -132,7 +132,7 @@ class ModelSpeech(): # 语音模型类 #y_pred, labels, input_length, label_length = args y_true, y_pred = args #print(y_pred) - y_pred = y_pred[:, :, 0:-2] + y_pred = y_pred[:, 2:, :] #return K.ctc_decode(y_pred,self.MS_OUTPUT_SIZE) return K.ctc_batch_cost(y_true, y_pred, y_true.shape[1], y_pred.shape[1]) diff --git a/SpeechModel4.py b/SpeechModel4.py index fa9cd59..cfb8000 100644 --- a/SpeechModel4.py +++ b/SpeechModel4.py @@ -321,12 +321,12 @@ class ModelSpeech(): # 语音模型类 print(shape) - #print(test_input_data) + print('test_input_data:',test_input_data) y_p = self.test_func([test_input_data]) print(type(y_p)) print('y_p:',y_p) - for j in range(0,200): + for j in range(0,0): mean = sum(y_p[0][0][j])/len(y_p[0][0][j]) print('max y_p:',max(y_p[0][0][j]),'min y_p:',min(y_p[0][0][j]),'mean y_p:',mean,'mid y_p:',y_p[0][0][j][100]) print('argmin:',np.argmin(y_p[0][0][j]),'argmax:',np.argmax(y_p[0][0][j])) @@ -338,15 +338,30 @@ class ModelSpeech(): # 语音模型类 print(K.is_sparse(y_p)) - y_p = K.to_dense(y_p) + #y_p = K.to_dense(y_p) print(K.is_sparse(y_p)) + + _list = [] + for i in y_p: + list_i = [] + for j in i: + list_j = [] + for k in j: + list_j.append(np.argmin(k)) + list_i.append(list_j) + _list .append(list_i) + + #y_p = np.array(_list, dtype = np.float) + y_p = _list + #print(y_p,type(y_p),y_p.shape) #y_p = tf.sparse_to_dense(y_p,(2,397),1417,0) print(test_input_length.T) test_input_length = test_input_length.reshape(2,1) func_in_len = self.test_func_input_length([test_input_length]) print(type(func_in_len)) + print(func_in_len) #in_len = np.ones(shape[0]) * shape[1] - ctc_decoded = K.ctc_decode(y_p, input_length = func_in_len) + ctc_decoded = K.ctc_decode(y_p[0][0], input_length = tf.squeeze(func_in_len[0][0][0])) print(ctc_decoded) #ctc_decoded = ctc_decoded[0][0] diff --git a/SpeechModel5.py b/SpeechModel5.py index 63750d1..5227c20 100644 --- a/SpeechModel5.py +++ b/SpeechModel5.py @@ -63,31 +63,31 @@ class ModelSpeech(): # 语音模型类 # 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 input_data = Input(name='the_input', shape=(self.AUDIO_LENGTH, self.AUDIO_FEATURE_LENGTH)) - layer_h1_c = Conv1D(filters=256, kernel_size=5, strides=1, use_bias=True, padding="same")(input_data) # 卷积层 + layer_h1_c = Conv1D(filters=256, kernel_size=5, strides=1, use_bias=True, kernel_initializer='he_normal', padding="same")(input_data) # 卷积层 #layer_h1_a = Activation('relu', name='relu0')(layer_h1_c) layer_h1_a = LeakyReLU(alpha=0.3)(layer_h1_c) # 高级激活层 layer_h1 = MaxPooling1D(pool_size=2, strides=None, padding="valid")(layer_h1_a) # 池化层 layer_h2 = BatchNormalization()(layer_h1) - layer_h3_c = Conv1D(filters=256, kernel_size=5, strides=1, use_bias=True, padding="same")(layer_h2) # 卷积层 + layer_h3_c = Conv1D(filters=256, kernel_size=5, strides=1, use_bias=True, kernel_initializer='he_normal', padding="same")(layer_h2) # 卷积层 layer_h3_a = LeakyReLU(alpha=0.3)(layer_h3_c) # 高级激活层 #layer_h3_a = Activation('relu', name='relu1')(layer_h3_c) layer_h3 = MaxPooling1D(pool_size=2, strides=None, padding="valid")(layer_h3_a) # 池化层 layer_h4 = Dropout(0.1)(layer_h3) # 随机中断部分神经网络连接，防止过拟合 - layer_h5 = Dense(256, use_bias=True, activation="softmax")(layer_h4) # 全连接层 - layer_h6 = Dense(256, use_bias=True, activation="softmax")(layer_h5) # 全连接层 + layer_h5 = Dense(256, use_bias=True, kernel_initializer='he_normal', activation="relu")(layer_h4) # 全连接层 + layer_h6 = Dense(256, use_bias=True, kernel_initializer='he_normal', activation="relu")(layer_h5) # 全连接层 #layer_h4 = Activation('softmax', name='softmax0')(layer_h4_d1) - layer_h7a = LSTM(256, activation='softmax', use_bias=True, return_sequences=True)(layer_h6) # LSTM层 - layer_h7b = LSTM(256, activation='softmax', use_bias=True, return_sequences=True)(layer_h6) # LSTM层 + layer_h7a = LSTM(256, activation='tanh', use_bias=True, return_sequences=True, kernel_initializer='he_normal')(layer_h6) # LSTM层 + layer_h7b = LSTM(256, activation='tanh', use_bias=True, return_sequences=True, go_backwards=True, kernel_initializer='he_normal')(layer_h6) # LSTM层 layer_h7_merged = add([layer_h7a, layer_h7b]) - layer_h8a = LSTM(256, activation='softmax', use_bias=True, return_sequences=True)(layer_h7_merged) # LSTM层 - layer_h8b = LSTM(256, activation='softmax', use_bias=True, return_sequences=True)(layer_h7_merged) # LSTM层 + layer_h8a = LSTM(256, activation='tanh', use_bias=True, return_sequences=True, kernel_initializer='he_normal')(layer_h7_merged) # LSTM层 + layer_h8b = LSTM(256, activation='tanh', use_bias=True, return_sequences=True, go_backwards=True, kernel_initializer='he_normal')(layer_h7_merged) # LSTM层 layer_h8 = concatenate([layer_h8a, layer_h8b]) #layer_h10 = Activation('softmax', name='softmax1')(layer_h9) @@ -95,7 +95,7 @@ class ModelSpeech(): # 语音模型类 #layer_h10_dropout = Dropout(0.1)(layer_h10) # 随机中断部分神经网络连接，防止过拟合 #layer_h11 = Dense(512, use_bias=True, activation="softmax")(layer_h8) # 全连接层 - layer_h12 = Dense(self.MS_OUTPUT_SIZE, use_bias=True, activation="softmax")(layer_h8) # 全连接层 + layer_h12 = Dense(self.MS_OUTPUT_SIZE, use_bias=True, kernel_initializer='he_normal')(layer_h8) # 全连接层 #layer_h6 = Dense(1283, activation="softmax")(layer_h5) # 全连接层 y_pred = Activation('softmax', name='softmax2')(layer_h12) @@ -348,7 +348,7 @@ if(__name__=='__main__'): ms = ModelSpeech(datapath) #ms.LoadModel(modelpath + 'speech_model_e_0_step_1.model') - ms.TrainModel(datapath, epoch = 2, batch_size = 8, save_step = 1) + ms.TrainModel(datapath, epoch = 2, batch_size = 8, save_step = 10) #ms.TestModel(datapath, str_dataset='dev', data_count = 32) #r = ms.RecognizeSpeech_FromFile('E:\\语音数据集\\wav\\test\\D4\\D4_750.wav') #print('*[提示] 语音识别结果：\n',r)