我刚接触LSTM模型，但是我的网络很小。我从音频文件中提取了MFCC功能，并将其展平并作为输入。但是验证的准确性停留在2个值之间，而我的准确性却在不断下降。 我使用的RMSprop的学习率为0.001。 我尝试过更改优化器，添加退出和批处理规范化。 数据集也均衡。
Model: "model_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, 3460, 1) 0
_________________________________________________________________
cu_dnnlstm_1 (CuDNNLSTM) (None, 3460, 1024) 4206592
_________________________________________________________________
cu_dnnlstm_2 (CuDNNLSTM) (None, 1024) 8396800
_________________________________________________________________
dense_1 (Dense) (None, 512) 524800
_________________________________________________________________
batch_normalization_1 (Batch (None, 512) 2048
_________________________________________________________________
dropout_1 (Dropout) (None, 512) 0
_________________________________________________________________
dense_2 (Dense) (None, 256) 131328
_________________________________________________________________
batch_normalization_2 (Batch (None, 256) 1024
_________________________________________________________________
dropout_2 (Dropout) (None, 256) 0
_________________________________________________________________
dense_3 (Dense) (None, 1) 257
=================================================================
Total params: 13,262,849
Trainable params: 13,261,313
Non-trainable params: 1,536
_________________________________________________________________
Train on 385 samples, validate on 165 samples
Epoch 1/10
385/385 [==============================] - 61s 160ms/step - loss: 1.0811 - accuracy: 0.5143 - val_loss: 0.6917 - val_accuracy: 0.5273
Epoch 2/10
385/385 [==============================] - 55s 142ms/step - loss: 0.7536 - accuracy: 0.5169 - val_loss: 0.6980 - val_accuracy: 0.4727
Epoch 3/10
385/385 [==============================] - 55s 142ms/step - loss: 0.7484 - accuracy: 0.5039 - val_loss: 0.7002 - val_accuracy: 0.4727
Epoch 4/10
385/385 [==============================] - 55s 142ms/step - loss: 0.7333 - accuracy: 0.5091 - val_loss: 0.7030 - val_accuracy: 0.5273
Epoch 5/10
385/385 [==============================] - 55s 142ms/step - loss: 0.7486 - accuracy: 0.4675 - val_loss: 0.6917 - val_accuracy: 0.5273
Epoch 6/10
385/385 [==============================] - 55s 142ms/step - loss: 0.7222 - accuracy: 0.4935 - val_loss: 0.6917 - val_accuracy: 0.5273
Epoch 7/10
385/385 [==============================] - 55s 143ms/step - loss: 0.7208 - accuracy: 0.4883 - val_loss: 0.6919 - val_accuracy: 0.5273
Epoch 8/10
385/385 [==============================] - 55s 142ms/step - loss: 0.7134 - accuracy: 0.4805 - val_loss: 0.6919 - val_accuracy: 0.5273
Epoch 9/10
385/385 [==============================] - 55s 143ms/step - loss: 0.7168 - accuracy: 0.4987 - val_loss: 0.6927 - val_accuracy: 0.5273
Epoch 10/10
385/385 [==============================] - 55s 143ms/step - loss: 0.7089 - accuracy: 0.4909 - val_loss: 0.6926 - val_accuracy: 0.5273
这是我的代码：
def build_model():
input = Input((20*173,1))
x = Conv1D(filters=16, kernel_size=4, activation='relu')(input)
x = AveragePooling1D(pool_size=2)(x)
x = Conv1D(filters=16, kernel_size=3, activation='relu')(x)
x = AveragePooling1D(pool_size=2)(x)
x = Flatten()(x)
x = keras.layers.Reshape((13808, 1))(x)
x = CuDNNLSTM(1024, return_sequences=True)(x)
x = CuDNNLSTM(512)(x)
x = Dense(256,activation='relu')(x)
x = Dropout(0.3)(x)
x = Dense(128,activation='relu')(x)
x = Dropout(0.3)(x)
x = Dense(1,activation='sigmoid')(x)
model = Model(inputs=input, outputs=x)
return model
reduce_lr = ReduceLROnPlateau(monitor='val_accuracy', factor=0.2,patience=3, min_lr=0.001)
opt = RMSprop(lr=0.0001)
m2 = build_model()
m2.compile(loss = "binary_crossentropy", metrics=['accuracy'],optimizer = opt)
m2.fit(X, y, batch_size=16, epochs=10, validation_split=0.3,callbacks = [reduce_lr])