KERAS:每个输入的精度非常低，非常高的损失和预测是相同的

Question

我正在开发一个用于使用keras的通用音频标记的系统。

我有以下数据输入：x_train对于每个输入(data_leng，max，min等(有10个不同的数据，y_train代表41个可能的标签(吉他，低音等(

x_train shape = (7104, 10)
y_train shape = (41,)
print(x_train[0])
[ 3.75732000e+05 -2.23437546e-05 -1.17187500e-02  1.30615234e-02
  2.65964586e-03  2.65973969e-03  9.80024859e-02  1.13624850e+00
  1.00003528e+00 -1.11458333e+00] 
print(y_train[0])
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0.]

我的模型是：

from keras.models import Sequential
from keras.optimizers import SGD
from keras.layers import Dense, Dropout, Activation
model = Sequential()
model.add(Dense(units=128, activation='relu', input_dim=10))
model.add(Dropout(0.5))
model.add(Dense(units=64, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(units=32, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(41, activation='softmax'))
opt = SGD(lr=0.0001, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
model.fit(np.array(x_train), np.array(y_train), epochs=5, batch_size=8)

这是我的结果：

Epoch 1/5
7104/7104 [==============================] - 1s 179us/step - loss: 15.7392 - acc: 0.0235
Epoch 2/5
7104/7104 [==============================] - 1s 132us/step - loss: 15.7369 - acc: 0.0236
Epoch 3/5
7104/7104 [==============================] - 1s 133us/step - loss: 15.7415 - acc: 0.0234
Epoch 4/5
7104/7104 [==============================] - 1s 132us/step - loss: 15.7262 - acc: 0.0242
Epoch 5/5
7104/7104 [==============================] - 1s 132us/step - loss: 15.6484 - acc: 0.0291

您可以看到，我的结果显示出很高的数据丢失和非常低的精度，但主要问题是当我尝试预测结果时，每个输入的原因是输出相同。我怎样才能解决这个问题 ？

pre = model.predict(np.array(x_train), batch_size=8, verbose=0)
for i in pre:
    print(i)
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
...

Answer 1

在密集层中，您需要仅针对第一层指定input_dim。

keras照顾其他层中的昏暗。

所以尝试：

model = Sequential()
model.add(Dense(units=128, activation='relu', input_dim=10))
model.add(Dropout(0.5))
model.add(Dense(units=64, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(units=32, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(41, activation='softmax'))

也许您的正则化太强了，对于这种数据，请尝试辍学少或根本没有辍学。

您可以做的最后一件事是提高学习率，从1e-3之类的东西开始，看看是否有所改变。

希望我帮助您

Answer 2

您可以尝试测试其他优化器，然后尝试更改最后一层激活。我已经遇到了同样的问题，我在最后一个致密层中使用了SoftMax激活，我更改为sigmoid，效果很好。

一个好的策略正在修改模型的体系结构，添加更多层，更改辍学值等...

希望我能帮助你。祝你好运！