Keras: very low accuracy, very high loss and predictions are the same for every input

Question

i am developing a system for general-purpose audio tagging using keras.

I have the following data input: x_train has 10 different data for each input (data_leng,max,min,etc) and y_train represents 41 possible labels (guitar,bass,etc)

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.]

My model is:

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)

This is my result:

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

As you can see, my results show very high data loss and very low accuracy but the main problem is when i try to predict the result, cause for each one input the output is the same. How can i fix this ?

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

In your Dense layers, you need to specify the Input_dim only for the first layer.

Keras take care of the Dim in the other layers.

So try :

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'))

And maybe your regularization is too strong for this kind of data, try a dropout less strong or no dropout at all.

Last thing you can do is increasing your learning rate, start with something like 1e-3 and see if something change.

Hope i helped you

Answer 2

You can try testing others optimizers and try do change your last layer activation. I already haved the same problem, i was using Softmax activation in last Dense layer, i changed to Sigmoid and works well.

A good strategy are modifying the model's architecture, add more layers, change dropout values, etc...

Hope i helped you. Good luck!