How to stack LSTM layers to classify speech files

Question

I have been trying to implement a LSTM based classifier to classify descrete speech. I have created feature vectors with 13 mfcc. For a given file have 2D vector of [99, 13]. After following the mnist_irnn example I could set up single layer RNN to classify my speech files. But now I want to add more layers to the network. Therefore, I have been trying to implement the network with two LSTM layers and softmax layer as the output layer. After going through number of posts here I could set up the network as follows, where it doesn't throw any exceptions during model building time.

from __future__ import print_function
import numpy as np

from keras.optimizers import SGD
from keras.utils.visualize_util import plot

np.random.seed(1337)  # for reproducibility
from keras.preprocessing import sequence
from keras.utils import np_utils
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation, TimeDistributedDense
from keras.layers.recurrent import LSTM
from SpeechResearch import loadData

batch_size = 5
hidden_units = 100
nb_classes = 10
print('Loading data...')
(X_train, y_train), (X_test, y_test) = loadData.load_mfcc(10, 2)

print(len(X_train), 'train sequences')
print(len(X_test), 'test sequences')
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)
print('y_train shape:', y_train.shape)
print('y_test shape:', y_test.shape)
print('Build model...')

Y_train = np_utils.to_categorical(y_train, nb_classes)
Y_test = np_utils.to_categorical(y_test, nb_classes)
print(batch_size, 99, X_train.shape[2])
print(X_train.shape[1:])
print(X_train.shape[2])
model = Sequential()

model.add(LSTM(output_dim=hidden_units, init='uniform', inner_init='uniform',
               forget_bias_init='one', activation='tanh', inner_activation='sigmoid', return_sequences=True,
               stateful=True, batch_input_shape=(batch_size, 99, X_train.shape[2])))
# model.add(Dropout(0.5))
model.add(LSTM(output_dim=hidden_units, init='uniform', inner_init='uniform',
               forget_bias_init='one', activation='tanh', inner_activation='sigmoid', return_sequences=True,
               stateful=True, input_length=X_train.shape[2]))

model.add(TimeDistributedDense(input_dim=hidden_units, output_dim=nb_classes))
model.add(Activation('softmax'))

# try using different optimizers and different optimizer configs
sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=sgd)

print("Train...")
model.fit(X_train, Y_train, batch_size=batch_size, nb_epoch=3, validation_data=(X_test, Y_test), show_accuracy=True)
score, acc = model.evaluate(X_test, Y_test,
                            batch_size=batch_size,
                            show_accuracy=True)
print('Test score:', score)
print('Test accuracy:', acc)

I have been trying different values at different points. (For the moment I have been trying with a small sample, therefore values are very small) But, now it is throwing exception during training. Some dimension mismatch.

Using Theano backend.
Loading data...
100 train sequences
20 test sequences
X_train shape: (100, 99, 13)
X_test shape: (20, 99, 13)
y_train shape: (100,)
y_test shape: (20,)
Build model...
5 99 13
(99, 13)
13
Train...
Train on 100 samples, validate on 20 samples
Epoch 1/3

Traceback (most recent call last):
  File "/home/udani/PycharmProjects/testResearch/SpeechResearch/lstmNetwork.py", line 54, in <module>
    model.fit(X_train, Y_train, batch_size=batch_size, nb_epoch=3, validation_data=(X_test, Y_test), show_accuracy=True)
  File "/usr/local/lib/python2.7/dist-packages/keras/models.py", line 581, in fit
    shuffle=shuffle, metrics=metrics)
  File "/usr/local/lib/python2.7/dist-packages/keras/models.py", line 239, in _fit
    outs = f(ins_batch)
  File "/usr/local/lib/python2.7/dist-packages/keras/backend/theano_backend.py", line 365, in __call__
    return self.function(*inputs)
  File "/home/udani/Documents/ResearchSW/Theano/theano/compile/function_module.py", line 786, in __call__
    allow_downcast=s.allow_downcast)
  File "/home/udani/Documents/ResearchSW/Theano/theano/tensor/type.py", line 177, in filter
    data.shape))
TypeError: ('Bad input argument to theano function with name "/usr/local/lib/python2.7/dist-packages/keras/backend/theano_backend.py:362"  at index 1(0-based)', 'Wrong number of dimensions: expected 3, got 2 with shape (5, 10).')

I would like to know what I am doing wrong here. I have been going through the code whole day, but still I am unable to figure out the reason for dimension mismatch.

Further, I would be really thankful if someone can explain what it is meant by output_dim. (Is that the shape of the vector output by a single node, when we have n number of nodes in a given layer? Should it be equal to the number of nodes in the next layer? )

Answer 1

You have a problem with Ydimension, the output should be something like (100, 99, 10), that is a set of sequences of outputs, same as features, just 1 in output. It seems your Y vector is different. Method to_categorical is not really applicable to a sequences, it expects a vector.

Alternatively you can output a single vector and feed it into a dense layer in in the last LSTM layer with return_sequences=False

You do not need stateful network as well.