How to use Keras TimeseriesGenerator

user_11666502 0
Viewed 3117

I am having trouble implementing Keras TimeseriesGenerator. What I want is to experiment with different values for look_back, which is a variable that determines the lag length for X in terms of each y. Right now, I have it set to 3, but would like to be able to test multiple values. Essentially I want to see if using the last n rows to predict a value increases the accuracy. Here is my code:

### trying with timeseries generator
from keras.preprocessing.sequence import TimeseriesGenerator

look_back = 3

train_data_gen = TimeseriesGenerator(X_train, X_train,
    length=look_back, sampling_rate=1,stride=1,
    batch_size=3)
test_data_gen = TimeseriesGenerator(X_test, X_test,
    length=look_back, sampling_rate=1,stride=1,
    batch_size=1)

### Bi_LSTM
Bi_LSTM = Sequential()
Bi_LSTM.add(layers.Bidirectional(layers.LSTM(512, input_shape=(look_back, 11))))
Bi_LSTM.add(layers.Dropout(.5))
# Bi_LSTM.add(layers.Flatten())
Bi_LSTM.add(Dense(11, activation='softmax'))
Bi_LSTM.compile(optimizer='rmsprop',
              loss='categorical_crossentropy',
              metrics=['accuracy'])
### fitting a small normal model seems to be necessary for compile
Bi_LSTM.fit(X_train[:1],
              y_train[:1],
              epochs=1,
              batch_size=32,
              validation_data=(X_test[:1], y_test[:1]),
              class_weight=class_weights)
print('ignore above, necessary to run custom generator...')
Bi_LSTM_history = Bi_LSTM.fit_generator(Bi_LSTM.fit_generator(generator,
                                                    steps_per_epoch=1,
                                                    epochs=20,
                                                    verbose=0,
                                                    class_weight=class_weights))

Which yields the following error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-35-11561ec7fb92> in <module>()
     26               batch_size=32,
     27               validation_data=(X_test[:1], y_test[:1]),
---> 28               class_weight=class_weights)
     29 print('ignore above, necessary to run custom generator...')
     30 Bi_LSTM_history = Bi_LSTM.fit_generator(Bi_LSTM.fit_generator(generator,

2 frames
/usr/local/lib/python3.6/dist-packages/keras/engine/training_utils.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)
    143                             ': expected ' + names[i] + ' to have shape ' +
    144                             str(shape) + ' but got array with shape ' +
--> 145                             str(data_shape))
    146     return data
    147 

ValueError: Error when checking input: expected lstm_16_input to have shape (3, 11) but got array with shape (1, 11)

If I change the BiLSTM input shape to (1,11) like listed above, then I get this error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-36-7360e3790518> in <module>()
     31                                                     epochs=20,
     32                                                     verbose=0,
---> 33                                                     class_weight=class_weights))
     34 

5 frames
/usr/local/lib/python3.6/dist-packages/keras/engine/training_utils.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)
    143                             ': expected ' + names[i] + ' to have shape ' +
    144                             str(shape) + ' but got array with shape ' +
--> 145                             str(data_shape))
    146     return data
    147 

ValueError: Error when checking input: expected lstm_17_input to have shape (1, 11) but got array with shape (3, 11)

What is going on here?

If needed, my data is read from a df where each row (observation) is a (1,11) float vector and each label is an int, which I convert to a 1 hot vector shape (1,11).

lstm keras tensorflow machine-learning python
1 Answers

I found a lot of mistakes in the code... for this reason, I want to provide a dummy example that you can follow to carry out your task. Please pay attention to the original dimension of your data and the dimension of data generated by the TimeSeriesGenerator. This is important to understand how to build the network

# utility variable
look_back = 3
batch_size = 3
n_feat = 11
n_class = 11
n_train = 200
n_test = 60

# data simulation
X_train = np.random.uniform(0,1, (n_train,n_feat)) # 2D!
X_test = np.random.uniform(0,1, (n_test,n_feat)) # 2D!
y_train = np.random.randint(0,2, (n_train,n_class)) # 2D!
y_test = np.random.randint(0,2, (n_test,n_class)) # 2D!


train_data_gen = TimeseriesGenerator(X_train, y_train, length=look_back, batch_size=batch_size)
test_data_gen = TimeseriesGenerator(X_test, y_test, length=look_back, batch_size=batch_size)

# check generator dimensions
for i in range(len(train_data_gen)):
    x, y = train_data_gen[i]
    print(x.shape, y.shape)

Bi_LSTM = Sequential()
Bi_LSTM.add(Bidirectional(LSTM(512), input_shape=(look_back, n_feat)))
Bi_LSTM.add(Dropout(.5))
Bi_LSTM.add(Dense(n_class, activation='softmax'))
print(Bi_LSTM.summary())

Bi_LSTM.compile(optimizer='rmsprop',
              loss='categorical_crossentropy',
              metrics=['accuracy'])

Bi_LSTM_history = Bi_LSTM.fit_generator(train_data_gen,
                                        steps_per_epoch=50,
                                        epochs=3,
                                        verbose=1,
                                        validation_data=test_data_gen) # class_weight=class_weights)
user_103750490
Related