same as 2a but with 5 time steps

seq2c.py

from keras.models import Sequential
from keras.layers import LSTM, TimeDistributed, Dense, Activation
import numpy as np


# predict whether two numbers in sequence are same or not.
# the only difference between this an seq2a is that it has 5 time steps.
# this works, except for first number in sequence (should output 0)...effectively there is only a single training example for the model to learn this
# since step 1 from all batches other than the first are initialized with state from last step of previous batch.

train_input = [np.random.randint(0,2) for r in xrange(1000)]
train_output = [0]

for i in range(1, len(train_input)):
    if(train_input[i]==train_input[i-1]):
        train_output.append(1)
    else:
        train_output.append(0)

train_input = np.reshape(train_input, (200, 5, 1))
train_output = np.reshape(train_output, (200,5,1))


print "test and training data loaded"

model = Sequential()
model.add(LSTM(10,  return_sequences=True, batch_input_shape=(1, 5, 1), stateful=True))
model.add(TimeDistributed(Dense(1)))
model.add(Activation('sigmoid'))

model.compile(loss='binary_crossentropy', optimizer='sgd', metrics=['accuracy'])

model.fit(np.asarray(train_input), np.asarray(train_output), epochs=200, batch_size=1, shuffle=False)


print model.predict(np.asarray([1,0,0,1,1]).reshape(1,5,1))
print model.predict(np.asarray([1,1,0,0,1]).reshape(1,5,1))