Simple Keras recurrent neural network skeleton for sequence-to-sequence mapping

seq2seqRNN.py

__author__ = 'Conan'

import numpy as np
import random
from keras.models import Sequential
from keras.layers.recurrent import SimpleRNN



# Toy dictionary of 1000 indices to random length 10 vectors
toywords = {}
toywords[0] = np.zeros(10)
for i in range(1,1000):
    toywords[i] = np.random.rand(1,10)

# toy sequences from above dictionary/list:
toyinput = []
for i in range(100):
    wordnum = np.random.randint(3,20)
    toyinput.append(random.sample(xrange(1000), wordnum))

# 100 sets of output vectors to correspond to those sequences
toyoutput = []
for ti in toyinput:
    toyoutput.append([])
    for t in ti:
        toyoutput[-1].insert(-1,np.random.rand(1,5))


X = np.zeros((100, 20, 10))
for r in range(len(toyinput)):
    for c in range(20):
        try:
            X[r,c,:] = toywords[toyinput[r][c]]
        except:
            X[r,c,:] = toywords[0]


Y = np.zeros((100, 20, 5))
for r in range(len(toyoutput)):
    for c in range(20):
        try:
            Y[r,c,:] = toyoutput[r][c]
        except:
            Y[r,c,:] = np.zeros(5)



# length 10 toy vectors
embedding_size = 10
# Not sure what difference it makes
batch_size = 16
# Hidden dimension 100
hidden_size = 100
# Maximum length  ??
maxlen = 200
# Maximum features  ??
output_size = 5


# Sequential
model = Sequential()
model.add(SimpleRNN(10, 5,
        init='glorot_uniform', inner_init='orthogonal', activation='sigmoid', weights=None,
        truncate_gradient=-1, return_sequences=True))


model.compile(loss='binary_crossentropy', optimizer='adam')

for e in range(100):
    results = model.fit(X, Y, batch_size=batch_size, nb_epoch=1, validation_split=0.1, show_accuracy=True, verbose=1)
    preds = model.predict_proba(X, batch_size=batch_size)