''' The training model learns to generate a "question" which contains all the same words as the original question. So it isn't really learning a sequence, but the result is interesting. ''' from __future__ import print_function import os import sys import numpy as np from keras.engine import Input from keras.layers import RepeatVector, TimeDistributed, Dense, Activation, merge, ActivityRegularization, GRU, Embedding, \ regularizers from keras.models import Model from keras.optimizers import Adam try: import cPickle as pickle except: import pickle data_path = os.environ['INSURANCE_QA'] model_save = os.path.join(os.environ['MODEL_PATH'], 'model.h5') class InsuranceQA: def __init__(self): self.vocab = self.load('vocabulary') self.table = InsuranceQA.VocabularyTable(self.vocab.values()) def load(self, name): return pickle.load(open(os.path.join(data_path, name), 'rb')) class VocabularyTable: ''' Identical to CharacterTable from Keras example ''' def __init__(self, words): self.words = sorted(set(words)) self.words_indices = dict((c, i) for i, c in enumerate(self.words)) self.indices_words = dict((i, c) for i, c in enumerate(self.words)) def encode(self, sentence, maxlen, one_hot=False): if one_hot: indices = np.zeros((maxlen, len(self.words)), dtype=np.int32) for i, w in enumerate(sentence): if i == maxlen: break indices[i, self.words_indices[w]] = 1 return indices else: indices = np.zeros((maxlen,), dtype=np.int32) for i, w in enumerate(sentence): if i == maxlen: break indices[i] = self.words_indices[w] return indices def decode(self, indices, calc_argmax=True): if calc_argmax: indices = np.argmax(indices, axis=-1) # indices = [self.sample(i) for i in indices] return ' '.join(self.indices_words[x] for x in indices if x != 0) def sample(self, index, noise=0.2): index = np.log(index) / noise index = np.exp(index) / np.sum(np.exp(index)) index = np.argmax(np.random.multinomial(1, index, 1)) return index def get_model(question_maxlen, answer_maxlen, vocab_len, n_hidden, load_save=False): answer = Input(shape=(answer_maxlen,), dtype='int32') embedded = Embedding(input_dim=vocab_len, output_dim=n_hidden, mask_zero=True)(answer) # masked = Masking(mask_value=0.)(answer) # encoder rnn encode_rnn = GRU(n_hidden, return_sequences=True, dropout_U=0.2)(embedded) encode_rnn = GRU(n_hidden, return_sequences=False, dropout_U=0.2)(encode_rnn) encode_brnn = GRU(n_hidden, return_sequences=True, go_backwards=True, dropout_U=0.2)(embedded) encode_brnn = GRU(n_hidden, return_sequences=False, go_backwards=True, dropout_U=0.2)(encode_brnn) # repeat it maxlen times repeat_encoding_rnn = RepeatVector(question_maxlen)(encode_rnn) repeat_encoding_brnn = RepeatVector(question_maxlen)(encode_brnn) # decoder rnn decode_rnn = GRU(n_hidden, return_sequences=True, dropout_U=0.2, dropout_W=0.5)(repeat_encoding_rnn) decode_rnn = GRU(n_hidden, return_sequences=True, dropout_U=0.2)(decode_rnn) decode_brnn = GRU(n_hidden, return_sequences=True, go_backwards=True, dropout_U=0.2, dropout_W=0.5)( repeat_encoding_brnn) decode_brnn = GRU(n_hidden, return_sequences=True, go_backwards=True, dropout_U=0.2)(decode_brnn) merged_output = merge([decode_rnn, decode_brnn], mode='concat', concat_axis=-1) # output dense = TimeDistributed(Dense(vocab_len, activity_regularizer=regularizers.activity_l1(1e-4)))(merged_output) softmax = Activation('softmax')(dense) # compile the prediction model model = Model([answer], [softmax]) model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy']) if os.path.exists(model_save) and load_save: try: model.load_weights(model_save) except Exception: print('A model exists at "%s", but it is not compatible' % model_save) return model if __name__ == '__main__': question_maxlen, answer_maxlen = 20, 60 qa = InsuranceQA() batch_size = 50 n_test = 5 nb_epoch = 20 nb_iteration = 200 print('Generating data...') answers = qa.load('answers') def gen_questions(batch_size, test=False): if test: questions = qa.load('test1') else: questions = qa.load('train') while True: i = 0 question_idx = np.zeros(shape=(batch_size, question_maxlen, len(qa.vocab))) answer_idx = np.zeros(shape=(batch_size, answer_maxlen)) for s in questions: if test: ans = s['good'] else: ans = s['answers'] for a in ans: answer = qa.table.encode([qa.vocab[x] for x in answers[a]], answer_maxlen, one_hot=False) question = qa.table.encode([qa.vocab[x] for x in s['question']], question_maxlen, one_hot=True) # question = np.amax(question, axis=0, keepdims=False) answer_idx[i] = answer question_idx[i] = question i += 1 if i == batch_size: yield ([answer_idx], [question_idx]) i = 0 gen = gen_questions(batch_size) test_gen = gen_questions(n_test, test=True) print('Generating model...') model = get_model(question_maxlen=question_maxlen, answer_maxlen=answer_maxlen, vocab_len=len(qa.vocab), n_hidden=256, load_save=False) print('Training model...') for iteration in range(1, nb_iteration + 1): print('\n' + '-' * 50 + '\nIteration %d' % iteration) model.fit_generator(gen, samples_per_epoch=100 * batch_size, nb_epoch=nb_epoch) model.save_weights(model_save, overwrite=True) # test this iteration on some sample data x, y = next(test_gen) pred = model.predict(x, verbose=0) y = y[0] x = x[0] for i in range(n_test): print('Answer: {}'.format(qa.table.decode(x[i], calc_argmax=False))) print(' Expected: {}'.format(qa.table.decode(y[i]))) print(' Predicted: {}'.format(qa.table.decode(pred[i])))