import numpy as np import tensorflow as tf import os import datetime from DataReader import Reader class BaseC(): def __init__(self, FLAGS): self.FLAGS = FLAGS def formData(self, sentence_tf, aspect_tf): with tf.name_scope('Data_formation'): sentence_tf_exp = tf.expand_dims(sentence_tf, 2) aspect_tf_exp = tf.expand_dims(aspect_tf, 1) mask = tf.where(tf.equal(sentence_tf_exp, aspect_tf_exp)) idx_start_aspect = tf.cast(tf.segment_min(mask[:,1], mask[:,0]), dtype=tf.int64) aspect_length = tf.reduce_sum(tf.cast(tf.not_equal(aspect_tf, tf.constant(0, dtype=tf.int64)), dtype=tf.int64), 1) idx_end_aspect = idx_start_aspect + aspect_length idx_end = tf.reduce_sum(tf.cast(tf.not_equal(sentence_tf, tf.constant(0, dtype=tf.int64)), dtype=tf.int64), 1) indices = tf.cast(tf.reshape(tf.tile(tf.range(0, self.FLAGS.N), [tf.shape(sentence_tf)[0]]), [tf.shape(sentence_tf)[0], self.FLAGS.N]),tf.int64) idx_lc = tf.tile(tf.expand_dims(idx_start_aspect, 1), [1, self.FLAGS.N]) idx_a = tf.tile(tf.expand_dims(idx_end_aspect, 1), [1, self.FLAGS.N]) idx_rc = tf.tile(tf.expand_dims(idx_end, 1), [1, self.FLAGS.N]) pad_length = self.FLAGS.N - tf.shape(sentence_tf)[1] paddings = [tf.zeros([2], dtype=tf.int64), [tf.zeros([1], dtype=tf.int64)[0], pad_length]] whole = tf.cast(tf.pad(sentence_tf, paddings, 'CONSTANT', constant_values=0), dtype=tf.int64) LC = tf.where(tf.less(indices, idx_lc), whole, tf.zeros_like(whole, dtype=tf.int64)) A = tf.where(tf.logical_and(tf.greater_equal(indices, idx_lc), tf.less(indices, idx_a)), whole, tf.zeros_like(whole, dtype=tf.int64)) RC = tf.where(tf.logical_and(tf.greater_equal(indices, idx_a), tf.less(indices, idx_rc)), whole, tf.zeros_like(whole, dtype=tf.int64)) return LC, A, RC def getMinibatches(self, batch_size, sentence_data, aspect_data, Y): with tf.name_scope('Mini_batch'): batch_size = tf.cond(batch_size > tf.cast(tf.shape(Y)[0], tf.int64), lambda: tf.cast(tf.shape(Y)[0], tf.int64), lambda: batch_size) df = tf.data.Dataset.from_tensor_slices((sentence_data, aspect_data, Y)).repeat().batch(batch_size) iterator = df.make_initializable_iterator() next_batch = iterator.get_next() nr_batches = tf.identity(tf.cast(tf.math.ceil(tf.shape(sentence_data)[0]/tf.cast(batch_size, tf.int32)), tf.int64), name='nr_batches') LC, A, RC = self.formData(sentence_data, aspect_data) return iterator, next_batch, nr_batches, LC, A, RC def getAccuracy(self, pred, y, data_len, data_len_glove): acc = tf.reduce_mean(tf.cast(tf.equal(tf.math.argmax(y,1), tf.math.argmax(pred,1)), tf.float64)) * (data_len_glove / data_len) return tf.metrics.mean(acc) def getLoss(self, logits, y): return tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=y, name='loss') def getLossOp(self, loss): return tf.metrics.mean(loss, name='loss_op') def getSummaries(self, logits, prediction, Y, data_len, data_len_glove): with tf.name_scope('Summaries'): loss = self.getLoss(logits, Y) loss_log, loss_log_update_op = self.getLossOp(loss) loss_scalar = tf.summary.scalar('loss', loss_log) accuracy, accuracy_update_op = self.getAccuracy(prediction, Y, data_len, data_len_glove) accuracy_scalar = tf.summary.scalar('accuracy', accuracy) return loss, loss_log, loss_log_update_op, loss_scalar, accuracy, accuracy_update_op, accuracy_scalar def initWeights(self): with tf.name_scope('weights'): weights = {\ ### Sentence Level Content Attention Module 'W_1' : tf.get_variable(initializer=tf.random_normal([1, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_1'), 'W_2' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_2'), 'W_3' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_3'), 'W_4' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_4'), 'b_1' : tf.get_variable(initializer=tf.zeros([1, self.d]), dtype=tf.float32, name='b_1'), ### MLP 'W_5' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_5'), 'b_2' : tf.get_variable(initializer=tf.zeros([1, self.d]), dtype=tf.float32, name='b_2'), ### Linear Layer 'W_6' : tf.get_variable(initializer=tf.random_normal([self.FLAGS.nr_cat, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_6'), 'b_3' : tf.get_variable(initializer=tf.zeros([1, self.FLAGS.nr_cat]), dtype=tf.float32, name='b_3')} return weights def getLambda(self, E, E_LC, E_A, E_RC): mask_left = tf.not_equal(E_LC, tf.constant(0.0), name='mask_left') mask_left_with_aspect = tf.not_equal(E_LC+E_A, tf.constant(0.0), name='mask_left_with_aspect') mask_right = tf.not_equal(E_RC, tf.constant(0.0), name='mask_right') mask_right_with_aspect = tf.not_equal(E_RC+E_A, tf.constant(0.0), name='mask_right_with_aspect') mask_sentence = tf.not_equal(E, tf.constant(0.0), name='mask_sentence') left_length = tf.reduce_sum(tf.cast(mask_left_with_aspect[:,:,0], tf.int32), 1, name='left_length') # (S) right_length = tf.reduce_sum(tf.cast(mask_right_with_aspect[:,:,0], tf.int32), 1, name='right_length') # (S) sentence_length = tf.reduce_sum(tf.cast(mask_sentence[:,:,0], tf.int32), 1, name='sentence_length') # (S) left_tiled = tf.identity(tf.tile(tf.reshape(tf.range(self.FLAGS.N+1, 1, -1), [1, -1]), [tf.shape(E)[0], 1]) - tf.reshape(right_length, [-1,1]), name='left_tiled') # (S x N) left_reversed = tf.identity(left_tiled - (self.FLAGS.N+1 - tf.reshape(sentence_length, [-1,1])), name='left_reversed') # (S x N) left_masked = tf.where(mask_left[:,:,0], tf.cast(left_reversed, tf.float32), tf.zeros_like(mask_left[:,:,0], dtype=tf.float32), name='left_masked') # (S x N) right_tiled = tf.identity(tf.tile(tf.reshape(tf.range(1, self.FLAGS.N+1), [1,-1]), [tf.shape(E)[0],1]) - tf.reshape(left_length, [-1,1]), name='right_tiled') # (S x N) right_masked = tf.where(mask_right[:,:,0], tf.cast(right_tiled, tf.float32), tf.zeros_like(mask_right[:,:,0], dtype=tf.float32), name='right_masked') # (S x N) sentence_locations = tf.identity((left_masked + right_masked) / tf.cast(tf.reshape(sentence_length, [-1,1]), tf.float32), name='sentence_location') # (S x N) sentence_masked = tf.where(mask_sentence[:,:,0], 1-sentence_locations, tf.zeros_like(mask_sentence[:,:,0], dtype=tf.float32), name='sentence_masked') # (S x N) Lambda = tf.identity(tf.tile(tf.expand_dims(sentence_masked, 2), [1,1,self.d]), name='Lambda') # (S x N x d) return Lambda def model(self, E, E_LC, E_A, E_RC, dropout_prob, weights): with tf.name_scope('Position_Attention_Module'): Lambda = self.getLambda(E, E_LC, E_A, E_RC) with tf.name_scope('position_attention_weighted_memory'): M_w = tf.identity(tf.multiply(Lambda, E), name='M_w') with tf.name_scope('Content_Attention_Module'): with tf.name_scope('scores_of_words'): v_a = tf.identity(tf.reduce_sum(E_A, 1) / tf.math.count_nonzero(E_A, 1, dtype=tf.float32), name='v_a') # (S x d) v_s = tf.identity(tf.reduce_sum(E, 1) / tf.math.count_nonzero(E, 1, dtype=tf.float32), name='v_s') # (S x d) v_a_matrix = tf.tile(tf.expand_dims(v_a, 1), [1, self.FLAGS.N, 1], name='v_a_matrix') # (S x N x d) v_s_matrix = tf.tile(tf.expand_dims(v_s, 1), [1, self.FLAGS.N, 1], name='v_s_matrix') # (S x N x d) C = tf.where(tf.not_equal(E, tf.constant(0.0))[:,:,0], tf.reshape(tf.tensordot(tf.nn.tanh( tf.tensordot(tf.nn.dropout(M_w, rate=dropout_prob, seed=self.FLAGS.seed), weights['W_2'], [2,1]) + tf.tensordot(tf.nn.dropout(v_a_matrix, rate=dropout_prob, seed=self.FLAGS.seed), weights['W_3'], [2,1]) + tf.tensordot(tf.nn.dropout(v_s_matrix, rate=dropout_prob, seed=self.FLAGS.seed), weights['W_4'], [2,1]) + tf.nn.dropout(weights['b_1'], rate=dropout_prob, seed=self.FLAGS.seed)), weights['W_1'], [2, 1]), [-1, self.FLAGS.N]), -1e10*tf.ones([tf.shape(E)[0], self.FLAGS.N]), name='C') # (S x N) with tf.name_scope('attention_weights_for_scores'): A = tf.nn.softmax(C, name='A') with tf.name_scope('sentence_representation'): v_ts = tf.transpose(tf.matmul(tf.transpose(E, [0,2,1]), tf.expand_dims(A, 2)), [0,2,1], name='v_ts') # (S x d x 1) v_ns = tf.identity(v_ts + tf.reshape(v_s, [-1, 1, self.d]), name='v_ns') with tf.name_scope('MLP'): v_ms = tf.nn.dropout(tf.nn.tanh(tf.tensordot(tf.nn.dropout(v_ns, rate=dropout_prob, seed=self.FLAGS.seed), weights['W_5'], [2,1]) + weights['b_2']), rate=dropout_prob, seed=self.FLAGS.seed, name='v_ms') # (S x d x 1) with tf.name_scope('Linear_Layer'): v_L = tf.reshape(tf.tensordot(v_ms, weights['W_6'], [2,1]) + weights['b_3'], [-1, self.FLAGS.nr_cat], name='v_L') # (S x nr_cat) return v_L def trainModel(self, data, words_dict, train_data_length, train_data_length_glove, data_test=None, test_data_length=None, test_data_length_glove=None, save_to_files=False): tf.reset_default_graph() sentence_data = tf.placeholder(tf.int64, [None, None], name='sentence_data') aspect_data = tf.placeholder(tf.int64, [None, None], name='aspect_data') Y = tf.placeholder(tf.int64, [None, 3], name='Y') batch_size = tf.placeholder(tf.int64, name='batch_size') dropout_prob = tf.placeholder(tf.float32, name='dropout_prob') dictionary = tf.placeholder(tf.float32, [self.V, self.d], name='dictionary') data_len = tf.placeholder(tf.int32, None, name='data_len') data_len_glove = tf.placeholder(tf.int32, None, name='data_len_glove') iterator, next_batch, nr_batches, LC, A, RC = self.getMinibatches(batch_size, sentence_data, aspect_data, Y) with tf.name_scope('Embeddings'): E = tf.nn.embedding_lookup(dictionary, LC+A+RC, name='E') E_LC = tf.nn.embedding_lookup(dictionary, LC, name='E_LC') E_A = tf.nn.embedding_lookup(dictionary, A, name='E_A') E_RC = tf.nn.embedding_lookup(dictionary, RC, name='E_RC') weights = self.initWeights() logits = tf.identity(self.model(E, E_LC, E_A, E_RC, dropout_prob, weights), name='logits') prediction = tf.nn.softmax(logits, name='prediction') loss, loss_log, loss_log_update_op, loss_scalar, accuracy, accuracy_update_op, accuracy_scalar = self.getSummaries(logits, prediction, Y, data_len, data_len_glove) with tf.name_scope('TrainOp'): global_step = tf.Variable(0, trainable=False, name='global_step') optimizer = tf.train.GradientDescentOptimizer(self.FLAGS.learning_rate) train_op = optimizer.minimize(loss, global_step=global_step) with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess: if save_to_files: writer_train = tf.summary.FileWriter('./Results/logs/{0}/train/{1}'.format(self.__class__.__name__, datetime.datetime.now().strftime('%Y_%m_%d_%H_%M_%S')), sess.graph) writer_test = tf.summary.FileWriter('./Results/logs/{0}/test/{1}'.format(self.__class__.__name__, datetime.datetime.now().strftime('%Y_%m_%d_%H_%M_%S')), sess.graph) summary_op = tf.summary.merge_all() saver = tf.train.Saver(max_to_keep = 3, save_relative_paths=True) sess.run(tf.global_variables_initializer()) batches, _ = sess.run([nr_batches, iterator.initializer], feed_dict={'sentence_data:0':data[0], 'aspect_data:0':data[1], 'Y:0':data[2], 'batch_size:0':self.FLAGS.batch_size}) acc_test_best = 0 loss_test_best = np.inf step_best = -1 for epoch in range(self.FLAGS.num_epochs): sess.run(tf.local_variables_initializer()) for _ in range(batches): sentence_batch, aspect_batch, Y_batch = sess.run(next_batch) sess.run(train_op, feed_dict={'sentence_data:0':sentence_batch, 'aspect_data:0':aspect_batch, 'Y:0':Y_batch, 'dropout_prob:0':self.FLAGS.dropout_prob, 'dictionary:0':words_dict}) sess.run([accuracy_update_op, loss_log_update_op], feed_dict={'sentence_data:0':sentence_batch, 'aspect_data:0':aspect_batch, 'Y:0':Y_batch, 'dropout_prob:0':self.FLAGS.dropout_prob, 'dictionary:0':words_dict, 'data_len:0':train_data_length, 'data_len_glove:0':train_data_length_glove}) acc, cost, step = sess.run([accuracy, loss_log, global_step]) if save_to_files: writer_train.add_summary(sess.run(summary_op), step) writer_train.flush() if data_test is not None: sess.run(tf.local_variables_initializer()) acc_test, cost_test = sess.run([accuracy_update_op, loss_log_update_op], feed_dict={'sentence_data:0':data_test[0], 'aspect_data:0':data_test[1], 'Y:0':data_test[2], 'dropout_prob:0':0.0, 'dictionary:0':words_dict, 'data_len:0':test_data_length, 'data_len_glove:0':test_data_length_glove}) if acc_test > acc_test_best: acc_test_best = acc_test loss_test_best = cost_test step_best = step if save_to_files: if not os.path.exists('./Results/ckpt/{}'.format(self.__class__.__name__)): os.makedirs('./Results/ckpt/{}'.format(self.__class__.__name__)) saver.save(sess, './Results/ckpt/{}/model.ckpt'.format(self.__class__.__name__), step) improved = '*' else: improved = '' print('Step: {0:>6}, Train-Batch Accuracy: {1:>6.1%}, Test Accuracy: {2:>6.1%} Train-Batch Loss: {3} Test Loss: {4} {5}'. format(step, acc, acc_test, np.round(cost,5), np.round(cost_test,5), improved)) if save_to_files: writer_test.add_summary(sess.run(loss_scalar), step) writer_test.add_summary(sess.run(accuracy_scalar), step) writer_test.flush() else: print('Step: {0:>6}, Train-Batch Accuracy: {1:>6.1%}, Train Loss: {2}'. format(step, acc, np.round(cost,5))) if step_best + 1000 <= step: print('Stopping in step: {}, the best result in step: {}'.format(step, step_best)) break if save_to_files: writer_train.close() writer_test.close() return acc_test_best, loss_test_best def predict(self, data, words_dict, test_data_length, test_data_length_glove): tf.reset_default_graph() with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess: ckpt = tf.train.latest_checkpoint('./Results/ckpt/{}/'.format(self.__class__.__name__)) saver = tf.train.import_meta_graph(ckpt+'.meta') sess.run(tf.global_variables_initializer()) saver.restore(sess, ckpt) graph = tf.get_default_graph() global_step = graph.get_tensor_by_name('TrainOp/global_step:0') data_len = graph.get_tensor_by_name('data_len:0') data_len_glove = graph.get_tensor_by_name('data_len_glove:0') Y = graph.get_tensor_by_name('Y:0') logits = graph.get_tensor_by_name('logits:0') prediction = graph.get_tensor_by_name('prediction:0') _, accuracy = self.getAccuracy(prediction, Y, data_len, data_len_glove) _, loss = self.getLossOp(self.getLoss(logits, Y)) sess.run(tf.local_variables_initializer()) pred, acc, cost, step = sess.run([prediction, accuracy, loss, global_step], feed_dict={'sentence_data:0' : data[0], 'aspect_data:0' : data[1], 'Y:0' : data[2], 'dropout_prob:0':0.0, 'dictionary:0' : words_dict, 'data_len:0' : test_data_length, 'data_len_glove:0':test_data_length_glove}) print('Step: {0:>6}, Accuracy: {1:>6.1%}, Loss: {2}'. format(step, acc, np.round(cost,5))) return pred ########################################## if __name__ == '__main__': np.set_printoptions(threshold=np.inf, edgeitems=3, linewidth=120) FLAGS = tf.app.flags.FLAGS for name in list(FLAGS): if name not in ('showprefixforinfo',): delattr(FLAGS, name) tf.app.flags.DEFINE_string('f', '', 'kernel') tf.app.flags.DEFINE_boolean('logtostderr', True, 'tensorboard') tf.app.flags.DEFINE_float('train_proportion', 1.0, 'Train proportion for train/validation split') #0.75 tf.app.flags.DEFINE_boolean('shuffle', True, 'Shuffle datasets') tf.app.flags.DEFINE_integer('num_epochs', 600, 'Number of iterations') tf.app.flags.DEFINE_integer('seed', None, 'Random seed') tf.app.flags.DEFINE_float('mean', 0.0, 'Mean of normally initialized variables') tf.app.flags.DEFINE_float('stddev', 0.05, 'Standard deviation of normally initialized variables') tf.app.flags.DEFINE_float('dropout_prob', 0.3, 'Dropout probability') tf.app.flags.DEFINE_float('learning_rate', 0.001, 'Adam optimzier learning rate') tf.app.flags.DEFINE_integer('batch_size', 128, 'Batch size') tf.app.flags.DEFINE_integer('nr_cat', 3, 'Number of classification categories') tf.app.flags.DEFINE_string('embeddings_path', './Embeddings/glove.42B.300d.txt', 'Embedding path') tf.app.flags.DEFINE_boolean('train_model', True, 'Run Ont on train data') tf.app.flags.DEFINE_boolean('save_to_files', True, 'Whether to save checkpoints') tf.app.flags.DEFINE_boolean('predict_values', True, 'Run Ont on test data') tf.app.flags.DEFINE_string('train_data_path', './Data/ABSA-16_SB1_Restaurants_Train_Data.xml', 'Train data path') tf.app.flags.DEFINE_string('test_data_path', './Data/ABSA-16_SB1_Restaurants_Test_Gold.xml', 'Test data path') reader = Reader(FLAGS) embeddings, word2idx, idx2word = reader.readEmbeddings(FLAGS.embeddings_path) data_train, longest_sentence_train = reader.readData(FLAGS.train_data_path) data_test, longest_sentence_test = reader.readData(FLAGS.test_data_path) longest_sentence = max(longest_sentence_train, longest_sentence_test) tf.app.flags.DEFINE_integer('N', longest_sentence, 'Length of the longest sentence') data_idx_train = reader.transformSent2idx(data_train, word2idx) data_idx_test = reader.transformSent2idx(data_test, word2idx) model = BaseC(FLAGS) model.V, model.d = embeddings.shape model.longest_sentence = longest_sentence ## Train and evaluate models if FLAGS.train_model: print('Training...') a = datetime.datetime.now() model.trainModel(data_idx_train, embeddings, len(data_train), len(data_idx_train[0]), data_idx_test, len(data_test), len(data_idx_test[0]), save_to_files=FLAGS.save_to_files) b = datetime.datetime.now() print('Time:', b-a) ## Predict using trained models if FLAGS.predict_values: print('Prediction...') pred = model.predict(data_idx_test, embeddings, len(data_test), len(data_idx_test[0])) print('')