# Checkpoint directory. Tensorflow assumes this directory already exists so we need to create it checkpoint_dir = os.path.abspath(os.path.join(out_dir, "checkpoints")) checkpoint_prefix = os.path.join(checkpoint_dir, "model") if not os.path.exists(checkpoint_dir): # 如果路径不存在就创建一个 os.makedirs(checkpoint_dir) saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints) # Write vocabulary vocab_processor.save(os.path.join(checkpoint_dir, "vocab")) # Initialize all variables sess.run(tf.global_variables_initializer()) if not FLAGS.random: vocabulary = vocab_processor.vocabulary_ initW = None initW = data_helpers.load_embedding_vectors_word2vec(vocabulary, "./data/GoogleNews-vectors-negative300.bin", True) print("word2vec file has been loaded") # initW = data_helpers.load_embedding_vectors_glove(vocabulary, # "./data/glove.6B.300d.txt", FLAGS.embedding_dim) # print("glove file has been loaded") sess.run(rnn.W.assign(initW)) def train_step(x_batch, y_batch): """ A single training step """ feed_dict = { # 给placeholder喂数据 rnn.input_x: x_batch, rnn.input_y: y_batch, rnn.dropout_keep_prob: FLAGS.dropout_keep_prob, }
# Write vocabulary vocab_processor.save(os.path.join(out_dir, "vocab")) # Initialize all variables #from here to line 203 added sess.run(tf.global_variables_initializer()) if FLAGS.enable_word_embeddings and cfg['word_embeddings'][ 'default'] is not None: vocabulary = vocab_processor.vocabulary_ initW = None if embedding_name == 'word2vec': # load embedding vectors from the word2vec print("Load word2vec file {}".format( cfg['word_embeddings']['word2vec']['path'])) initW = data_helpers.load_embedding_vectors_word2vec( vocabulary, cfg['word_embeddings']['word2vec']['path'], cfg['word_embeddings']['word2vec']['binary']) print("word2vec file has been loaded") elif embedding_name == 'glove': # load embedding vectors from the glove print("Load glove file {}".format( cfg['word_embeddings']['glove']['path'])) initW = data_helpers.load_embedding_vectors_glove( vocabulary, cfg['word_embeddings']['glove']['path'], embedding_dimension) print("glove file has been loaded\n") sess.run(cnn.W.assign(initW)) # learning rate was added as one of the parameters def train_step(x_batch, y_batch, learning_rate): """
if not os.path.exists(checkpoint_dir): os.makedirs(checkpoint_dir) saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints) # Write vocabulary vocab_processor.save(os.path.join(out_dir, "vocab")) # Initialize all variables sess.run(tf.global_variables_initializer()) # Load word2vec embeddings (Uncomment for word2vec stuff, if it actually works) if FLAGS.embedding_path: print('Loading word2vec embeddings...') vocabulary = vocab_processor.vocabulary_ initW = data_helpers.load_embedding_vectors_word2vec( vocabulary, FLAGS.embedding_path, True) sess.run(cnn.W.assign(initW)) else: warnings.warn( 'Pre-trained word vectors are not being used for this session.' ) def train_step(x_batch, y_batch): """ A single training step """ feed_dict = { cnn.input_x: x_batch, cnn.input_y: y_batch, cnn.dropout_keep_prob: FLAGS.dropout_keep_prob }
num_filters=num_filters, l2_reg_lambda=l2_reg_lambda) # Define Training procedure global_step = tf.Variable(0, name="global_step", trainable=False) optimizer = tf.train.AdamOptimizer(1e-3) grads_and_vars = optimizer.compute_gradients(cnn.loss) train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step) # Initialize all variables sess.run(tf.global_variables_initializer()) initW = None # load embedding vectors from the word2vec initW = data_helpers.load_embedding_vectors_word2vec( vocabulary, word2vec_model_path, word2vec_binary) print("word2vec file has been loaded") sess.run(cnn.W.assign(initW)) def train_step(x_batch, y_batch): feed_dict = { cnn.input_x: x_batch, cnn.input_y: y_batch, cnn.dropout_keep_prob: dropout_keep_prob } _, step, loss, accuracy = sess.run( [train_op, global_step, cnn.loss, cnn.accuracy], feed_dict) time_str = datetime.datetime.now().isoformat() print("{}: step {}, loss {:g}, acc {:g}".format( time_str, step, loss, accuracy))
def train(x_train, y_train, vocab_processor, x_dev, y_dev): # Training # ================================================== with tf.Graph().as_default(): session_conf = tf.ConfigProto( allow_soft_placement=FLAGS.allow_soft_placement, log_device_placement=FLAGS.log_device_placement) sess = tf.Session(config=session_conf) with sess.as_default(): cnn = TextCNN( sequence_length=x_train.shape[1], #56 num_classes=y_train.shape[1], #2 vocab_size=len(vocab_processor.vocabulary_), #18765? 18758 √ embedding_size=FLAGS.embedding_dim, #128 filter_sizes=list(map( int, FLAGS.filter_sizes.split(","))), #[3,4,5] num_filters=FLAGS.num_filters, #128 l2_reg_lambda=FLAGS.l2_reg_lambda) #0 # Define Training procedure global_step = tf.Variable(0, name="global_step", trainable=False) optimizer = tf.train.AdamOptimizer(1e-3) grads_and_vars = optimizer.compute_gradients(cnn.loss) train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step) # Keep track of gradient values and sparsity (optional) grad_summaries = [] for g, v in grads_and_vars: if g is not None: grad_hist_summary = tf.summary.histogram( "{}/grad/hist".format(v.name), g) sparsity_summary = tf.summary.scalar( "{}/grad/sparsity".format(v.name), tf.nn.zero_fraction(g)) grad_summaries.append(grad_hist_summary) grad_summaries.append(sparsity_summary) grad_summaries_merged = tf.summary.merge(grad_summaries) # Output directory for models and summaries timestamp = str(int(time.time())) out_dir = os.path.abspath( os.path.join(os.path.curdir, "runs", timestamp)) print("Writing to {}\n".format(out_dir)) # Summaries for loss and accuracy loss_summary = tf.summary.scalar("loss", cnn.loss) acc_summary = tf.summary.scalar("accuracy", cnn.accuracy) # Train Summaries train_summary_op = tf.summary.merge( [loss_summary, acc_summary, grad_summaries_merged]) train_summary_dir = os.path.join(out_dir, "summaries", "train") train_summary_writer = tf.summary.FileWriter( train_summary_dir, sess.graph) # Dev summaries dev_summary_op = tf.summary.merge([loss_summary, acc_summary]) dev_summary_dir = os.path.join(out_dir, "summaries", "dev") dev_summary_writer = tf.summary.FileWriter(dev_summary_dir, sess.graph) # Checkpoint directory. Tensorflow assumes this directory already exists so we need to create it checkpoint_dir = os.path.abspath( os.path.join(out_dir, "checkpoints")) checkpoint_prefix = os.path.join(checkpoint_dir, "model") if not os.path.exists(checkpoint_dir): os.makedirs(checkpoint_dir) saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints) # Write vocabulary vocab_processor.save(os.path.join(out_dir, "vocab")) # Initialize all variables sess.run(tf.global_variables_initializer()) vocabulary = vocab_processor.vocabulary_ initW = None initW = data_helpers.load_embedding_vectors_word2vec( vocabulary, './GoogleNews-vectors-negative300.bin', True) print("word2vec file has been loaded") sess.run(cnn.W.assign(initW)) def train_step(x_batch, y_batch): """ A single training step """ feed_dict = { cnn.input_x: x_batch, cnn.input_y: y_batch, cnn.dropout_keep_prob: FLAGS.dropout_keep_prob #0.5 } _, step, summaries, loss, accuracy = sess.run([ train_op, global_step, train_summary_op, cnn.loss, cnn.accuracy ], feed_dict) time_str = datetime.datetime.now().isoformat() print("{}: step {}, loss {:g}, acc {:g}".format( time_str, step, loss, accuracy)) train_summary_writer.add_summary(summaries, step) def dev_step(x_batch, y_batch, writer=None): """ Evaluates model on a dev set """ feed_dict = { cnn.input_x: x_batch, cnn.input_y: y_batch, cnn.dropout_keep_prob: 1.0 } step, summaries, loss, accuracy = sess.run( [global_step, dev_summary_op, cnn.loss, cnn.accuracy], feed_dict) time_str = datetime.datetime.now().isoformat() print("{}: step {}, loss {:g}, acc {:g}".format( time_str, step, loss, accuracy)) if writer: writer.add_summary(summaries, step) # Generate batches batches = data_helpers.batch_iter(list(zip(x_train, y_train)), FLAGS.batch_size, FLAGS.num_epochs) # Training loop. For each batch... for batch in batches: #print (batch[0]) x_batch, y_batch = zip( *batch) #x_batch = 64 * 56 y_batch = 64 * 2 train_step(x_batch, y_batch) current_step = tf.train.global_step(sess, global_step) if current_step % FLAGS.evaluate_every == 0: #100次 print("\nEvaluation:") dev_step(x_dev, y_dev, writer=dev_summary_writer) print("") if current_step % FLAGS.checkpoint_every == 0: #100次 path = saver.save(sess, checkpoint_prefix, global_step=current_step) print("Saved model checkpoint to {}\n".format(path))
def train(self): # Training # ================================================== if self.x_train is None: self.preprocess() with tf.Graph().as_default(): session_conf = tf.ConfigProto( allow_soft_placement=FLAGS.allow_soft_placement, log_device_placement=FLAGS.log_device_placement) sess = tf.Session(config=session_conf) with sess.as_default(): cnn = TextCNN(sequence_length=self.x_train.shape[1], num_classes=self.y_train.shape[1], vocab_size=len(self.vocab_processor.vocabulary_), embedding_size=self.embedding_dimension, filter_sizes=list( map(int, FLAGS.filter_sizes.split(","))), num_filters=FLAGS.num_filters, l2_reg_lambda=FLAGS.l2_reg_lambda) # Define Training procedure print("Define Training procedure") global_step = tf.Variable(0, name="global_step", trainable=False) optimizer = tf.train.GradientDescentOptimizer( cnn.learning_rate) grads_and_vars = optimizer.compute_gradients(cnn.loss) train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step) # Keep track of gradient values and sparsity (optional) print("Keep track of gradient values and sparsity (optional)") grad_summaries = [] for g, v in grads_and_vars: if g is not None: grad_hist_summary = tf.summary.histogram( "{}/grad/hist".format(v.name), g) sparsity_summary = tf.summary.scalar( "{}/grad/sparsity".format(v.name), tf.nn.zero_fraction(g)) grad_summaries.append(grad_hist_summary) grad_summaries.append(sparsity_summary) grad_summaries_merged = tf.summary.merge(grad_summaries) # Output directory for models and summaries print("Output directory for models and summaries") out_dir = os.path.abspath( os.path.join(os.path.curdir, "runs", self.dataset_name, str(self.embedding_name))) print("Writing to {}\n".format(out_dir)) # Summaries for loss and accuracy loss_summary = tf.summary.scalar("loss", cnn.loss) acc_summary = tf.summary.scalar("accuracy", cnn.accuracy) # Train Summaries train_summary_op = tf.summary.merge( [loss_summary, acc_summary, grad_summaries_merged]) train_summary_dir = os.path.join(out_dir, "summaries", "train") train_summary_writer = tf.summary.FileWriter( train_summary_dir, sess.graph) # Dev summaries dev_summary_op = tf.summary.merge([loss_summary, acc_summary]) dev_summary_dir = os.path.join(out_dir, "summaries", "dev") dev_summary_writer = tf.summary.FileWriter( dev_summary_dir, sess.graph) # Checkpoint directory. Tensorflow assumes this directory already exists so we need to create it checkpoint_dir = os.path.abspath( os.path.join(out_dir, "checkpoints")) checkpoint_prefix = os.path.join(checkpoint_dir, "model") if not os.path.exists(checkpoint_dir): os.makedirs(checkpoint_dir) saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints) # Write vocabulary print("Write vocabulary") self.vocab_processor.save(os.path.join(out_dir, "vocab")) # Initialize all variables print("Initialize all variables") sess.run(tf.global_variables_initializer()) if self.embedding_name is not None and self.cfg is not None: vocabulary = self.vocab_processor.vocabulary_ initW = None if self.embedding_name == 'word2vec': # load embedding vectors from the word2vec print("Load word2vec file {}".format( self.cfg['word_embeddings']['word2vec']['path'])) initW = data_helpers.load_embedding_vectors_word2vec( vocabulary, self.cfg['word_embeddings']['word2vec']['path'], self.cfg['word_embeddings']['word2vec']['binary']) print("word2vec file has been loaded") elif self.embedding_name == 'glove': # load embedding vectors from the glove print("Load glove file {}".format( self.cfg['word_embeddings']['glove']['path'])) initW = data_helpers.load_embedding_vectors_glove( vocabulary, self.cfg['word_embeddings']['glove']['path'], self.embedding_dimension) print("glove file has been loaded\n") if initW is not None: sess.run(cnn.W.assign(initW)) else: print( "HIGH ALERT - cnn.W not assigned. initW is None\n") def train_step(x_batch, y_batch, learning_rate): """ A single training step """ feed_dict = { cnn.input_x: x_batch, cnn.input_y: y_batch, cnn.dropout_keep_prob: FLAGS.dropout_keep_prob, cnn.learning_rate: learning_rate } _, step, summaries, loss, accuracy = sess.run([ train_op, global_step, train_summary_op, cnn.loss, cnn.accuracy ], feed_dict) time_str = datetime.datetime.now().isoformat() if step % 50 == 0: print( "{}: step {}, loss {:g}, acc {:g}, learning_rate {:g}" .format(time_str, step, loss, accuracy, learning_rate)) train_summary_writer.add_summary(summaries, step) def dev_step(x_batch, y_batch, writer=None): """ Evaluates model on a dev set """ feed_dict = { cnn.input_x: x_batch, cnn.input_y: y_batch, cnn.dropout_keep_prob: 1.0 } step, summaries, loss, accuracy = sess.run( [global_step, dev_summary_op, cnn.loss, cnn.accuracy], feed_dict) time_str = datetime.datetime.now().isoformat() print("{}: step {}, loss {:g}, acc {:g}".format( time_str, step, loss, accuracy)) if writer: writer.add_summary(summaries, step) def get_learning_rate(decay_speed, counter): # # It uses dynamic learning rate with a high value at the beginning to speed up the training # max_learning_rate = 0.005 # min_learning_rate = 0.0001 # learning_rate = min_learning_rate + (max_learning_rate - min_learning_rate) * 0.25 * math.exp( # -counter / decay_speed) # # print("decay speed: {}. counter: {}. learning_rate: {}".format(decay_speed, counter, learning_rate)) # return learning_rate return 0.0005 # Generate batches print("Generate batches") batches = data_helpers.batch_iter( list(zip(self.x_train, self.y_train)), FLAGS.batch_size, FLAGS.num_epochs) decay_speed = FLAGS.decay_coefficient * len( self.y_train) / FLAGS.batch_size # Training loop. For each batch... print("Training loop. For each batch...") counter = 0 for batch in batches: learning_rate = get_learning_rate(decay_speed, counter) counter += 1 x_batch, y_batch = zip(*batch) train_step(x_batch, y_batch, learning_rate) current_step = tf.train.global_step(sess, global_step) if current_step % FLAGS.evaluate_every == 0: print("\nEvaluation:") dev_step(self.x_eval, self.y_eval, writer=dev_summary_writer) print() if current_step % FLAGS.checkpoint_every == 0: path = saver.save(sess, checkpoint_prefix, global_step=current_step) print("\tSaved model checkpoint to {}\n".format(path)) print("End training. counter: {}. batch size: {}\n".format( counter, FLAGS.batch_size))
def main(): import time start_time = time.time() FLAGS = flagClass() with open("config.yml", 'r') as ymlfile: cfg = yaml.load(ymlfile) dataset_name = cfg["datasets"]["default"] if FLAGS.enable_word_embeddings and cfg['word_embeddings'][ 'default'] is not None: embedding_name = cfg['word_embeddings']['default'] embedding_dimension = cfg['word_embeddings'][embedding_name][ 'dimension'] else: embedding_dimension = FLAGS.embedding_dim # Data Preparation # ================================================== # Load data print("Loading data...") datasets = None if dataset_name == "mrpolarity": datasets = data_helpers.get_datasets_mrpolarity( cfg["datasets"][dataset_name]["positive_data_file"]["path"], cfg["datasets"][dataset_name]["negative_data_file"]["path"]) elif dataset_name == 'spamham': datasets = data_helpers.get_datasets_mrpolarity( cfg["datasets"][dataset_name]["spam_file"]["path"], cfg["datasets"][dataset_name]["ham_file"]["path"]) elif dataset_name == "20newsgroup": datasets = data_helpers.get_datasets_20newsgroup( subset="train", categories=cfg["datasets"][dataset_name]["categories"], shuffle=cfg["datasets"][dataset_name]["shuffle"], random_state=cfg["datasets"][dataset_name]["random_state"]) elif dataset_name == "dbpedia": datasets = data_helpers.get_datasets_dbpedia( cfg["datasets"][dataset_name]["train_file"]["path"], cfg["datasets"][dataset_name]["train_file"]["limit"]) elif dataset_name == "email": datasets = data_helpers.get_datasets_email( container_path=cfg["datasets"][dataset_name]["container_path"], categories=cfg["datasets"][dataset_name]["categories"], shuffle=cfg["datasets"][dataset_name]["shuffle"], random_state=cfg["datasets"][dataset_name]["random_state"]) elif dataset_name == "localdata": datasets = data_helpers.get_datasets_localdata( container_path=cfg["datasets"][dataset_name]["container_path"], categories=cfg["datasets"][dataset_name]["categories"], shuffle=cfg["datasets"][dataset_name]["shuffle"], random_state=cfg["datasets"][dataset_name]["random_state"]) x_text, y = data_helpers.load_data_labels(datasets) # Build vocabulary # To limit memory usage, you can cut off input text to first 40 words # Other research has shown that first 40 words in text (IMDB dataset?) # were representative of the content of the sentence for classification # purposes - Comment out one of the two lines below # max_document_length = max([len(x.split(" ")) for x in x_text]) max_document_length = 40 # read up to 40 words from each sentence vocab_processor = learn.preprocessing.VocabularyProcessor( max_document_length) x = np.array(list(vocab_processor.fit_transform(x_text))) # Randomly shuffle data np.random.seed(10) shuffle_indices = np.random.permutation(np.arange(len(y))) x_shuffled = x[shuffle_indices] y_shuffled = y[shuffle_indices] # Split train/test set # TODO: This is very crude, should use cross-validation dev_sample_index = -1 * int(FLAGS.dev_sample_percentage * float(len(y))) x_train, x_dev = x_shuffled[:dev_sample_index], x_shuffled[ dev_sample_index:] y_train, y_dev = y_shuffled[:dev_sample_index], y_shuffled[ dev_sample_index:] print("Vocabulary Size: {:d}".format(len(vocab_processor.vocabulary_))) print("Train/Dev split: {:d}/{:d}".format(len(y_train), len(y_dev))) print('Sequence_length={}'.format(x_train.shape[1])) # Training # ================================================== with tf.Graph().as_default(): session_conf = tf.ConfigProto( allow_soft_placement=FLAGS.allow_soft_placement, log_device_placement=FLAGS.log_device_placement) sess = tf.Session(config=session_conf) with sess.as_default(): cnn = TextCNN(sequence_length=x_train.shape[1], num_classes=y_train.shape[1], vocab_size=len(vocab_processor.vocabulary_), embedding_size=embedding_dimension, filter_sizes=list( map(int, FLAGS.filter_sizes.split(","))), num_filters=FLAGS.num_filters, l2_reg_lambda=FLAGS.l2_reg_lambda) # Define Training procedure global_step = tf.Variable(0, name="global_step", trainable=False) optimizer = tf.train.AdamOptimizer(cnn.learning_rate) grads_and_vars = optimizer.compute_gradients(cnn.loss) train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step) # Keep track of gradient values and sparsity (optional) grad_summaries = [] for g, v in grads_and_vars: if g is not None: grad_hist_summary = tf.summary.histogram( "{}/grad/hist".format(v.name), g) sparsity_summary = tf.summary.scalar( "{}/grad/sparsity".format(v.name), tf.nn.zero_fraction(g)) grad_summaries.append(grad_hist_summary) grad_summaries.append(sparsity_summary) grad_summaries_merged = tf.summary.merge(grad_summaries) # Output directory for models and summaries timestamp = str(int(time.time())) out_dir = os.path.abspath( os.path.join(os.path.curdir, "runs", timestamp)) print("Writing to {}\n".format(out_dir)) # Summaries for loss and accuracy loss_summary = tf.summary.scalar("loss", cnn.loss) acc_summary = tf.summary.scalar("accuracy", cnn.accuracy) # Train Summaries train_summary_op = tf.summary.merge( [loss_summary, acc_summary, grad_summaries_merged]) train_summary_dir = os.path.join(out_dir, "summaries", "train") train_summary_writer = tf.summary.FileWriter( train_summary_dir, sess.graph) # Dev summaries dev_summary_op = tf.summary.merge([loss_summary, acc_summary]) dev_summary_dir = os.path.join(out_dir, "summaries", "dev") dev_summary_writer = tf.summary.FileWriter(dev_summary_dir, sess.graph) # Checkpoint directory. Tensorflow assumes this directory already exists so we need to create it checkpoint_dir = os.path.abspath( os.path.join(out_dir, "checkpoints")) checkpoint_prefix = os.path.join(checkpoint_dir, "model") if not os.path.exists(checkpoint_dir): os.makedirs(checkpoint_dir) saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints) # Write vocabulary vocab_processor.save(os.path.join(out_dir, "vocab")) # Initialize all variables sess.run(tf.global_variables_initializer()) if FLAGS.enable_word_embeddings and cfg['word_embeddings'][ 'default'] is not None: vocabulary = vocab_processor.vocabulary_ initW = None if embedding_name == 'word2vec': # load embedding vectors from the word2vec print("Load word2vec file {}".format( cfg['word_embeddings']['word2vec']['path'])) initW = data_helpers.load_embedding_vectors_word2vec( vocabulary, cfg['word_embeddings']['word2vec']['path'], cfg['word_embeddings']['word2vec']['binary']) print("word2vec file has been loaded") elif embedding_name == 'glove': # load embedding vectors from the glove print("Load glove file {}".format( cfg['word_embeddings']['glove']['path'])) initW = data_helpers.load_embedding_vectors_glove( vocabulary, cfg['word_embeddings']['glove']['path'], embedding_dimension) print("glove file has been loaded\n") sess.run(cnn.W.assign(initW)) def train_step(x_batch, y_batch, learning_rate): """ A single training step """ feed_dict = { cnn.input_x: x_batch, cnn.input_y: y_batch, cnn.dropout_keep_prob: FLAGS.dropout_keep_prob, cnn.learning_rate: learning_rate } _, step, summaries, loss, accuracy = sess.run([ train_op, global_step, train_summary_op, cnn.loss, cnn.accuracy ], feed_dict) time_str = datetime.datetime.now().isoformat() print("{}: step {}, loss {:g}, acc {:g}, learning_rate {:g}". format(time_str, step, loss, accuracy, learning_rate)) train_summary_writer.add_summary(summaries, step) def dev_step(x_batch, y_batch, writer=None): """ Evaluates model on a dev set """ feed_dict = { cnn.input_x: x_batch, cnn.input_y: y_batch, cnn.dropout_keep_prob: 1.0 } step, summaries, loss, accuracy, gr = sess.run([ global_step, dev_summary_op, cnn.loss, cnn.accuracy, cnn.grad ], feed_dict) time_str = datetime.datetime.now().isoformat() print("{}: step {}, loss {:g}, acc {:g}, gr {}".format( time_str, step, loss, accuracy, gr)) if writer: writer.add_summary(summaries, step) # Generate batches batches = data_helpers.batch_iter(list(zip(x_train, y_train)), FLAGS.batch_size, FLAGS.num_epochs) print("Number of epochs: {}".format(FLAGS.num_epochs)) num_batches_per_epoch = int( (len(list(zip(x_train, y_train))) - 1) / FLAGS.batch_size) + 1 print("Batches per epoch: {}".format(num_batches_per_epoch)) print("Batch size: {}".format(FLAGS.batch_size)) # It uses dynamic learning rate with a high value at the beginning to speed up the training max_learning_rate = 0.005 min_learning_rate = 0.0001 decay_speed = FLAGS.decay_coefficient * len( y_train) / FLAGS.batch_size # Training loop. For each batch... counter = 0 for batch in batches: learning_rate = min_learning_rate + ( max_learning_rate - min_learning_rate) * math.exp( -counter / decay_speed) counter += 1 x_batch, y_batch = zip(*batch) train_step(x_batch, y_batch, learning_rate) current_step = tf.train.global_step(sess, global_step) if current_step % FLAGS.evaluate_every == 0: print("\nEvaluation:") dev_step(x_dev, y_dev, writer=dev_summary_writer) print("") if current_step % FLAGS.checkpoint_every == 0: path = saver.save(sess, checkpoint_prefix, global_step=current_step) print("Saved model checkpoint to {}\n".format(path)) print("runtime was " + str(time.time() - start_time))
def train(x_train, y_train, vocab_processor, x_test, y_test, x_valid, y_valid, report_df, current_fold, df_valid): # Training # ================================================== with tf.Graph().as_default(): session_conf = tf.ConfigProto( allow_soft_placement=FLAGS.allow_soft_placement, log_device_placement=FLAGS.log_device_placement) sess = tf.Session(config=session_conf) with sess.as_default(): cnn = TextCNN(sequence_length=x_train.shape[1], num_classes=y_train.shape[1], vocab_size=len(vocab_processor.vocabulary_), embedding_size=FLAGS.embedding_dim, filter_sizes=list( map(int, FLAGS.filter_sizes.split(","))), num_filters=FLAGS.num_filters, l2_reg_lambda=FLAGS.l2_reg_lambda) # Define Training procedure global_step = tf.Variable(0, name="global_step", trainable=False) optimizer = tf.train.AdamOptimizer(1e-3) grads_and_vars = optimizer.compute_gradients(cnn.loss) train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step) # Keep track of gradient values and sparsity (optional) grad_summaries = [] for g, v in grads_and_vars: if g is not None: grad_hist_summary = tf.summary.histogram( "{}/grad/hist".format(v.name), g) sparsity_summary = tf.summary.scalar( "{}/grad/sparsity".format(v.name), tf.nn.zero_fraction(g)) grad_summaries.append(grad_hist_summary) grad_summaries.append(sparsity_summary) grad_summaries_merged = tf.summary.merge(grad_summaries) # Output directory for models and summaries timestamp = str(int(time.time())) out_dir = os.path.abspath( os.path.join(os.path.curdir, "runs", timestamp)) print("Writing to {}\n".format(out_dir)) # Summaries for loss and accuracy loss_summary = tf.summary.scalar("loss", cnn.loss) acc_summary = tf.summary.scalar("accuracy", cnn.accuracy) # Train Summaries train_summary_op = tf.summary.merge( [loss_summary, acc_summary, grad_summaries_merged]) train_summary_dir = os.path.join(out_dir, "summaries", "train") train_summary_writer = tf.summary.FileWriter( train_summary_dir, sess.graph) # Dev summaries dev_summary_op = tf.summary.merge([loss_summary, acc_summary]) dev_summary_dir = os.path.join(out_dir, "summaries", "dev") dev_summary_writer = tf.summary.FileWriter(dev_summary_dir, sess.graph) # Checkpoint directory. Tensorflow assumes this directory already exists so we need to create it checkpoint_dir = os.path.abspath( os.path.join(out_dir, "checkpoints")) checkpoint_prefix = os.path.join(checkpoint_dir, "model") if not os.path.exists(checkpoint_dir): os.makedirs(checkpoint_dir) saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints) # Write vocabulary vocab_processor.save(os.path.join(out_dir, "vocab")) # Initialize all variables sess.run(tf.global_variables_initializer()) ## code by Sven vocabulary = vocab_processor.vocabulary_ initW = None # load embedding vectors from the word2vec print("Load word2vec file {}".format(FLAGS.embedding_path)) initW = data_helpers.load_embedding_vectors_word2vec( vocabulary, FLAGS.embedding_path, FLAGS.embedding_bin) print("word2vec file has been loaded") sess.run(cnn.W.assign(initW)) ## end of change def train_step(x_batch, y_batch): """ A single training step """ feed_dict = { cnn.input_x: x_batch, cnn.input_y: y_batch, cnn.dropout_keep_prob: FLAGS.dropout_keep_prob } _, step, summaries, loss, accuracy = sess.run([ train_op, global_step, train_summary_op, cnn.loss, cnn.accuracy ], feed_dict) time_str = datetime.datetime.now().isoformat() print("{}: step {}, loss {:g}, acc {:g}".format( time_str, step, loss, accuracy)) train_summary_writer.add_summary(summaries, step) def dev_step(x_batch, y_batch, writer=None): """ Evaluates model on a dev set """ feed_dict = { cnn.input_x: x_batch, cnn.input_y: y_batch, cnn.dropout_keep_prob: 1.0 } step, summaries, loss, accuracy = sess.run( [global_step, dev_summary_op, cnn.loss, cnn.accuracy], feed_dict) # compute f1 score labels = np.array(y_batch)[:, 1] prediction = cnn.predictions.eval(feed_dict) f1 = f1_score(labels, prediction) time_str = datetime.datetime.now().isoformat() print("{}: step {}, loss {:g}, acc {:g} f1 {:g}".format( time_str, step, loss, accuracy, f1)) if writer: writer.add_summary(summaries, step) return accuracy, f1 # Generate batches batches = data_helpers.batch_iter(list(zip(x_train, y_train)), FLAGS.batch_size, FLAGS.num_epochs) # Training loop. For each batch... for batch in batches: x_batch, y_batch = zip(*batch) train_step(x_batch, y_batch) current_step = tf.train.global_step(sess, global_step) if current_step % FLAGS.evaluate_every == 0: print("\nTest evaluation:") dev_step(x_test, y_test, writer=dev_summary_writer) print("") if current_step % FLAGS.checkpoint_every == 0: path = saver.save(sess, checkpoint_prefix, global_step=current_step) print("Saved model checkpoint to {}\n".format(path)) print("\nEnd test evaluation:") acc_score_in_cat, f1_score_in_cat = dev_step( x_test, y_test, writer=dev_summary_writer) print("\nEnd val evaluation:") acc_score_out_of_cat, f1_score_out_of_cat = dev_step( x_valid, y_valid, writer=dev_summary_writer) cat_dict = {} for category in df_valid.category.unique(): mask = df_valid['category'] == category category_df = df_valid[mask] category_array = np.array(category_df['text'].values.tolist()) category_label = np.array( category_df['labels'].values.tolist()) print(category_array.shape) print(x_valid.shape) print(category_label.shape) print(y_valid.shape) cat_acc, cat_f1 = dev_step(category_array, category_label) cat_dict[category + "-f1"] = cat_f1 cat_dict[category + "-acc"] = cat_acc name = 'fold-' + str(current_fold) report_df.append([ name, acc_score_in_cat, f1_score_in_cat, acc_score_out_of_cat, f1_score_out_of_cat ] + list(cat_dict.values())) return list(cat_dict.keys())
checkpoint_prefix = os.path.join(checkpoint_dir, "model") if not os.path.exists(checkpoint_dir): os.makedirs(checkpoint_dir) saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints) # Write vocabulary vocab_processor.save(os.path.join(out_dir, "vocab")) # Initialize all variables sess.run(tf.global_variables_initializer()) if FLAGS.model_type != "rand": if FLAGS.word2vec != None: print("Load word2vec file {}\n".format(FLAGS.word2vec)) initW = data_helpers.load_embedding_vectors_word2vec( vocab_processor.vocabulary_, FLAGS.word2vec, True) print("The word2vec file {} is loaded!\n".format( FLAGS.word2vec)) sess.run(cnn.W.assign(initW)) if FLAGS.model_type == "multichannel": sess.run(cnn.W_MultiChannel.assign(initW)) def train_step(x_batch, y_batch): """ A single training step """ feed_dict = { cnn.input_x: x_batch, cnn.input_y: y_batch,