def main(_):
assert FLAGS.checkpoint_dir, "--checkpoint_dir is required."
assert FLAGS.source_test_path, "--source_test_path is required."
assert FLAGS.target_test_path, "--target_test_path is required."
assert FLAGS.reference_test_path, "--reference_test_path is required."
assert FLAGS.source_vocab_path, "--souce_vocab_path is required."
assert FLAGS.target_vocab_path, "--target_vocab_path is required."
# Read vocabularies.
source_vocab, _ = utils.initialize_vocabulary(FLAGS.source_vocab_path)
target_vocab, _ = utils.initialize_vocabulary(FLAGS.target_vocab_path)
# Read test set.
source_sentences, target_sentences, references = utils.read_data_with_ref(
FLAGS.source_test_path, FLAGS.target_test_path,
FLAGS.reference_test_path)
# Convert sentences to token ids sequences.
source_sentences_ids = [
utils.sentence_to_token_ids(sent, source_vocab, FLAGS.max_seq_length)
for sent in source_sentences
]
target_sentences_ids = [
utils.sentence_to_token_ids(sent, target_vocab, FLAGS.max_seq_length)
for sent in target_sentences
]
utils.reset_graph()
with tf.Session() as sess:
# Restore saved model.
utils.restore_model(sess, FLAGS.checkpoint_dir)
# Recover placeholders and ops for evaluation.
x_source = sess.graph.get_tensor_by_name("x_source:0")
source_seq_length = sess.graph.get_tensor_by_name(
"source_seq_length:0")
x_target = sess.graph.get_tensor_by_name("x_target:0")
target_seq_length = sess.graph.get_tensor_by_name(
"target_seq_length:0")
labels = sess.graph.get_tensor_by_name("labels:0")
placeholders = [
x_source, source_seq_length, x_target, target_seq_length, labels
]
probs = sess.graph.get_tensor_by_name("feed_forward/output/probs:0")
# Run evaluation.
evaluate(sess, source_sentences, target_sentences, references,
source_sentences_ids, target_sentences_ids, probs,
placeholders)
def main(_):
assert FLAGS.checkpoint_dir, "--checkpoint_dir is required."
assert FLAGS.extract_dir, "--extract_dir is required."
assert FLAGS.source_vocab_path, "--source_vocab_path is required."
assert FLAGS.target_vocab_path, "--target_vocab_path is required."
assert FLAGS.source_output_path, "--source_output_path is required."
assert FLAGS.target_output_path, "--target_output_path is required."
assert FLAGS.score_output_path, "--score_output_path is required."
assert FLAGS.source_language, "--source_language is required."
assert FLAGS.target_language, "--target_language is required."
# Read vocabularies.
source_vocab, _ = utils.initialize_vocabulary(FLAGS.source_vocab_path)
target_vocab, _ = utils.initialize_vocabulary(FLAGS.target_vocab_path)
source_vocab_words = read_vocabulary(FLAGS.source_vocab_path)
target_vocab_words = read_vocabulary(FLAGS.target_vocab_path)
utils.reset_graph()
with tf.Session() as sess:
# Restore saved model.
utils.restore_model(sess, FLAGS.checkpoint_dir)
# Recover placeholders and ops for extraction.
x_source = sess.graph.get_tensor_by_name("x_source:0")
source_seq_length = sess.graph.get_tensor_by_name(
"source_seq_length:0")
x_target = sess.graph.get_tensor_by_name("x_target:0")
target_seq_length = sess.graph.get_tensor_by_name(
"target_seq_length:0")
labels = sess.graph.get_tensor_by_name("labels:0")
placeholders = [
x_source, source_seq_length, x_target, target_seq_length, labels
]
probs = sess.graph.get_tensor_by_name("feed_forward/output/probs:0")
with open(FLAGS.source_output_path, mode="w", encoding="utf-8") as source_output_file, \
open(FLAGS.target_output_path, mode="w", encoding="utf-8") as target_output_file, \
open(FLAGS.score_output_path, mode="w", encoding="utf-8") as score_output_file:
source_docs, target_docs = read_docs(FLAGS.extract_dir,
source_vocab, target_vocab)
pairs = extract_pairs(sess, source_docs, target_docs,
source_sentences_ids, target_sentences_ids,
probs, placeholders)
#for source_path, target_path in zip(source_paths, target_paths):
for source_path, target_path in itertools.product(
source_paths, target_paths):
#print("paths", source_path, target_path)
# Read sentences from articles.
source_sentences, target_sentences = read_articles(
source_path, target_path)
# Convert sentences to token ids sequences.
source_sentences_ids = [
utils.sentence_to_token_ids(sent, source_vocab,
FLAGS.max_seq_length)
for sent in source_sentences
]
target_sentences_ids = [
utils.sentence_to_token_ids(sent, target_vocab,
FLAGS.max_seq_length)
for sent in target_sentences
]
# Extract sentence pairs.
pairs = extract_pairs(sess, source_sentences, target_sentences,
source_sentences_ids,
target_sentences_ids, probs,
placeholders)
if not pairs:
continue
for source_sentence, target_sentence, score in pairs:
source_output_file.write(source_sentence)
target_output_file.write(target_sentence)
score_output_file.write(str(score) + "\n")
def build_graph(self):
# Reset previous graph.
reset_graph()
# Placeholders.
x_source = tf.placeholder(tf.int32,
shape=[None, None],
name="x_source")
source_seq_length = tf.placeholder(tf.int32,
shape=[None],
name="source_seq_length")
x_target = tf.placeholder(tf.int32,
shape=[None, None],
name="x_target")
target_seq_length = tf.placeholder(tf.int32,
shape=[None],
name="target_seq_length")
labels = tf.placeholder(tf.float32,
shape=[None],
name="labels")
input_dropout = tf.placeholder_with_default(1.0,
shape=[],
name="input_dropout")
output_dropout = tf.placeholder_with_default(1.0,
shape=[],
name="output_dropout")
decision_threshold = tf.placeholder_with_default(0.5,
shape=[],
name="decision_threshold")
# Embedding layer.
with tf.variable_scope("embeddings"):
if self.config.source_embeddings_path is not None and self.config.target_embeddings_path is not None:
source_pretrained_embeddings,\
target_pretrained_embeddings = get_pretrained_embeddings(
source_embeddings_path,
target_embeddings_path,
source_vocab,
target_vocab)
assert source_pretrained_embeddings.shape[1] == target_pretrained_embeddings.shape[1]
self.config.embedding_size = source_pretrained_embeddings.shape[1]
if self.config.fix_pretrained:
source_embeddings = tf.get_variable(
name="source_embeddings_matrix",
shape=[self.config.source_vocab_size, self.config.embedding_size],
initializer=tf.constant_initializer(source_pretrained_embeddings),
trainable=False)
target_embeddings = tf.get_variable(
name="target_embeddings_matrix",
shape=[self.config.target_vocab_size, self.config.embedding_size],
initializer=tf.constant_initializer(target_pretrained_embeddings),
trainable=False)
else:
source_embeddings = tf.get_variable(
name="source_embeddings_matrix",
shape=[self.config.source_vocab_size, self.config.embedding_size],
initializer=tf.constant_initializer(source_pretrained_embeddings))
target_embeddings = tf.get_variable(
name="target_embeddings_matrix",
shape=[self.config.target_vocab_size, self.config.embedding_size],
initializer=tf.constant_initializer(target_pretrained_embeddings))
else:
source_embeddings = tf.get_variable(
name="source_embeddings_matrix",
shape=[self.config.source_vocab_size, self.config.embedding_size])
target_embeddings = tf.get_variable(
name="target_embeddings_matrix",
shape=[self.config.target_vocab_size, self.config.embedding_size])
source_rnn_inputs = tf.nn.embedding_lookup(source_embeddings, x_source)
target_rnn_inputs = tf.nn.embedding_lookup(target_embeddings, x_target)
source_rnn_inputs = tf.nn.dropout(source_rnn_inputs,
keep_prob=input_dropout,
name="source_seq_embeddings")
target_rnn_inputs = tf.nn.dropout(target_rnn_inputs,
keep_prob=input_dropout,
name="target_seq_embeddings")
# BiRNN encoder.
with tf.variable_scope("birnn") as scope:
if self.config.use_lstm:
cell_fw = tf.nn.rnn_cell.LSTMCell(self.config.state_size, use_peepholes=True)
cell_bw = tf.nn.rnn_cell.LSTMCell(self.config.state_size, use_peepholes=True)
else:
cell_fw = tf.nn.rnn_cell.GRUCell(self.config.state_size)
cell_bw = tf.nn.rnn_cell.GRUCell(self.config.state_size)
cell_fw = tf.nn.rnn_cell.DropoutWrapper(cell_fw, output_keep_prob=output_dropout)
cell_bw = tf.nn.rnn_cell.DropoutWrapper(cell_bw, output_keep_prob=output_dropout)
if self.config.num_layers > 1:
if self.config.use_lstm:
cell_fw = tf.nn.rnn_cell.MultiRNNCell([tf.nn.rnn_cell.LSTMCell(self.config.state_size,
use_peepholes=True)
for _ in range(self.config.num_layers)])
cell_bw = tf.nn.rnn_cell.MultiRNNCell([tf.nn.rnn_cell.LSTMCell(self.config.state_size,
use_peepholes=True)
for _ in range(self.config.num_layers)])
else:
cell_fw = tf.nn.rnn_cell.MultiRNNCell([tf.nn.rnn_cell.GRUCell(self.config.state_size)
for _ in range(self.config.num_layers)])
cell_bw = tf.nn.rnn_cell.MultiRNNCell([tf.nn.rnn_cell.GRUCell(self.config.state_size)
for _ in range(self.config.num_layers)])
with tf.variable_scope(scope):
source_rnn_outputs, source_final_state = tf.nn.bidirectional_dynamic_rnn(
cell_fw=cell_fw,
cell_bw=cell_bw,
inputs=source_rnn_inputs,
sequence_length=source_seq_length,
dtype=tf.float32)
with tf.variable_scope(scope, reuse=True):
target_rnn_outputs, target_final_state = tf.nn.bidirectional_dynamic_rnn(
cell_fw=cell_fw,
cell_bw=cell_bw,
inputs=target_rnn_inputs,
sequence_length=target_seq_length,
dtype=tf.float32)
self.config.state_size *= 2
# Mean and max pooling only work for 1 layer BiRNN.
if self.config.use_mean_pooling:
source_final_state = self.average_pooling(source_rnn_outputs, source_seq_length)
target_final_state = self.average_pooling(target_rnn_outputs, target_seq_length)
elif self.config.use_max_pooling:
source_final_state = self.max_pooling(source_rnn_outputs)
target_final_state = self.max_pooling(target_rnn_outputs)
else:
source_final_state_fw, source_final_state_bw = source_final_state
target_final_state_fw, target_final_state_bw = target_final_state
if self.config.num_layers > 1:
source_final_state_fw = source_final_state_fw[-1]
source_final_state_bw = source_final_state_bw[-1]
target_final_state_fw = target_final_state_fw[-1]
target_final_state_bw = target_final_state_bw[-1]
if self.config.use_lstm:
source_final_state_fw = source_final_state_fw.h
source_final_state_bw = source_final_state_bw.h
target_final_state_fw = target_final_state_fw.h
target_final_state_bw = target_final_state_bw.h
source_final_state = tf.concat([source_final_state_fw, source_final_state_bw],
axis=1, name="source_final_state_ph")
target_final_state = tf.concat([target_final_state_fw, target_final_state_bw],
axis=1)
# Feed-forward neural network.
with tf.variable_scope("feed_forward"):
h_multiply = tf.multiply(source_final_state, target_final_state)
h_abs_diff = tf.abs(tf.subtract(source_final_state, target_final_state))
W_1 = tf.get_variable(name="W_1",
shape=[self.config.state_size, self.config.hidden_size])
W_2 = tf.get_variable(name="W_2",
shape=[self.config.state_size, self.config.hidden_size])
b_1 = tf.get_variable(name="b_1",
shape=[self.config.hidden_size],
initializer=tf.constant_initializer(0.0))
h_semantic = tf.tanh(tf.matmul(h_multiply, W_1) + tf.matmul(h_abs_diff, W_2) + b_1)
W_3 = tf.get_variable(name="W_3",
shape=[self.config.hidden_size, 1])
b_2 = tf.get_variable(name="b_2",
shape=[1],
initializer=tf.constant_initializer(0.0))
logits = tf.matmul(h_semantic, W_3) + b_2
logits = tf.squeeze(logits,
name="logits")
# Sigmoid output layer.
with tf.name_scope("output"):
probs = tf.sigmoid(logits,
name="probs")
predicted_class = tf.cast(tf.greater(probs, decision_threshold),
tf.float32,
name="predicted_class")
# Loss.
with tf.name_scope("cross_entropy"):
losses = tf.nn.sigmoid_cross_entropy_with_logits(
logits=logits,
labels=labels,
name="cross_entropy_per_sequence")
mean_loss = tf.reduce_mean(losses,
name="cross_entropy_loss")
# Optimization.
with tf.name_scope("optimization"):
global_step = tf.Variable(initial_value=0,
trainable=False,
name="global_step")
optimizer = tf.train.AdamOptimizer(self.config.learning_rate)
trainable_variables = tf.trainable_variables()
gradients = tf.gradients(mean_loss, trainable_variables,
name="gradients")
clipped_gradients, global_norm = tf.clip_by_global_norm(
gradients,
clip_norm=self.config.max_gradient_norm,
name="clipped_gradients")
train_op = optimizer.apply_gradients(zip(clipped_gradients, trainable_variables),
global_step=global_step)
# Evaluation metrics.
accuracy = tf.metrics.accuracy(labels, predicted_class,
name="accuracy")
precision = tf.metrics.precision(labels, predicted_class,
name="precision")
recall = tf.metrics.recall(labels, predicted_class,
name="recall")
# Add summaries.
tf.summary.scalar("loss", mean_loss)
tf.summary.scalar("global_norm", global_norm)
tf.summary.scalar("accuracy", accuracy[0])
tf.summary.scalar("precision", precision[0])
tf.summary.scalar("recall", recall[0])
tf.summary.scalar("logits" + "/sparsity", tf.nn.zero_fraction(logits))
tf.summary.histogram("logits" + "/activations", logits)
tf.summary.histogram("probs", probs)
# Add histogram for trainable variables.
for var in trainable_variables:
tf.summary.histogram(var.op.name, var)
# Add histogram for gradients.
for grad, var in zip(clipped_gradients, trainable_variables):
if grad is not None:
tf.summary.histogram(var.op.name + "/gradients", grad)
# Assign placeholders and operations.
self.x_source = x_source
self.x_target = x_target
self.source_seq_length = source_seq_length
self.target_seq_length = target_seq_length
self.labels = labels
self.input_dropout = input_dropout
self.output_dropout = output_dropout
self.decision_threshold = decision_threshold
self.train_op = train_op
self.probs = probs
self.predicted_class = predicted_class
self.mean_loss = mean_loss
self.accuracy = accuracy
self.precision = precision
self.recall = recall
self.summaries = tf.summary.merge_all()
self.saver = tf.train.Saver(tf.global_variables(), max_to_keep=5)
def main(_):
assert FLAGS.checkpoint_dir, "--checkpoint_dir is required."
assert FLAGS.extract_dir, "--extract_dir is required."
assert FLAGS.source_vocab_path, "--source_vocab_path is required."
assert FLAGS.target_vocab_path, "--target_vocab_path is required."
assert FLAGS.source_output_path, "--source_output_path is required."
assert FLAGS.target_output_path, "--target_output_path is required."
assert FLAGS.score_output_path, "--score_output_path is required."
assert FLAGS.source_language, "--source_language is required."
assert FLAGS.target_language, "--target_language is required."
# Read vocabularies.
source_vocab, _ = utils.initialize_vocabulary(FLAGS.source_vocab_path)
target_vocab, _ = utils.initialize_vocabulary(FLAGS.target_vocab_path)
# Read source and target paths for sentence extraction.
source_paths = []
target_paths = []
for file in os.listdir(FLAGS.extract_dir):
if file.endswith(FLAGS.source_language):
source_paths.append(os.path.join(FLAGS.extract_dir, file))
elif file.endswith(FLAGS.target_language):
target_paths.append(os.path.join(FLAGS.extract_dir, file))
source_paths.sort()
target_paths.sort()
utils.reset_graph()
with tf.Session() as sess:
# Restore saved model.
utils.restore_model(sess, FLAGS.checkpoint_dir)
# Recover placeholders and ops for extraction.
x_source = sess.graph.get_tensor_by_name("x_source:0")
source_seq_length = sess.graph.get_tensor_by_name("source_seq_length:0")
x_target = sess.graph.get_tensor_by_name("x_target:0")
target_seq_length = sess.graph.get_tensor_by_name("target_seq_length:0")
labels = sess.graph.get_tensor_by_name("labels:0")
placeholders = [x_source, source_seq_length, x_target, target_seq_length, labels]
probs = sess.graph.get_tensor_by_name("feed_forward/output/probs:0")
source_final_state_ph = sess.graph.get_tensor_by_name("birnn/source_final_state_ph:0")
with open(FLAGS.source_output_path, mode="w", encoding="utf-8") as source_output_file,\
open(FLAGS.target_output_path, mode="w", encoding="utf-8") as target_output_file,\
open(FLAGS.score_output_path, mode="w", encoding="utf-8") as score_output_file:
for source_path, target_path in zip(source_paths, target_paths):
# Read sentences from articles.
source_sentences, target_sentences = read_articles(source_path, target_path)
# Convert sentences to token ids sequences.
source_sentences_ids = [utils.sentence_to_token_ids(sent, source_vocab, FLAGS.max_seq_length)
for sent in source_sentences]
target_sentences_ids = [utils.sentence_to_token_ids(sent, target_vocab, FLAGS.max_seq_length)
for sent in target_sentences]
# Extract sentence pairs.
pairs = extract_pairs(sess, source_sentences, target_sentences,
source_sentences_ids, target_sentences_ids,
probs, placeholders, source_final_state_ph)
if not pairs:
continue
for source_sentence, target_sentence, score in pairs:
source_output_file.write(source_sentence)
target_output_file.write(target_sentence)
score_output_file.write(str(score) + "\n")
