def dev_step(x_batch1, x_batch2, y_batch):
total_loss = []
step = 0
test_batch = datahelper.batch_iter(
list(zip(x_batch1, x_batch2, y_batch)), FLAGS.batch_size,
1)
for test_data in test_batch:
x_dev_batch1, x_dev_batch2, y_dev_batch = zip(*test_data)
weights_d = []
for i in range(len(y_dev_batch)):
weights_d.append(1.0)
feed_dict = {
rnn.is_training: False,
rnn.input_x1: x_dev_batch1,
rnn.input_x2: x_dev_batch2,
rnn.input_y: y_dev_batch,
rnn.dropout_keep_prob: 1.0,
rnn.weights: weights_d
}
summaries, step, loss = sess.run(
[dev_summary_op, rnn.global_step, rnn.loss], feed_dict)
total_loss.append(loss)
loss = np.mean(total_loss, axis=0)
time_str = datetime.datetime.now().isoformat()
print("Test: {}:step {}, loss: {:g}".format(
time_str, step, loss))
log_write.write("Test: {}: step {}, loss {:g}.\n".format(
time_str, step, loss))
log_write.flush()
dev_summary_writer.add_summary(summaries, step)
def predict(x_pre1, x_pre2):
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():
x_test1 = x_pre1 # tf.nn.embedding_lookup(word_embedding, x_pre1)
x_test2 = x_pre2 # tf.nn.embedding_lookup(word_embedding, x_pre2)
checkpoint_dir = os.path.abspath(
os.path.join(os.path.curdir, "checkpoint", "1532884392_image"))
model_path = os.path.join(checkpoint_dir, "model-15100.meta")
saver = tf.train.import_meta_graph(model_path)
if os.path.exists(checkpoint_dir):
print("Restoring Variables from Checkpoint for rnn")
saver.restore(sess, tf.train.latest_checkpoint(checkpoint_dir))
else:
print("Can't find the checkpoint.going to stop")
return
graph = tf.get_default_graph()
batches = datahelper.batch_iter(list(zip(x_test1, x_test2)), FLAGS.batch_size,
1, shuffle=False)
result_list = []
for test_data in batches:
x_dev_batch1, x_dev_batch2 = zip(*test_data)
weights = []
for i in range(len(x_dev_batch1)):
weights.append(1.0)
result = graph.get_tensor_by_name("outputs/predict:0")
logits = sess.run(result,
{"is_training:0": False, "input_x1:0": x_dev_batch1, "input_x2:0": x_dev_batch2,
"dropout_keep_prob:0": 1.0, "weights:0": weights})
result_list = np.concatenate([np.array(result_list), np.array(logits)], axis=0)
return np.array(result_list)
def train(x_train1, x_dev1, x_train2, x_dev2, y_train, y_dev, word_embedding,
max_len, vocab_size):
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():
rnn = MV_RNN(FLAGS.learning_rate, FLAGS.batch_size,
FLAGS.decay_steps, FLAGS.decay_rate, max_len,
FLAGS.hidden_size, vocab_size, FLAGS.embedding_dim)
learning_rate = tf.train.exponential_decay(rnn.learning_rate,
rnn.global_step,
rnn.decay_steps,
rnn.decay_rate,
staircase=True)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
grads_and_vars = optimizer.compute_gradients(rnn.loss_val)
train_op = optimizer.apply_gradients(grads_and_vars,
global_step=rnn.global_step)
embedding = tf.constant(word_embedding, tf.float32)
t_assign_embedding = tf.assign(rnn.Embedding, embedding)
sess.run(t_assign_embedding)
# Keep track of gradient values and sparity
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)
timestamp = str(int(time.time()))
out_dir = os.path.abspath(
os.path.join(os.path.curdir, "runs", timestamp))
loss_summary = tf.summary.scalar("loss", rnn.loss_val)
train_summary_op = tf.summary.merge(
[loss_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_summary_op = tf.summary.merge([loss_summary])
dev_summary_dir = os.path.join(out_dir, "summaries", "dev")
dev_summary_writer = tf.summary.FileWriter(dev_summary_dir,
sess.graph)
sess.run(tf.global_variables_initializer())
checkpoint_dir = os.path.abspath(
os.path.join(os.path.curdir, "checkpoint",
timestamp + "_image"))
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
log_file = checkpoint_dir + "\\log.txt"
log_write = open(log_file, 'w', encoding='utf-8')
checkpoint_prefix = os.path.join(checkpoint_dir, "model")
saver = tf.train.Saver(max_to_keep=FLAGS.num_checkpoints)
def train_step(x_batch1, x_batch2, y_batch):
weights_t = []
for i in range(len(y_batch)):
if y_batch[i] == 1:
weights_t.append(1.0)
else:
weights_t.append(1)
feed_dict = {
rnn.is_training: True,
rnn.input_x1: x_batch1,
rnn.input_x2: x_batch2,
rnn.input_y: y_batch,
rnn.dropout_keep_prob: FLAGS.dropout_keep_prob,
rnn.weights: weights_t
}
summaries, _, step, loss = sess.run(
[train_summary_op, train_op, rnn.global_step, rnn.loss],
feed_dict)
time_str = datetime.datetime.now().isoformat()
print("Train: {}: step {}, loss {:g}. ".format(
time_str, step, loss))
tmp = "Train: {}: step {}, loss {:g}.\n".format(
time_str, step, loss)
log_write.write(tmp)
log_write.flush()
train_summary_writer.add_summary(summaries, step)
def dev_step(x_batch1, x_batch2, y_batch):
total_loss = []
step = 0
test_batch = datahelper.batch_iter(
list(zip(x_batch1, x_batch2, y_batch)), FLAGS.batch_size,
1)
for test_data in test_batch:
x_dev_batch1, x_dev_batch2, y_dev_batch = zip(*test_data)
weights_d = []
for i in range(len(y_dev_batch)):
weights_d.append(1.0)
feed_dict = {
rnn.is_training: False,
rnn.input_x1: x_dev_batch1,
rnn.input_x2: x_dev_batch2,
rnn.input_y: y_dev_batch,
rnn.dropout_keep_prob: 1.0,
rnn.weights: weights_d
}
summaries, step, loss = sess.run(
[dev_summary_op, rnn.global_step, rnn.loss], feed_dict)
total_loss.append(loss)
loss = np.mean(total_loss, axis=0)
time_str = datetime.datetime.now().isoformat()
print("Test: {}:step {}, loss: {:g}".format(
time_str, step, loss))
log_write.write("Test: {}: step {}, loss {:g}.\n".format(
time_str, step, loss))
log_write.flush()
dev_summary_writer.add_summary(summaries, step)
batches = datahelper.batch_iter(
list(zip(x_train1, x_train2, y_train)), FLAGS.batch_size,
FLAGS.num_epochs)
for batch in batches:
x_batch1, x_batch2, y_batch = zip(*batch)
train_step(x_batch1, x_batch2, y_batch)
current_step = tf.train.global_step(sess, rnn.global_step)
if current_step % FLAGS.evaluate_every == 0:
print("\nEvaluation:")
dev_step(x_dev1, x_dev2, y_dev)
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))
log_write.close()
