def train_loop_sl(self, sess, batcher, valid_batcher, summary_writer, flags):
"""Runs model training."""
step, losses = 0, []
while step < flags.max_run_steps:
next_batch = batcher.next()
summaries, loss, train_step = self.run_train_step(sess, next_batch)
losses.append(loss)
summary_writer.add_summary(summaries, train_step)
step += 1
# Display current training loss
if step % flags.display_freq == 0:
avg_loss = lib.compute_avg(losses, summary_writer, "avg_loss",
train_step)
tf.logging.info("Train step %d: avg_loss %f" % (train_step, avg_loss))
losses = []
summary_writer.flush()
# Run evaluation on validation set
if step % flags.valid_freq == 0:
valid_losses = []
for _ in xrange(flags.num_valid_batch):
next_batch = valid_batcher.next()
valid_loss = self.run_eval_step(sess, next_batch)
valid_losses.append(valid_loss)
gstep = self.get_global_step(sess)
avg_valid_loss = lib.compute_avg(valid_losses, summary_writer,
"valid_loss", gstep)
tf.logging.info("\tValid step %d: avg_loss %f" % (gstep,
avg_valid_loss))
summary_writer.flush()
def train_loop(self, sess, batcher, valid_batcher, summary_writer):
"""Runs model training."""
step, losses, accuracies = 0, [], []
while step < FLAGS.max_run_steps:
next_batch = batcher.next()
summaries, loss, accuracy, train_step = self.run_train_step(
sess, next_batch)
losses.append(loss)
accuracies.append(accuracy)
summary_writer.add_summary(summaries, train_step)
step += 1
# Display current training loss
if step % FLAGS.display_freq == 0:
avg_loss = lib.compute_avg(losses, summary_writer, "avg_loss",
train_step)
avg_acc = lib.compute_avg(accuracies, summary_writer,
"avg_acc", train_step)
tf.logging.info("Train step %d: avg_loss %f avg_acc %f" %
(train_step, avg_loss, avg_acc))
losses, accuracies = [], []
summary_writer.flush()
# Run evaluation on validation set
if step % FLAGS.valid_freq == 0:
valid_losses, valid_accs = [], []
for _ in xrange(FLAGS.num_valid_batch):
next_batch = valid_batcher.next()
valid_loss, valid_acc = self.run_eval_step(
sess, next_batch)
valid_losses.append(valid_loss)
valid_accs.append(valid_acc)
gstep = self.get_global_step(sess)
avg_valid_loss = lib.compute_avg(valid_losses, summary_writer,
"valid_loss", gstep)
avg_valid_acc = lib.compute_avg(valid_accs, summary_writer,
"valid_accuracy", gstep)
tf.logging.info("\tValid step %d: avg_loss %f avg_acc %f" %
(step, avg_valid_loss, avg_valid_acc))
summary_writer.flush()
def run_valid_steps(self, sess, data_batcher, num_valid_batch,
summary_writer):
losses = []
for _ in xrange(num_valid_batch):
next_batch = data_batcher.next()
loss = self.run_eval_step(sess, next_batch)
losses.append(loss)
step = self.get_global_step(sess)
valid_loss = lib.compute_avg(losses, summary_writer, "valid_loss", step)
tf.logging.info("\tValid step %d: avg_loss %f" % (step, valid_loss))
def TrainLoop(model, sess, batcher, valid_batcher, summary_writer, flags):
"""Runs model training."""
step, losses = 0, []
while step < flags.max_run_steps:
next_batch = batcher.next()
summaries, loss, train_step = model.run_train_step(sess, next_batch)
losses.append(loss)
summary_writer.add_summary(summaries, train_step)
step += 1
# Display current training loss
if step % flags.display_freq == 0:
avg_loss = lib.compute_avg(losses, summary_writer, "avg_loss", train_step)
tf.logging.info("Train step %d: avg_loss %f" % (train_step, avg_loss))
losses = []
summary_writer.flush()
# Run evaluation on validation set
if step % flags.valid_freq == 0:
model.run_valid_steps(sess, valid_batcher, flags.num_valid_batch,
summary_writer)
summary_writer.flush()