def run_evaluate(self, inputs, targets, epoch):
"""Runs evaluation on validation inputs and targets.
Optionally: - writes summary to Tensorboard.
"""
if self.summary_dir is not None:
eval_losses = []
for inps, trgts in summarizer_model_utils.minibatches(
inputs, targets, self.batch_size):
fd, sl, s2 = self.get_feed_dict(inps, trgts)
eval_loss = self.sess.run([self.eval_loss], feed_dict=fd)
eval_losses.append(eval_loss)
avg_eval_loss = self.sess.run(tf.reduce_mean(eval_losses))
print('Eval_loss: {}\n'.format(avg_eval_loss))
eval_summ = self.sess.run([self.eval_summary], feed_dict=fd)
self.eval_writer.add_summary(eval_summ, epoch)
else:
eval_losses = []
for inps, trgts in summarizer_model_utils.minibatches(
inputs, targets, self.batch_size):
fd, sl, s2 = self.get_feed_dict(inps, trgts)
eval_loss = self.sess.run([self.eval_loss], feed_dict=fd)
eval_losses.append(eval_loss)
avg_eval_loss = self.sess.run(tf.reduce_mean(eval_losses))
print('Eval_loss: {}\n'.format(avg_eval_loss))
def run_epoch(self, inputs, targets, epoch):
"""Runs a single epoch.
Returns the average loss value on the epoch."""
batch_size = self.batch_size
nbatches = (len(inputs) + batch_size - 1) // batch_size
losses = []
for i, (inps, trgts) in enumerate(
summarizer_model_utils.minibatches(inputs, targets,
batch_size)):
if inps is not None and trgts is not None:
fd, sl, s2 = self.get_feed_dict(inps, trgts=trgts)
if i % 10 == 0 and self.summary_dir is not None:
_, train_loss, training_summ = self.sess.run([
self.train_op, self.train_loss, self.training_summary
],
feed_dict=fd)
self.training_writer.add_summary(training_summ,
epoch * nbatches + i)
else:
_, train_loss = self.sess.run(
[self.train_op, self.train_loss], feed_dict=fd)
if i % 2 == 0 or i == (nbatches - 1):
print('Iteration: {} of {}\ttrain_loss: {:.4f}'.format(
i, nbatches - 1, train_loss))
losses.append(train_loss)
else:
print('Minibatch empty.')
continue
avg_loss = self.sess.run(tf.reduce_mean(losses))
print('Average Score for this Epoch: {}'.format(avg_loss))
return avg_loss