def test_eval(self):
x_dev = np.random.randint(0, VOCABULARY_SIZE, [BATCH_SIZE * 5, SEQUENCE_LENGTH])
y_dev = np.eye(2)[np.random.randint(0, 1, [BATCH_SIZE * 5])]
def make_eval_iter():
data_iter = tfmodels.data.utils.batch_iter(
list(zip(x_dev, y_dev)), BATCH_SIZE, 1, fill=True, seed=42)
return map(lambda batch: zip(*batch), data_iter)
with self.graph.as_default(), self.sess.as_default():
cnn = self._build_classifier()
ev = CNNClassifierEvaluator(cnn)
self.sess.run(tf.initialize_all_variables())
loss, acc, current_step = ev.eval(make_eval_iter())
loss2, acc2, current_step = ev.eval(make_eval_iter())
self.assertGreater(loss, 0)
self.assertGreater(acc, 0)
self.assertEqual(loss, loss2)
self.assertEqual(acc, acc2)
def test_eval(self):
x_dev = np.random.randint(0, VOCABULARY_SIZE,
[BATCH_SIZE * 5, SEQUENCE_LENGTH])
y_dev = np.eye(2)[np.random.randint(0, 1, [BATCH_SIZE * 5])]
def make_eval_iter():
data_iter = tfmodels.data.utils.batch_iter(list(zip(x_dev, y_dev)),
BATCH_SIZE,
1,
fill=True,
seed=42)
return map(lambda batch: zip(*batch), data_iter)
with self.graph.as_default(), self.sess.as_default():
cnn = self._build_classifier()
ev = CNNClassifierEvaluator(cnn)
self.sess.run(tf.initialize_all_variables())
loss, acc, current_step = ev.eval(make_eval_iter())
loss2, acc2, current_step = ev.eval(make_eval_iter())
self.assertGreater(loss, 0)
self.assertGreater(acc, 0)
self.assertEqual(loss, loss2)
self.assertEqual(acc, acc2)
print("Restoring checkpoint from {}".format(latest_checkpoint))
saver.restore(sess, latest_checkpoint)
# Initialize variables
sess.run(tf.initialize_all_variables())
# Training loop
for train_loss, train_acc, current_step, time_delta in trainer.train_loop(
train_data_iter):
examples_per_second = FLAGS.batch_size / time_delta
print("{}: step {}, loss {:g}, acc {:g} ({:g} examples/sec)".format(
datetime.now().isoformat(), current_step, train_loss, train_acc,
examples_per_second))
# Evaluate dev set
if current_step % FLAGS.evaluate_every == 0:
dev_iter = tfmodels.data.utils.xy_iter(x_dev, y_dev,
FLAGS.batch_size, 1)
mean_loss, acc, _ = evaluator.eval(dev_iter,
global_step=trainer.global_step)
print(
"{}: Step {}, Dev Accuracy: {:g}, Dev Mean Loss: {:g}".format(
datetime.now().isoformat(), current_step, acc, mean_loss))
# Checkpoint Model
if current_step % FLAGS.checkpoint_every == 0:
save_path = saver.save(sess,
checkpoint_file,
global_step=trainer.global_step)
print("Saved {}".format(save_path))
checkpoint_file = os.path.join(checkpoint_dir, "model.ckpt")
saver = tf.train.Saver(keep_checkpoint_every_n_hours=2)
# Initialization, optinally load from checkpoint
sess.run(tf.initialize_all_variables())
latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
if latest_checkpoint:
print("Restoring checkpoint from {}".format(latest_checkpoint))
saver.restore(sess, latest_checkpoint)
# Initialize variables
sess.run(tf.initialize_all_variables())
# Training loop
for train_loss, train_acc, current_step, time_delta in trainer.train_loop(train_data_iter):
examples_per_second = FLAGS.batch_size/time_delta
print("{}: step {}, loss {:g}, acc {:g} ({:g} examples/sec)".format(
datetime.now().isoformat(), current_step, train_loss, train_acc, examples_per_second))
# Evaluate dev set
if current_step % FLAGS.evaluate_every == 0:
dev_iter = tfmodels.data.utils.xy_iter(x_dev, y_dev, FLAGS.batch_size, 1)
mean_loss, acc, _ = evaluator.eval(dev_iter, global_step=trainer.global_step)
print("{}: Step {}, Dev Accuracy: {:g}, Dev Mean Loss: {:g}".format(
datetime.now().isoformat(), current_step, acc, mean_loss))
# Checkpoint Model
if current_step % FLAGS.checkpoint_every == 0:
save_path = saver.save(sess, checkpoint_file, global_step=trainer.global_step)
print("Saved {}".format(save_path))
