def test_generate_train_batches(self):
get_train_gen, train_size = self.data_manager.get_train_data_from_file(
[self.TRAIN_FILE])
batch_gen = DataManager.get_batch_generator(get_train_gen, 2)
new_batch_gen = DataManager.get_batch_generator(get_train_gen, 2)
# Assert that the new generator is a different object
# than the old generator.
assert new_batch_gen != batch_gen
assert train_size == 3
first_batch = batch_gen.__next__()
new_first_batch = new_batch_gen.__next__()
inputs, labels = first_batch
new_inputs, new_labels = new_first_batch
assert len(inputs) == len(new_inputs) == 2
assert len(labels) == len(new_labels) == 1
# Ensure output matches ground truth
assert_allclose(inputs[0], np.array([[2, 0], [5, 0]]))
assert_allclose(inputs[1], np.array([[3, 4], [6, 0]]))
assert_allclose(labels[0], np.array([[1, 0], [0, 1]]))
# Ensure both generators produce same results.
assert_allclose(inputs[0], new_inputs[0])
assert_allclose(inputs[1], new_inputs[1])
assert_allclose(labels[0], labels[0])
second_batch = batch_gen.__next__()
new_second_batch = new_batch_gen.__next__()
inputs, labels = second_batch
new_inputs, new_labels = new_second_batch
assert len(inputs) == len(new_inputs) == 2
assert len(labels) == len(new_labels) == 1
# Ensure output matches ground truth
assert_allclose(inputs[0], np.array([[7, 0]]))
assert_allclose(inputs[1], np.array([[8, 0]]))
assert_allclose(labels[0], np.array([[1, 0]]))
# Ensure both generators produce same results.
assert_allclose(inputs[0], new_inputs[0])
assert_allclose(inputs[1], new_inputs[1])
assert_allclose(labels[0], labels[0])
# Should raise a StopIteration
with self.assertRaises(StopIteration):
batch_gen.__next__()
new_batch_gen.__next__()
def _evaluate_on_validation(self, get_val_feature_generator, batch_size,
num_val_steps, session):
val_batch_gen = DataManager.get_batch_generator(
get_val_feature_generator, batch_size)
# Calculate the mean of the validation metrics
# over the validation set.
val_accuracies = []
val_losses = []
for val_batch in tqdm(val_batch_gen,
total=num_val_steps,
desc="Validation Batches Completed",
leave=False):
#Ignore the last batch if the size doesn't match
# if(len(val_batch) != batch_size):
# continue
feed_dict = self._get_validation_feed_dict(val_batch)
val_batch_acc, val_batch_loss = session.run(
[self.eval_metric, self.loss], feed_dict=feed_dict)
val_accuracies.append(val_batch_acc)
val_losses.append(val_batch_loss)
# Take the mean of the accuracies and losses.
# TODO/FIXME this assumes each batch is same shape, which
# is not necessarily true.
mean_val_accuracy = np.mean(val_accuracies)
mean_val_loss = np.mean(val_losses)
# Create a new Summary object with mean_val accuracy
# and mean_val_loss and add it to Tensorboard.
val_summary = tf.Summary(value=[
tf.Summary.Value(tag="_val_summaries/loss",
simple_value=mean_val_loss),
tf.Summary.Value(tag="_val_summaries/accuracy",
simple_value=mean_val_accuracy)
])
return mean_val_accuracy, mean_val_loss, val_summary
def train_gym(self,
gym_env: GymEnv,
batch_size=5,
num_episodes=100000,
expected_reward=200):
reward_sum = 0
running_reward = None
features = []
batched_features = None
gym_env.reset()
previous_mode = self._mode
self._mode = 'train'
global_step = 0
init_op = tf.global_variables_initializer()
gpu_options = tf.GPUOptions(allow_growth=True)
sess_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True,
log_device_placement=True)
with tf.Session(config=sess_config) as sess:
sess.run(init_op)
# self._setup_summaries(sess=sess)
feature = gym_env.reset_get_initial_observation()
for episode_number in tqdm(range(1, num_episodes + 1),
desc="Episodes completed"):
global_step = sess.run(self.global_step) + 1
i = 0
while (gym_env.is_done is False):
batched_features = DataManager.to_batch(feature)
model_prediction = sess.run(
self.prediction,
feed_dict=self._get_train_feed_dict(
batched_features, gym_env.is_done))
feature = gym_env.next_action(model_prediction)
features.append(feature)
reward_sum += gym_env.reward
gym_env.render()
i += 1
if episode_number % batch_size == 0:
batched_features = DataManager.to_batch(features)
loss = sess.run([self.optimizer],
feed_dict=self._get_train_feed_dict(
batched_features, gym_env.is_done))
running_reward = reward_sum if running_reward is None else running_reward * 0.99 + reward_sum * 0.01
print(
'Episode:%d Average reward for episode:%f Total average reward:%f'
% (episode_number, reward_sum / batch_size,
running_reward / batch_size))
if reward_sum / batch_size > 200:
print("Task solved in", episode_number, 'episodes')
break
features = []
reward_sum = 0
running_reward = None
feature = gym_env.reset_get_initial_observation()
gym_env.close()
self._mode = previous_mode
def test(self,
get_test_instance_generator,
model_load_dir,
batch_size,
num_test_steps=None):
"""
Load a serialized model and use it for prediction on a test
set (from a finite generator).
:param get_test_instance_generator: Function returning generator
This function should return a finite generator that produces features
for use in training.
:param model_load_dir: str
Path to a directory with serialized tensorflow checkpoints for the
model to be run. The most recent checkpoint will be loaded and used
for prediction.
:param batch_size: int
The number of features per batch produced by the generator.
:param num_test_steps: int
The number of steps (calculated by ceil(total # test examples / batch_size))
in testing. This does not have any effect on how much of the test data
is read; inference keeps going until the generator is exhausted. It
is used to set a total for the progress bar.
"""
previous_mode = self._mode
self._mode = 'predict'
if num_test_steps is None:
logger.info("num_test_steps is not set, pass in a value "
"to show a progress bar.")
gpu_options = tf.GPUOptions(allow_growth=True)
sess_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True,
log_device_placement=True)
with tf.Session(config=sess_config) as sess:
saver = tf.train.Saver()
logger.info(
"Getting latest checkpoint in {}".format(model_load_dir))
last_checkpoint = tf.train.latest_checkpoint(model_load_dir)
logger.info(
"Attempting to load checkpoint at {}".format(last_checkpoint))
saver.restore(sess, last_checkpoint)
logger.info("Successfully loaded {}!".format(last_checkpoint))
# Get a generator of test batches
test_batch_gen = DataManager.get_batch_generator(
get_test_instance_generator, batch_size)
y_pred = []
for batch in tqdm(test_batch_gen,
total=num_test_steps,
desc="Test Batches Completed"):
feed_dict = self._get_test_feed_dict(batch)
y_pred_batch = sess.run(self.prediction, feed_dict=feed_dict)
y_pred.append(y_pred_batch)
y_pred_flat = np.concatenate(y_pred, axis=0)
self._mode = previous_mode
return y_pred_flat
def train(self,
get_train_feature_generator,
get_val_feature_generator,
batch_size,
num_epochs,
num_train_steps_per_epoch,
num_val_steps,
val_period,
log_period,
save_period,
max_ckpts_to_keep=10,
patience=0):
"""
Train the model.
:param get_train_instance_generator: Function returning generator
This function should return a finite generator that produces
features for use in training.
:param get_val_feature_generator: Function returning generator
This function should return a finite generator that produces
features for use in validation.
:param batch_size: int
The number of features per batch produced by the generator.
:param num_train_steps_per_epoch: int
The number of training steps after which an epoch has passed.
:param num_epochs: int
The number of epochs to train for.
:param num_val_steps: int
The number of batches generated by the validation batch generator.
:param save_path: str
The input path to the tensorflow Saver responsible for
checkpointing.
:param log_path: str
The input path to the tensorflow SummaryWriter responsible for
logging the progress.
:param val_period: int, optional (default=250)
Number of steps between each evaluation of performance on the
held-out validation set.
:param log_period: int, optional (default=10)
Number of steps between each summary op evaluation.
:param save_period: int, optional (default=250)
Number of steps between each model checkpoint.
:param max_ckpts_to_keep: int, optional (default=10)
The maximum number of model to checkpoints to keep.
:param patience: int, optional (default=0)
The number of epochs with no improvement in validation loss
after which training will be stopped.
"""
previous_mode = self._mode
self._mode = 'train'
global_step = 0
init_op = tf.global_variables_initializer()
gpu_options = tf.GPUOptions(allow_growth=True)
sess_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True,
log_device_placement=True)
with tf.Session(config=sess_config) as sess:
sess.run(init_op)
self._setup_summaries(sess=sess)
epoch_validation_losses = []
# Iterate over a generator that returns batches.
for epoch in tqdm(range(num_epochs), desc="Epochs Completed"):
# Get a generator of train batches
train_batch_gen = DataManager.get_batch_generator(
get_train_feature_generator, batch_size)
# Iterate over the generated batches
for train_batch in tqdm(train_batch_gen,
total=num_train_steps_per_epoch,
desc="Train Batches Completed",
leave=False):
global_step = sess.run(self.global_step) + 1
feed_dict = self._get_train_feed_dict(train_batch)
# Do a gradient update, and log results to Tensorboard
# if necessary.
if global_step % log_period == 0:
# Record summary with gradient update
train_loss, _, train_summary = sess.run(
[self.loss, self.optimizer, self.train_summary_op],
feed_dict=feed_dict)
self._train_summary_writer.add_summary(
train_summary, global_step)
else:
# Do a gradient update without recording anything.
train_loss, _ = sess.run([self.loss, self.optimizer],
feed_dict=feed_dict)
if global_step % val_period == 0:
# Evaluate on validation data
val_acc, val_loss, val_summary = self._evaluate_on_validation(
get_val_feature_generator=get_val_feature_generator,
batch_size=batch_size,
num_val_steps=num_val_steps,
session=sess)
self._val_summary_writer.add_summary(
val_summary, global_step)
# Write a model checkpoint if necessary.
if global_step % save_period == 0:
ret = self._saver.save(sess,
self._save_dir + '/' +
self._name,
global_step=global_step)
logger.info('Saving final model @ ' +
os.path.abspath(ret))
# End of the epoch, so save the model and check validation loss,
# stopping if applicable.
model_path = self._saver.save(sess,
self._save_dir + '/' +
self._name,
global_step=global_step)
logger.info('Saving final model @ ' +
os.path.abspath(model_path))
val_acc, val_loss, val_summary = self._evaluate_on_validation(
get_val_feature_generator=get_val_feature_generator,
batch_size=batch_size,
num_val_steps=num_val_steps,
session=sess)
self._val_summary_writer.add_summary(val_summary, global_step)
epoch_validation_losses.append(val_loss)
# Get the lowest validation loss, with regards to the patience
# threshold.
patience_val_losses = epoch_validation_losses[:-(patience + 1)]
if patience_val_losses:
min_patience_val_loss = min(patience_val_losses)
else:
min_patience_val_loss = math.inf
if min_patience_val_loss <= val_loss:
# past loss was lower, so stop
logger.info("Validation loss of {} in last {} "
"epochs, which is lower than current "
"epoch validation loss of {}; stopping "
"early.".format(min_patience_val_loss,
patience, val_loss))
break
#Evaluate model specific evaluations
self._evaluate_model_parameters(sess)
# Done training!
logger.info("Finished {} epochs!".format(epoch + 1))
self._mode = previous_mode
return os.path.abspath(model_path)
def setUp(self):
super(TestDataManagerTest, self).setUp()
self.write_duplicate_questions_train_file()
self.write_duplicate_questions_test_file()
self.data_manager = DataManager(PairFeature)
self.data_manager.get_train_data_from_file([self.TRAIN_FILE])
class TestDataManagerTest(DuplicateTestCase):
@overrides
def setUp(self):
super(TestDataManagerTest, self).setUp()
self.write_duplicate_questions_train_file()
self.write_duplicate_questions_test_file()
self.data_manager = DataManager(PairFeature)
self.data_manager.get_train_data_from_file([self.TRAIN_FILE])
def test_get_test_data_default(self):
get_test_gen, test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE])
assert test_size == 3
test_gen = get_test_gen()
inputs1, labels1 = test_gen.__next__()
assert_allclose(inputs1[0], np.array([2, 1]))
assert_allclose(inputs1[1], np.array([1, 0]))
inputs2, labels2 = test_gen.__next__()
assert_allclose(inputs2[0], np.array([4, 0]))
assert_allclose(inputs2[1], np.array([5, 1]))
inputs3, labels3 = test_gen.__next__()
assert_allclose(inputs3[0], np.array([6, 0]))
assert_allclose(inputs3[1], np.array([7, 0]))
# Should raise a StopIteration
with self.assertRaises(StopIteration):
test_gen.__next__()
# Test that we can make a new test generator
new_test_gen = get_test_gen()
# Verify that the new and old generator are not the same object
assert new_test_gen != test_gen
new_inputs1, new_labels1 = new_test_gen.__next__()
assert_allclose(new_inputs1, inputs1)
assert_allclose(new_labels1, labels1)
new_inputs2, new_labels2 = new_test_gen.__next__()
assert_allclose(new_inputs2, inputs2)
assert_allclose(new_labels2, labels2)
new_inputs3, new_labels3 = new_test_gen.__next__()
assert_allclose(new_inputs3, inputs3)
assert_allclose(new_labels3, labels3)
# Should raise a StopIteration
with self.assertRaises(StopIteration):
new_test_gen.__next__()
def test_get_test_data_default_character(self):
get_test_gen, test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE], mode="character")
test_gen = get_test_gen()
assert test_size == 3
inputs1, labels = test_gen.__next__()
assert_allclose(
inputs1[0],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 10, 0, 0, 0],
[6, 9, 2, 7, 8, 3, 5, 4, 9, 4, 1, 10]]))
assert_allclose(
inputs1[1],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 9, 4, 1, 11],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]))
assert len(labels) == 0
inputs2, labels = test_gen.__next__()
assert_allclose(
inputs2[0],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 12, 19, 17, 18],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]))
assert_allclose(
inputs2[1],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 13, 0, 0, 0],
[6, 9, 2, 7, 8, 3, 5, 4, 9, 4, 1, 12]]))
assert len(labels) == 0
inputs3, labels = test_gen.__next__()
assert_allclose(
inputs3[0],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 14, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]))
assert_allclose(
inputs3[1],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 15, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]))
assert len(labels) == 0
# Should raise a StopIteration
with self.assertRaises(StopIteration):
test_gen.__next__()
def test_get_test_data_default_word_and_character(self):
get_test_gen, test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE], mode="word+character")
test_gen = get_test_gen()
assert test_size == 3
inputs1, labels = test_gen.__next__()
assert_allclose(inputs1[0], np.array([2, 1]))
assert_allclose(
inputs1[1],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 10, 0, 0, 0],
[6, 9, 2, 7, 8, 3, 5, 4, 9, 4, 1, 10]]))
assert_allclose(inputs1[2], np.array([1, 0]))
assert_allclose(
inputs1[3],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 9, 4, 1, 11],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]))
assert len(labels) == 0
inputs2, labels = test_gen.__next__()
assert_allclose(inputs2[0], np.array([4, 0]))
assert_allclose(
inputs2[1],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 12, 19, 17, 18],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]))
assert_allclose(inputs2[2], np.array([5, 1]))
assert_allclose(
inputs2[3],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 13, 0, 0, 0],
[6, 9, 2, 7, 8, 3, 5, 4, 9, 4, 1, 12]]))
assert len(labels) == 0
inputs3, labels = test_gen.__next__()
assert_allclose(inputs3[0], np.array([6, 0]))
assert_allclose(
inputs3[1],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 14, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]))
assert_allclose(inputs3[2], np.array([7, 0]))
assert_allclose(
inputs3[3],
np.array([[6, 9, 2, 7, 8, 3, 5, 4, 15, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]))
assert len(labels) == 0
# Should raise a StopIteration
with self.assertRaises(StopIteration):
test_gen.__next__()
def test_get_test_data_pad_with_max_lens(self):
get_test_gen, test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE], max_lengths={"num_sentence_words": 1})
test_gen = get_test_gen()
assert test_size == 3
inputs, labels = test_gen.__next__()
assert_allclose(inputs[0], np.array([2]))
assert_allclose(inputs[1], np.array([1]))
assert len(labels) == 0
inputs, labels = test_gen.__next__()
assert_allclose(inputs[0], np.array([4]))
assert_allclose(inputs[1], np.array([5]))
assert len(labels) == 0
inputs, labels = test_gen.__next__()
assert_allclose(inputs[0], np.array([6]))
assert_allclose(inputs[1], np.array([7]))
assert len(labels) == 0
# Should raise a StopIteration
with self.assertRaises(StopIteration):
test_gen.__next__()
def test_get_test_data_with_max_features(self):
get_test_gen, test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE], max_features=2)
test_gen = get_test_gen()
assert test_size == 2
inputs, labels = test_gen.__next__()
assert_allclose(inputs[0], np.array([2, 1]))
assert_allclose(inputs[1], np.array([1, 0]))
assert len(labels) == 0
inputs, labels = test_gen.__next__()
assert_allclose(inputs[0], np.array([4, 0]))
assert_allclose(inputs[1], np.array([5, 1]))
assert len(labels) == 0
# Should raise a StopIteration
with self.assertRaises(StopIteration):
test_gen.__next__()
def test_get_test_data_errors(self):
with self.assertRaises(ValueError):
self.data_manager.get_test_data_from_file(
[self.TEST_FILE],
max_lengths={"num_sentence_words": 1},
pad=False)
with self.assertRaises(ValueError):
self.data_manager.get_test_data_from_file(
[self.TEST_FILE], max_lengths={"some wrong key": 1})
def test_get_test_data_no_pad(self):
get_test_gen, test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE], pad=False)
test_gen = get_test_gen()
assert test_size == 3
inputs, labels = test_gen.__next__()
assert_allclose(inputs[0], np.array([2, 1, 2]))
assert_allclose(inputs[1], np.array([1]))
assert len(labels) == 0
inputs, labels = test_gen.__next__()
assert_allclose(inputs[0], np.array([4]))
assert_allclose(inputs[1], np.array([5, 1]))
assert len(labels) == 0
inputs, labels = test_gen.__next__()
assert_allclose(inputs[0], np.array([6]))
assert_allclose(inputs[1], np.array([7]))
assert len(labels) == 0
# Should raise a StopIteration
with self.assertRaises(StopIteration):
test_gen.__next__()
def test_generate_test_batches(self):
get_test_gen, test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE])
batch_gen = self.data_manager.get_batch_generator(get_test_gen, 2)
new_batch_gen = DataManager.get_batch_generator(get_test_gen, 2)
# Assert that the new generator is a different object
# than the old generator.
assert new_batch_gen != batch_gen
assert test_size == 3
first_batch = batch_gen.__next__()
new_first_batch = new_batch_gen.__next__()
inputs, labels = first_batch
new_inputs, new_labels = new_first_batch
assert len(inputs) == 2
assert len(labels) == 0
# Ensure output matches ground truth
assert_allclose(inputs[0], np.array([[2, 1], [4, 0]]))
assert_allclose(inputs[1], np.array([[1, 0], [5, 1]]))
# Ensure both generators produce same results.
assert_allclose(inputs[0], new_inputs[0])
assert_allclose(inputs[1], new_inputs[1])
second_batch = batch_gen.__next__()
new_second_batch = new_batch_gen.__next__()
inputs, labels = second_batch
new_inputs, new_labels = new_second_batch
assert len(inputs) == 2
assert len(labels) == 0
# Ensure output matches ground truth
assert_allclose(inputs[0], np.array([[6, 0]]))
assert_allclose(inputs[1], np.array([[7, 0]]))
# Ensure both generators produce same results.
assert_allclose(inputs[0], new_inputs[0])
assert_allclose(inputs[1], new_inputs[1])
with self.assertRaises(StopIteration):
batch_gen.__next__()
new_batch_gen.__next__()
def setUp(self):
super(TestDataManagerTrain, self).setUp()
self.write_duplicate_questions_train_file()
self.data_manager = DataManager(PairFeature)
from dhira.data.dataset.quora import QuoraDataset
from dhira.data.features.quora_feature import QuoraFeature
quora_dataset = QuoraDataset(name='quora-spacy',
train_files='../data/quora/processed/train_cleaned_train_split.csv',
val_files='../data/quora/processed/train_cleaned_val_split.csv',
test_files='../data/quora/processed/test_final.csv',
feature_type=QuoraFeature,
pad=True,
pickle_dir='../models/',
max_lengths={"num_sentence_words": 30})
# In[ ]:
data_manager = DataManager(dataset=quora_dataset)
# In[ ]:
get_train_data_gen, train_data_size = data_manager.get_train_data_from_file()
# In[ ]:
get_val_data_gen, val_data_size = data_manager.get_validation_data_from_file()
# In[ ]: