def setUp(self):
super(TestBiMPM, self).setUp()
self.write_duplicate_questions_train_file()
self.write_duplicate_questions_validation_file()
self.write_duplicate_questions_test_file()
self.data_manager = DataManager(STSInstance)
self.batch_size = 3
self.get_train_gen, self.train_size = self.data_manager.get_train_data_from_file(
[self.TRAIN_FILE], mode="word+character")
self.get_val_gen, self.val_size = self.data_manager.get_validation_data_from_file(
[self.VALIDATION_FILE], mode="word+character")
self.get_test_gen, self.test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE], mode="word+character")
self.embedding_manager = EmbeddingManager(
self.data_manager.data_indexer)
self.word_embedding_dim = 5
self.word_embedding_matrix = self.embedding_manager.get_embedding_matrix(
self.word_embedding_dim)
self.char_embedding_dim = 2
self.char_embedding_matrix = self.embedding_manager.get_embedding_matrix(
self.char_embedding_dim)
self.char_rnn_hidden_size = 6
self.context_rnn_hidden_size = 3
self.aggregation_rnn_hidden_size = 4
self.dropout_ratio = 0.1
self.config_dict = {
"mode":
"train",
"word_vocab_size":
self.data_manager.data_indexer.get_vocab_size(),
"word_embedding_dim":
self.word_embedding_dim,
"word_embedding_matrix":
self.word_embedding_matrix,
"char_vocab_size":
self.data_manager.data_indexer.get_vocab_size(
namespace="characters"),
"char_embedding_dim":
self.char_embedding_dim,
"char_embedding_matrix":
self.char_embedding_matrix,
"char_rnn_hidden_size":
self.char_rnn_hidden_size,
"fine_tune_embeddings":
False,
"context_rnn_hidden_size":
self.context_rnn_hidden_size,
"aggregation_rnn_hidden_size":
self.aggregation_rnn_hidden_size,
"dropout_ratio":
self.dropout_ratio
}
self.num_train_steps_per_epoch = int(
math.ceil(self.train_size / self.batch_size))
self.num_val_steps = int(math.ceil(self.val_size / self.batch_size))
self.num_test_steps = int(math.ceil(self.test_size / self.batch_size))
def setUp(self):
super(TestSiameseMatchingBiLSTM, self).setUp()
self.write_duplicate_questions_train_file()
self.write_duplicate_questions_validation_file()
self.write_duplicate_questions_test_file()
self.data_manager = DataManager(STSInstance)
self.batch_size = 2
self.get_train_gen, self.train_size = self.data_manager.get_train_data_from_file(
[self.TRAIN_FILE])
self.get_val_gen, self.val_size = self.data_manager.get_validation_data_from_file(
[self.VALIDATION_FILE])
self.get_test_gen, self.test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE])
self.embedding_manager = EmbeddingManager(self.data_manager.data_indexer)
self.word_embedding_dim = 5
self.embedding_matrix = self.embedding_manager.get_embedding_matrix(
self.word_embedding_dim)
self.rnn_hidden_size = 6
self.output_keep_prob = 1.0
self.share_encoder_weights = True
self.config_dict = {
"mode": "train",
"word_vocab_size": self.data_manager.data_indexer.get_vocab_size(),
"word_embedding_dim": self.word_embedding_dim,
"fine_tune_embeddings": False,
"word_embedding_matrix": self.embedding_matrix,
"rnn_hidden_size": self.rnn_hidden_size,
"output_keep_prob": self.output_keep_prob,
"share_encoder_weights": self.share_encoder_weights
}
self.num_train_steps_per_epoch = int(math.ceil(self.train_size / self.batch_size))
self.num_val_steps = int(math.ceil(self.val_size / self.batch_size))
self.num_test_steps = int(math.ceil(self.test_size / self.batch_size))
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 setUp(self):
super(TestSiameseBiLSTM, self).setUp()
self.write_duplicate_questions_train_file()
self.write_duplicate_questions_validation_file()
self.write_duplicate_questions_test_file()
self.data_manager = DataManager(STSInstance)
self.batch_size = 2
self.get_train_gen, self.train_size = self.data_manager.get_train_data_from_file(
[self.TRAIN_FILE])
self.get_val_gen, self.val_size = self.data_manager.get_validation_data_from_file(
[self.VALIDATION_FILE])
self.get_test_gen, self.test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE])
self.embedding_manager = EmbeddingManager(self.data_manager.data_indexer)
self.word_embedding_dim = 5
self.embedding_matrix = self.embedding_manager.get_embedding_matrix(
self.word_embedding_dim)
self.rnn_hidden_size = 6
self.rnn_output_mode = "last"
self.output_keep_prob = 1.0
self.share_encoder_weights = True
self.config_dict = {
"mode": "train",
"word_vocab_size": self.data_manager.data_indexer.get_vocab_size(),
"word_embedding_dim": self.word_embedding_dim,
"fine_tune_embeddings": False,
"word_embedding_matrix": self.embedding_matrix,
"rnn_hidden_size": self.rnn_hidden_size,
"rnn_output_mode": self.rnn_output_mode,
"output_keep_prob": self.output_keep_prob,
"share_encoder_weights": self.share_encoder_weights
}
self.num_train_steps_per_epoch = int(math.ceil(self.train_size / self.batch_size))
self.num_val_steps = int(math.ceil(self.val_size / self.batch_size))
self.num_test_steps = int(math.ceil(self.test_size / self.batch_size))
def main():
project_dir = os.path.join(os.path.dirname(__file__), os.pardir, os.pardir)
# Parse config arguments
argparser = argparse.ArgumentParser(
description=("Run the Bilateral Multi-Perspective "
"Matching (biMPM) model on the paraphrase "
"identification task."))
argparser.add_argument("mode",
type=str,
choices=["train", "predict"],
help=("One of {train|predict}, to "
"indicate what you want the model to do. "
"If you pick \"predict\", then you must also "
"supply the path to a pretrained model and "
"DataIndexer to load."))
argparser.add_argument("--model_load_dir",
type=str,
help=("The path to a directory with checkpoints to "
"load for evaluation or prediction. The "
"latest checkpoint will be loaded."))
argparser.add_argument("--dataindexer_load_path",
type=str,
help=("The path to the dataindexer fit on the "
"train data, so we can properly index the "
"test data for evaluation or prediction."))
argparser.add_argument("--train_file",
type=str,
default=os.path.join(
project_dir, "data/processed/quora/"
"train_cleaned_train_split.csv"),
help="Path to a file to train on.")
argparser.add_argument(
"--val_file",
type=str,
default=os.path.join(
project_dir, "data/processed/quora/"
"train_cleaned_val_split.csv"),
help="Path to a file to monitor validation acc. on.")
argparser.add_argument("--test_file",
type=str,
default=os.path.join(
project_dir, "data/processed/quora/"
"test_final.csv"))
argparser.add_argument("--batch_size",
type=int,
default=64,
help="Number of instances per batch.")
argparser.add_argument("--num_epochs",
type=int,
default=10,
help=("Number of epochs to perform in "
"training."))
argparser.add_argument("--early_stopping_patience",
type=int,
default=0,
help=("number of epochs with no validation "
"accuracy improvement after which training "
"will be stopped"))
argparser.add_argument("--num_sentence_words",
type=int,
default=100,
help=("The maximum length of a sentence. Longer "
"sentences will be truncated, and shorter "
"ones will be padded."))
argparser.add_argument("--num_word_characters",
type=int,
default=10,
help=("The maximum length of a word. Longer "
"words will be truncated, and shorter "
"ones will be padded."))
argparser.add_argument("--word_embedding_dim",
type=int,
default=300,
help="Dimensionality of the word embedding layer")
argparser.add_argument(
"--pretrained_word_embeddings_file_path",
type=str,
help="Path to a file with pretrained word embeddings.",
default=os.path.join(project_dir, "data/external/",
"glove.6B.300d.txt"))
argparser.add_argument("--char_embedding_dim",
type=int,
default=20,
help="Dimensionality of the char embedding layer")
argparser.add_argument("--fine_tune_embeddings",
action="store_true",
help=("Whether to train the embedding layer "
"(if True), or keep it fixed (False)."))
argparser.add_argument("--char_rnn_hidden_size",
type=int,
default=50,
help=("The output dimension of the character "
"encoder RNN."))
argparser.add_argument("--context_rnn_hidden_size",
type=int,
default=100,
help=("The output dimension of the context "
"encoding RNN."))
argparser.add_argument("--aggregation_rnn_hidden_size",
type=int,
default=100,
help=("The output dimension of the aggregation "
"encoding RNN."))
argparser.add_argument("--dropout_ratio",
type=float,
default=0.1,
help=("The proportion of RNN outputs to "
"drop out."))
argparser.add_argument("--log_period",
type=int,
default=10,
help=("Number of steps between each summary "
"op evaluation."))
argparser.add_argument("--val_period",
type=int,
default=2500,
help=("Number of steps between each evaluation of "
"validation performance."))
argparser.add_argument("--log_dir",
type=str,
default=os.path.join(project_dir, "logs/"),
help=("Directory to save logs to."))
argparser.add_argument("--save_period",
type=int,
default=2500,
help=("Number of steps between each "
"model checkpoint"))
argparser.add_argument("--save_dir",
type=str,
default=os.path.join(project_dir, "models/"),
help=("Directory to save model checkpoints to."))
argparser.add_argument("--run_id",
type=str,
required=True,
help=("Identifying run ID for this run. If "
"predicting, you probably want this "
"to be the same as the train run_id"))
argparser.add_argument("--model_name",
type=str,
required=True,
help=("Identifying model name for this run. If"
"predicting, you probably want this "
"to be the same as the train run_id"))
argparser.add_argument("--reweight_predictions_for_kaggle",
action="store_true",
help=("Only relevant when predicting. Whether to "
"reweight the prediction probabilities to "
"account for class proportion discrepancy "
"between train and test."))
config = argparser.parse_args()
model_name = config.model_name
run_id = config.run_id
mode = config.mode
# Get the data.
batch_size = config.batch_size
if mode == "train":
# Read the train data from a file, and use it to index the
# validation data
data_manager = DataManager(STSInstance)
num_sentence_words = config.num_sentence_words
num_word_characters = config.num_word_characters
get_train_data_gen, train_data_size = data_manager.get_train_data_from_file(
[config.train_file],
max_lengths={
"num_sentence_words": num_sentence_words,
"num_word_characters": num_word_characters
},
mode="word+character")
get_val_data_gen, val_data_size = data_manager.get_validation_data_from_file(
[config.val_file],
max_lengths={
"num_sentence_words": num_sentence_words,
"num_word_characters": num_word_characters
},
mode="word+character")
else:
# Load the fitted DataManager, and use it to index the test data
logger.info("Loading pickled DataManager from {}".format(
config.dataindexer_load_path))
data_manager = pickle.load(open(config.dataindexer_load_path, "rb"))
get_test_data_gen, test_data_size = data_manager.get_test_data_from_file(
[config.test_file], mode="word+character")
vars(config)["word_vocab_size"] = data_manager.data_indexer.get_vocab_size(
)
vars(config)["char_vocab_size"] = data_manager.data_indexer.get_vocab_size(
namespace="characters")
# Log the run parameters.
log_dir = config.log_dir
log_path = os.path.join(log_dir, model_name, run_id.zfill(2))
logger.info("Writing logs to {}".format(log_path))
if not os.path.exists(log_path):
logger.info("log path {} does not exist, "
"creating it".format(log_path))
os.makedirs(log_path)
params_path = os.path.join(log_path, mode + "params.json")
logger.info("Writing params to {}".format(params_path))
with open(params_path, 'w') as params_file:
json.dump(vars(config), params_file, indent=4)
# Get the embeddings.
embedding_manager = EmbeddingManager(data_manager.data_indexer)
word_embedding_matrix = embedding_manager.get_embedding_matrix(
config.word_embedding_dim, config.pretrained_word_embeddings_file_path)
vars(config)["word_embedding_matrix"] = word_embedding_matrix
char_embedding_matrix = embedding_manager.get_embedding_matrix(
config.char_embedding_dim, namespace="characters")
vars(config)["char_embedding_matrix"] = char_embedding_matrix
# Initialize the model.
model = BiMPM(vars(config))
model.build_graph()
if mode == "train":
# Train the model.
num_epochs = config.num_epochs
num_train_steps_per_epoch = int(math.ceil(train_data_size /
batch_size))
num_val_steps = int(math.ceil(val_data_size / batch_size))
log_period = config.log_period
val_period = config.val_period
save_period = config.save_period
save_dir = os.path.join(config.save_dir, model_name,
run_id.zfill(2) + "/")
save_path = os.path.join(save_dir, model_name + "-" + run_id.zfill(2))
logger.info("Checkpoints will be written to {}".format(save_dir))
if not os.path.exists(save_dir):
logger.info("save path {} does not exist, "
"creating it".format(save_dir))
os.makedirs(save_dir)
logger.info("Saving fitted DataManager to {}".format(save_dir))
data_manager_pickle_name = "{}-{}-DataManager.pkl".format(
model_name, run_id.zfill(2))
pickle.dump(
data_manager,
open(os.path.join(save_dir, data_manager_pickle_name), "wb"))
patience = config.early_stopping_patience
model.train(get_train_instance_generator=get_train_data_gen,
get_val_instance_generator=get_val_data_gen,
batch_size=batch_size,
num_train_steps_per_epoch=num_train_steps_per_epoch,
num_epochs=num_epochs,
num_val_steps=num_val_steps,
save_path=save_path,
log_path=log_path,
log_period=log_period,
val_period=val_period,
save_period=save_period,
patience=patience)
else:
# Predict with the model
model_load_dir = config.model_load_dir
num_test_steps = int(math.ceil(test_data_size / batch_size))
# Numpy array of shape (num_test_examples, 2)
raw_predictions = model.predict(
get_test_instance_generator=get_test_data_gen,
model_load_dir=model_load_dir,
batch_size=batch_size,
num_test_steps=num_test_steps)
# Remove the first column, so we're left with just the probabilities
# that a question is a duplicate.
is_duplicate_probabilities = np.delete(raw_predictions, 0, 1)
# The class balance between kaggle train and test seems different.
# This edits prediction probability to account for the discrepancy.
# See: https://www.kaggle.com/c/quora-question-pairs/discussion/31179
if config.reweight_predictions_for_kaggle:
positive_weight = 0.165 / 0.37
negative_weight = (1 - 0.165) / (1 - 0.37)
is_duplicate_probabilities = (
(positive_weight * is_duplicate_probabilities) /
(positive_weight * is_duplicate_probabilities +
negative_weight * (1 - is_duplicate_probabilities)))
# Write the predictions to an output submission file
output_predictions_path = os.path.join(
log_path,
model_name + "-" + run_id.zfill(2) + "-output_predictions.csv")
logger.info(
"Writing predictions to {}".format(output_predictions_path))
is_duplicate_df = pd.DataFrame(is_duplicate_probabilities)
is_duplicate_df.to_csv(output_predictions_path,
index_label="test_id",
header=["is_duplicate"])
def setUp(self):
super(TestDataManagerTest, self).setUp()
self.write_duplicate_questions_train_file()
self.write_duplicate_questions_test_file()
self.data_manager = DataManager(STSInstance)
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(STSInstance)
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_instances(self):
get_test_gen, test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE],
max_instances=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(STSInstance)
def main():
project_dir = os.path.join(os.path.dirname(__file__), os.pardir, os.pardir)
# Parse config arguments
argparser = argparse.ArgumentParser(
description=("Run the Siamese BiLSTM model with an added "
"matching layer for paraphase detection."))
argparser.add_argument("mode", type=str,
choices=["train", "predict"],
help=("One of {train|predict}, to "
"indicate what you want the model to do. "
"If you pick \"predict\", then you must also "
"supply the path to a pretrained model and "
"DataIndexer to load."))
argparser.add_argument("--model_load_dir", type=str,
help=("The path to a directory with checkpoints to "
"load for evaluation or prediction. The "
"latest checkpoint will be loaded."))
argparser.add_argument("--dataindexer_load_path", type=str,
help=("The path to the DataIndexer fit on the "
"train data, so we can properly index the "
"test data for evaluation or prediction."))
argparser.add_argument("--train_file", type=str,
default=os.path.join(project_dir,
"data/processed/quora/"
"train_cleaned_train_split.csv"),
help="Path to a file to train on.")
argparser.add_argument("--val_file", type=str,
default=os.path.join(project_dir,
"data/processed/quora/"
"train_cleaned_val_split.csv"),
help="Path to a file to monitor validation acc. on.")
argparser.add_argument("--test_file", type=str,
default=os.path.join(project_dir,
"data/processed/quora/"
"test_final.csv"))
argparser.add_argument("--batch_size", type=int, default=128,
help="Number of instances per batch.")
argparser.add_argument("--num_epochs", type=int, default=10,
help=("Number of epochs to perform in "
"training."))
argparser.add_argument("--early_stopping_patience", type=int, default=0,
help=("number of epochs with no validation "
"accuracy improvement after which training "
"will be stopped"))
argparser.add_argument("--num_sentence_words", type=int, default=30,
help=("The maximum length of a sentence. Longer "
"sentences will be truncated, and shorter "
"ones will be padded."))
argparser.add_argument("--word_embedding_dim", type=int, default=300,
help="Dimensionality of the word embedding layer")
argparser.add_argument("--pretrained_embeddings_file_path", type=str,
help="Path to a file with pretrained embeddings.",
default=os.path.join(project_dir,
"data/external/",
"glove.6B.300d.txt"))
argparser.add_argument("--fine_tune_embeddings", action="store_true",
help=("Whether to train the embedding layer "
"(if True), or keep it fixed (False)."))
argparser.add_argument("--rnn_hidden_size", type=int, default=256,
help=("The output dimension of the RNN."))
argparser.add_argument("--share_encoder_weights", action="store_true",
help=("Whether to use the same encoder on both "
"input sentences (thus sharing weights), "
"or a different one for each sentence"))
argparser.add_argument("--output_keep_prob", type=float, default=1.0,
help=("The proportion of RNN outputs to keep, "
"where the rest are dropped out."))
argparser.add_argument("--log_period", type=int, default=10,
help=("Number of steps between each summary "
"op evaluation."))
argparser.add_argument("--val_period", type=int, default=250,
help=("Number of steps between each evaluation of "
"validation performance."))
argparser.add_argument("--log_dir", type=str,
default=os.path.join(project_dir,
"logs/"),
help=("Directory to save logs to."))
argparser.add_argument("--save_period", type=int, default=250,
help=("Number of steps between each "
"model checkpoint"))
argparser.add_argument("--save_dir", type=str,
default=os.path.join(project_dir,
"models/"),
help=("Directory to save model checkpoints to."))
argparser.add_argument("--run_id", type=str, required=True,
help=("Identifying run ID for this run. If "
"predicting, you probably want this "
"to be the same as the train run_id"))
argparser.add_argument("--model_name", type=str, required=True,
help=("Identifying model name for this run. If"
"predicting, you probably want this "
"to be the same as the train run_id"))
argparser.add_argument("--reweight_predictions_for_kaggle", action="store_true",
help=("Only relevant when predicting. Whether to "
"reweight the prediction probabilities to "
"account for class proportion discrepancy "
"between train and test."))
config = argparser.parse_args()
model_name = config.model_name
run_id = config.run_id
mode = config.mode
# Get the data.
batch_size = config.batch_size
if mode == "train":
# Read the train data from a file, and use it to index the
# validation data
data_manager = DataManager(STSInstance)
num_sentence_words = config.num_sentence_words
get_train_data_gen, train_data_size = data_manager.get_train_data_from_file(
[config.train_file], max_lengths={"num_sentence_words": num_sentence_words})
get_val_data_gen, val_data_size = data_manager.get_validation_data_from_file(
[config.val_file], max_lengths={"num_sentence_words": num_sentence_words})
else:
# Load the fitted DataManager, and use it to index the test data
logger.info("Loading pickled DataManager from {}".format(
config.dataindexer_load_path))
data_manager = pickle.load(open(config.dataindexer_load_path, "rb"))
get_test_data_gen, test_data_size = data_manager.get_test_data_from_file(
[config.test_file])
vars(config)["word_vocab_size"] = data_manager.data_indexer.get_vocab_size()
# Log the run parameters.
log_dir = config.log_dir
log_path = os.path.join(log_dir, model_name, run_id.zfill(2))
logger.info("Writing logs to {}".format(log_path))
if not os.path.exists(log_path):
logger.info("log path {} does not exist, "
"creating it".format(log_path))
os.makedirs(log_path)
params_path = os.path.join(log_path, mode + "params.json")
logger.info("Writing params to {}".format(params_path))
with open(params_path, 'w') as params_file:
json.dump(vars(config), params_file, indent=4)
# Get the embeddings.
embedding_manager = EmbeddingManager(data_manager.data_indexer)
embedding_matrix = embedding_manager.get_embedding_matrix(
config.word_embedding_dim,
config.pretrained_embeddings_file_path)
vars(config)["word_embedding_matrix"] = embedding_matrix
# Initialize the model.
model = SiameseMatchingBiLSTM(vars(config))
model.build_graph()
if mode == "train":
# Train the model.
num_epochs = config.num_epochs
num_train_steps_per_epoch = int(math.ceil(train_data_size / batch_size))
num_val_steps = int(math.ceil(val_data_size / batch_size))
log_period = config.log_period
val_period = config.val_period
save_period = config.save_period
save_dir = os.path.join(config.save_dir, model_name, run_id.zfill(2) + "/")
save_path = os.path.join(save_dir, model_name + "-" + run_id.zfill(2))
logger.info("Checkpoints will be written to {}".format(save_dir))
if not os.path.exists(save_dir):
logger.info("save path {} does not exist, "
"creating it".format(save_dir))
os.makedirs(save_dir)
logger.info("Saving fitted DataManager to {}".format(save_dir))
data_manager_pickle_name = "{}-{}-DataManager.pkl".format(model_name,
run_id.zfill(2))
pickle.dump(data_manager,
open(os.path.join(save_dir, data_manager_pickle_name), "wb"))
patience = config.early_stopping_patience
model.train(get_train_instance_generator=get_train_data_gen,
get_val_instance_generator=get_val_data_gen,
batch_size=batch_size,
num_train_steps_per_epoch=num_train_steps_per_epoch,
num_epochs=num_epochs,
num_val_steps=num_val_steps,
save_path=save_path,
log_path=log_path,
log_period=log_period,
val_period=val_period,
save_period=save_period,
patience=patience)
else:
# Predict with the model
model_load_dir = config.model_load_dir
num_test_steps = int(math.ceil(test_data_size / batch_size))
# Numpy array of shape (num_test_examples, 2)
raw_predictions = model.predict(get_test_instance_generator=get_test_data_gen,
model_load_dir=model_load_dir,
batch_size=batch_size,
num_test_steps=num_test_steps)
# Remove the first column, so we're left with just the probabilities
# that a question is a duplicate.
is_duplicate_probabilities = np.delete(raw_predictions, 0, 1)
# The class balance between kaggle train and test seems different.
# This edits prediction probability to account for the discrepancy.
# See: https://www.kaggle.com/c/quora-question-pairs/discussion/31179
if config.reweight_predictions_for_kaggle:
positive_weight = 0.165 / 0.37
negative_weight = (1 - 0.165) / (1 - 0.37)
is_duplicate_probabilities = ((positive_weight * is_duplicate_probabilities) /
(positive_weight * is_duplicate_probabilities +
negative_weight *
(1 - is_duplicate_probabilities)))
# Write the predictions to an output submission file
output_predictions_path = os.path.join(log_path, model_name + "-" +
run_id.zfill(2) +
"-output_predictions.csv")
logger.info("Writing predictions to {}".format(output_predictions_path))
is_duplicate_df = pd.DataFrame(is_duplicate_probabilities)
is_duplicate_df.to_csv(output_predictions_path, index_label="test_id",
header=["is_duplicate"])
class TestBiMPM(DuplicateTestCase):
@overrides
def setUp(self):
super(TestBiMPM, self).setUp()
self.write_duplicate_questions_train_file()
self.write_duplicate_questions_validation_file()
self.write_duplicate_questions_test_file()
self.data_manager = DataManager(STSInstance)
self.batch_size = 3
self.get_train_gen, self.train_size = self.data_manager.get_train_data_from_file(
[self.TRAIN_FILE], mode="word+character")
self.get_val_gen, self.val_size = self.data_manager.get_validation_data_from_file(
[self.VALIDATION_FILE], mode="word+character")
self.get_test_gen, self.test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE], mode="word+character")
self.embedding_manager = EmbeddingManager(
self.data_manager.data_indexer)
self.word_embedding_dim = 5
self.word_embedding_matrix = self.embedding_manager.get_embedding_matrix(
self.word_embedding_dim)
self.char_embedding_dim = 2
self.char_embedding_matrix = self.embedding_manager.get_embedding_matrix(
self.char_embedding_dim)
self.char_rnn_hidden_size = 6
self.context_rnn_hidden_size = 3
self.aggregation_rnn_hidden_size = 4
self.dropout_ratio = 0.1
self.config_dict = {
"mode":
"train",
"word_vocab_size":
self.data_manager.data_indexer.get_vocab_size(),
"word_embedding_dim":
self.word_embedding_dim,
"word_embedding_matrix":
self.word_embedding_matrix,
"char_vocab_size":
self.data_manager.data_indexer.get_vocab_size(
namespace="characters"),
"char_embedding_dim":
self.char_embedding_dim,
"char_embedding_matrix":
self.char_embedding_matrix,
"char_rnn_hidden_size":
self.char_rnn_hidden_size,
"fine_tune_embeddings":
False,
"context_rnn_hidden_size":
self.context_rnn_hidden_size,
"aggregation_rnn_hidden_size":
self.aggregation_rnn_hidden_size,
"dropout_ratio":
self.dropout_ratio
}
self.num_train_steps_per_epoch = int(
math.ceil(self.train_size / self.batch_size))
self.num_val_steps = int(math.ceil(self.val_size / self.batch_size))
self.num_test_steps = int(math.ceil(self.test_size / self.batch_size))
def test_default_does_not_crash(self):
# Initialize the model
model = BiMPM(self.config_dict)
model.build_graph()
# Train the model
model.train(get_train_instance_generator=self.get_train_gen,
get_val_instance_generator=self.get_val_gen,
batch_size=self.batch_size,
num_train_steps_per_epoch=self.num_train_steps_per_epoch,
num_epochs=2,
num_val_steps=self.num_val_steps,
save_path=self.TEST_DIR,
log_path=self.TEST_DIR,
log_period=2,
val_period=2,
save_period=2,
patience=0)
tf.reset_default_graph()
# Load and predict with the model
self.config_dict["mode"] = "test"
del self.config_dict["word_embedding_matrix"]
del self.config_dict["char_embedding_matrix"]
loaded_model = BiMPM(self.config_dict)
loaded_model.build_graph()
loaded_model.predict(get_test_instance_generator=self.get_test_gen,
model_load_dir=self.TEST_DIR,
batch_size=self.batch_size,
num_test_steps=self.num_test_steps)
def main():
default_run_id = "01"
project_dir = os.path.join(os.path.dirname(__file__), os.pardir, os.pardir)
# Parse config arguments
argparser = configargparse.ArgumentParser(
default_config_files=['../../config/default.yml'],
description=("Run a baseline Siamese BiLSTM model "
"for paraphrase identification."))
argparser.add_argument("mode",
type=str,
choices=["train", "predict"],
help=("One of {train|predict}, to "
"indicate what you want the model to do. "
"If you pick \"predict\", then you must also "
"supply the path to a pretrained model and "
"DataIndexer to load."))
argparser.add_argument("--config_file",
is_config_file_arg=True,
help="The path to a config file.")
argparser.add_argument(
"--data_file_dir",
type=str,
default=os.path.join(project_dir, "data/processed/"),
help="Path of the dir to the (train, val, test).csv files.")
argparser.add_argument("--train_filename",
type=str,
default=os.path.join("train.csv"),
help="Basename of the train file.")
argparser.add_argument("--val_filename",
type=str,
default=os.path.join("train.csv"),
help="Basename of the validation file.")
argparser.add_argument("--test_filename",
type=str,
default=os.path.join("test.csv"),
help="Basename of the test file.")
argparser.add_argument("--batch_size",
type=int,
default=128,
help="Number of instances per batch.")
argparser.add_argument("--num_epochs",
type=int,
default=10,
help=("Number of epochs to perform in "
"training."))
argparser.add_argument("--early_stopping_patience",
type=int,
default=0,
help=("number of epochs with no validation "
"accuracy improvement after which training "
"will be stopped"))
argparser.add_argument("--num_sentence_words",
type=int,
default=30,
help=("The maximum length of a sentence. Longer "
"sentences will be truncated, and shorter "
"ones will be padded."))
argparser.add_argument("--embedding_file_path_template",
type=str,
help="Path to a file with pretrained embeddings.",
default=os.path.join(project_dir,
"data/external/bcb",
"{name}.{dim}d.txt"))
argparser.add_argument("--embedding_file_name",
type=str,
help="Name of the embedding file.")
argparser.add_argument("--word_embedding_dim",
type=int,
default=8,
help="Dimensionality of the word embedding layer")
argparser.add_argument("--fine_tune_embeddings",
action="store_true",
help=("Whether to train the embedding layer "
"(if True), or keep it fixed (False)."))
argparser.add_argument("--rnn_hidden_size",
type=int,
default=256,
help=("The output dimension of the RNN."))
argparser.add_argument("--share_encoder_weights",
action="store_true",
help=("Whether to use the same encoder on both "
"input sentences (thus sharing weights), "
"or a different one for each sentence"))
argparser.add_argument("--rnn_output_mode",
type=str,
default="last",
choices=["mean_pool", "last"],
help=("How to calculate the final sentence "
"representation from the RNN outputs. "
"\"mean_pool\" indicates that the outputs "
"will be averaged (with respect to padding), "
"and \"last\" indicates that the last "
"relevant output will be used as the "
"sentence representation."))
argparser.add_argument("--output_keep_prob",
type=float,
default=1.0,
help=("The proportion of RNN outputs to keep, "
"where the rest are dropped out."))
argparser.add_argument("--log_period",
type=int,
default=10,
help=("Number of steps between each summary "
"op evaluation."))
argparser.add_argument("--val_period",
type=int,
default=250,
help=("Number of steps between each evaluation of "
"validation performance."))
argparser.add_argument("--log_dir",
type=str,
default=os.path.join(project_dir, "logs/"),
help=("Directory to save logs to."))
argparser.add_argument("--save_period",
type=int,
default=250,
help=("Number of steps between each "
"model checkpoint"))
argparser.add_argument("--model_save_root",
type=str,
default=os.path.join(project_dir, "models/"),
help=("Directory to save model checkpoints to."))
argparser.add_argument("--token_file_dir",
type=str,
help=("Directory to token files."))
argparser.add_argument("--token_file_ext",
type=str,
help=("File extentions of token files."))
argparser.add_argument("--run_id",
type=str,
default=default_run_id,
help=("Identifying run ID for this run. If "
"predicting, you probably want this "
"to be the same as the train run_id"))
argparser.add_argument("--model_name",
type=str,
help=("Identifying model name for this run. If"
"predicting, you probably want this "
"to be the same as the train run_id"))
argparser.add_argument("--reweight_predictions_for_kaggle",
action="store_true",
help=("Only relevant when predicting. Whether to "
"reweight the prediction probabilities to "
"account for class proportion discrepancy "
"between train and test."))
config = argparser.parse_args()
# logger.info(config)
model_name = config.model_name
run_id = config.run_id.zfill(2)
mode = config.mode
batch_size = config.batch_size
paths = construct_paths(
model_name, run_id, config.data_file_dir, config.train_filename,
config.val_filename, config.test_filename, config.model_save_root,
config.log_dir, config.embedding_file_path_template,
config.embedding_file_name, config.word_embedding_dim)
model_save_file_path = paths['model_save_file_path']
model_save_dir = paths['model_save_dir']
data_manager_pickle_file_path = paths['data_manager_pickle_file_path']
if mode == "train":
# Read the train data from a file, and use it to index the validation data
# TODO: determine from config
# data_manager = DataManager(STSInstance)
CodeInstance.set_token_file(config.token_file_dir,
config.token_file_ext)
data_manager = DataManager(CodeInstance)
num_sentence_words = config.num_sentence_words
get_train_data_gen, train_data_size = data_manager.get_train_data_from_file(
[paths['train_file_path']],
max_lengths={"num_sentence_words": num_sentence_words})
get_val_data_gen, val_data_size = data_manager.get_validation_data_from_file(
[paths['val_file_path']],
max_lengths={"num_sentence_words": num_sentence_words})
else:
# Load the fitted DataManager, and use it to index the test data
logger.info("Loading pickled DataManager "
"from {}".format(data_manager_pickle_file_path))
data_manager = pickle.load(open(data_manager_pickle_file_path, "rb"))
test_data_gen, test_data_size = data_manager.get_test_data_from_file(
[paths['test_file_path']])
vars(config)["word_vocab_size"] = data_manager.data_indexer.get_vocab_size(
)
# Log the run parameters.
log_path = paths['log_file_path']
logger.info("Writing logs to {}".format(log_path))
if not os.path.exists(log_path):
logger.info("log path {} does not exist, "
"creating it".format(log_path))
os.makedirs(log_path)
params_path = os.path.join(log_path, mode + "params.json")
logger.info("Writing params to {}".format(params_path))
with open(params_path, 'w') as params_file:
json.dump(vars(config), params_file, indent=4)
# Get the embeddings.
embedding_manager = EmbeddingManager(data_manager.data_indexer)
embedding_matrix = embedding_manager.get_embedding_matrix(
config.word_embedding_dim, paths['embedding_file_path'])
vars(config)["word_embedding_matrix"] = embedding_matrix
# Initialize the model.
model = SiameseBiLSTM(vars(config))
model.build_graph()
if mode == "train":
# Train the model.
num_epochs = config.num_epochs
num_train_steps_per_epoch = int(math.ceil(train_data_size /
batch_size))
num_val_steps = int(math.ceil(val_data_size / batch_size))
log_period = config.log_period
val_period = config.val_period
save_period = config.save_period
logger.info("Checkpoints will be written to {}".format(model_save_dir))
if not os.path.exists(model_save_dir):
logger.info("save path {} does not exist, "
"creating it".format(model_save_dir))
os.makedirs(model_save_dir)
logger.info("Saving fitted DataManager to {}".format(model_save_dir))
pickle.dump(data_manager, open(data_manager_pickle_file_path, "wb"))
patience = config.early_stopping_patience
model.train(get_train_instance_generator=get_train_data_gen,
get_val_instance_generator=get_val_data_gen,
batch_size=batch_size,
num_train_steps_per_epoch=num_train_steps_per_epoch,
num_epochs=num_epochs,
num_val_steps=num_val_steps,
save_path=model_save_file_path,
log_path=log_path,
log_period=log_period,
val_period=val_period,
save_period=save_period,
patience=patience)
else:
# Predict with the model
num_test_steps = int(math.ceil(test_data_size / batch_size))
# Numpy array of shape (num_test_examples, 2)
raw_predictions, encodings = model.predict(
get_test_instance_generator=test_data_gen,
model_load_dir=model_save_dir,
batch_size=batch_size,
num_test_steps=num_test_steps)
# Remove the first column, so we're left with just the probabilities
# that a question is a duplicate.
is_duplicate_probabilities = np.delete(raw_predictions, 0, 1)
# The class balance between kaggle train and test seems different.
# This edits prediction probability to account for the discrepancy.
# See: https://www.kaggle.com/c/quora-question-pairs/discussion/31179
if config.reweight_predictions_for_kaggle:
positive_weight = 0.165 / 0.37
negative_weight = (1 - 0.165) / (1 - 0.37)
is_duplicate_probabilities = (
(positive_weight * is_duplicate_probabilities) /
(positive_weight * is_duplicate_probabilities +
negative_weight * (1 - is_duplicate_probabilities)))
# Write the predictions to an output submission file
predictions_file_path = paths['predictions_file_path']
logger.info("Writing predictions to {}".format(predictions_file_path))
is_duplicate_df = pd.DataFrame(is_duplicate_probabilities)
# is_duplicate_df.to_csv(predictions_file_path, index_label="test_id",
# header=["is_duplicate"])
# is_duplicate_df.to_csv(predictions_file_path, index=False, header=False)
encodings_df = pd.DataFrame(encodings)
pair_info_df = pd.read_csv(paths['test_file_path'], header=None)
# print(pair_info_df.shape, is_duplicate_df.shape, encodings_df.shape)
# result = pd.DataFrame(np.hstack((pair_info_df, is_duplicate_df, encodings_df)))
result = pd.DataFrame(np.hstack((pair_info_df, is_duplicate_df)))
result.to_csv(predictions_file_path, index=False, header=False)
plot_pairs(predictions_file_path)
class TestSiameseBiLSTM(DuplicateTestCase):
@overrides
def setUp(self):
super(TestSiameseBiLSTM, self).setUp()
self.write_duplicate_questions_train_file()
self.write_duplicate_questions_validation_file()
self.write_duplicate_questions_test_file()
self.data_manager = DataManager(STSInstance)
self.batch_size = 2
self.get_train_gen, self.train_size = self.data_manager.get_train_data_from_file(
[self.TRAIN_FILE])
self.get_val_gen, self.val_size = self.data_manager.get_validation_data_from_file(
[self.VALIDATION_FILE])
self.get_test_gen, self.test_size = self.data_manager.get_test_data_from_file(
[self.TEST_FILE])
self.embedding_manager = EmbeddingManager(self.data_manager.data_indexer)
self.word_embedding_dim = 5
self.embedding_matrix = self.embedding_manager.get_embedding_matrix(
self.word_embedding_dim)
self.rnn_hidden_size = 6
self.rnn_output_mode = "last"
self.output_keep_prob = 1.0
self.share_encoder_weights = True
self.config_dict = {
"mode": "train",
"word_vocab_size": self.data_manager.data_indexer.get_vocab_size(),
"word_embedding_dim": self.word_embedding_dim,
"fine_tune_embeddings": False,
"word_embedding_matrix": self.embedding_matrix,
"rnn_hidden_size": self.rnn_hidden_size,
"rnn_output_mode": self.rnn_output_mode,
"output_keep_prob": self.output_keep_prob,
"share_encoder_weights": self.share_encoder_weights
}
self.num_train_steps_per_epoch = int(math.ceil(self.train_size / self.batch_size))
self.num_val_steps = int(math.ceil(self.val_size / self.batch_size))
self.num_test_steps = int(math.ceil(self.test_size / self.batch_size))
def test_default_does_not_crash(self):
# Initialize the model
model = SiameseBiLSTM(self.config_dict)
model.build_graph()
# Train the model
model.train(get_train_instance_generator=self.get_train_gen,
get_val_instance_generator=self.get_val_gen,
batch_size=self.batch_size,
num_train_steps_per_epoch=self.num_train_steps_per_epoch,
num_epochs=2,
num_val_steps=self.num_val_steps,
save_path=self.TEST_DIR,
log_path=self.TEST_DIR,
log_period=2,
val_period=2,
save_period=2,
patience=0)
tf.reset_default_graph()
# Load and predict with the model
self.config_dict["mode"] = "test"
del self.config_dict["word_embedding_matrix"]
loaded_model = SiameseBiLSTM(self.config_dict)
loaded_model.build_graph()
loaded_model.predict(get_test_instance_generator=self.get_test_gen,
model_load_dir=self.TEST_DIR,
batch_size=self.batch_size,
num_test_steps=self.num_test_steps)
def test_mean_pool_does_not_crash(self):
# Initialize the model
self.config_dict["rnn_output_mode"] = "mean_pool"
model = SiameseBiLSTM(self.config_dict)
model.build_graph()
# Train the model
model.train(get_train_instance_generator=self.get_train_gen,
get_val_instance_generator=self.get_val_gen,
batch_size=self.batch_size,
num_train_steps_per_epoch=self.num_train_steps_per_epoch,
num_epochs=2,
num_val_steps=self.num_val_steps,
save_path=self.TEST_DIR,
log_path=self.TEST_DIR,
log_period=2,
val_period=2,
save_period=2,
patience=0)
tf.reset_default_graph()
# Load and predict with the model
self.config_dict["mode"] = "test"
del self.config_dict["word_embedding_matrix"]
loaded_model = SiameseBiLSTM(self.config_dict)
loaded_model.build_graph()
loaded_model.predict(get_test_instance_generator=self.get_test_gen,
model_load_dir=self.TEST_DIR,
batch_size=self.batch_size,
num_test_steps=self.num_test_steps)
def test_non_sharing_encoders_does_not_crash(self):
# Initialize the model
self.config_dict["share_encoder_weights"] = False
model = SiameseBiLSTM(self.config_dict)
model.build_graph()
# Train the model
model.train(get_train_instance_generator=self.get_train_gen,
get_val_instance_generator=self.get_val_gen,
batch_size=self.batch_size,
num_train_steps_per_epoch=self.num_train_steps_per_epoch,
num_epochs=2,
num_val_steps=self.num_val_steps,
save_path=self.TEST_DIR,
log_path=self.TEST_DIR,
log_period=2,
val_period=2,
save_period=2,
patience=0)
tf.reset_default_graph()
# Load and predict with the model
self.config_dict["mode"] = "test"
del self.config_dict["word_embedding_matrix"]
loaded_model = SiameseBiLSTM(self.config_dict)
loaded_model.build_graph()
loaded_model.predict(get_test_instance_generator=self.get_test_gen,
model_load_dir=self.TEST_DIR,
batch_size=self.batch_size,
num_test_steps=self.num_test_steps)
def main():
project_dir = os.path.join(os.path.dirname(__file__), os.pardir, os.pardir)
# Parse config arguments
argparser = argparse.ArgumentParser(
description=("Run the Siamese BiLSTM model with an added "
"matching layer for paraphase detection."))
argparser.add_argument("mode",
type=str,
choices=["train", "predict"],
help=("One of {train|predict}, to "
"indicate what you want the model to do. "
"If you pick \"predict\", then you must also "
"supply the path to a pretrained model and "
"DataIndexer to load."))
argparser.add_argument("--model_load_dir",
type=str,
default=os.path.join(project_dir, "models/"),
help=("The path to a directory with checkpoints to "
"load for evaluation or prediction. The "
"latest checkpoint will be loaded."))
argparser.add_argument("--dataindexer_load_path",
type=str,
default=os.path.join(project_dir, "models/"),
help=("The path to the DataIndexer fit on the "
"train data, so we can properly index the "
"test data for evaluation or prediction."))
argparser.add_argument("--train_file",
type=str,
default=os.path.join(
project_dir, "data/processed/weizhong/"
"weizhong_train_train_split.txt"),
help="Path to a file to train on.")
argparser.add_argument("--val_file",
type=str,
default=os.path.join(
project_dir, "data/processed/weizhong/"
"weizhong_train_val_split.txt"))
argparser.add_argument("--test_file",
type=str,
default=os.path.join(project_dir, "data/raw/"
"weizhong_dev.txt"))
argparser.add_argument("--batch_size",
type=int,
default=512,
help="Number of instances per batch.")
argparser.add_argument("--num_epochs",
type=int,
default=200,
help=("Number of epochs to perform in "
"training."))
argparser.add_argument("--early_stopping_patience",
type=int,
default=100,
help=("number of epochs with no validation "
"accuracy improvement after which training "
"will be stopped"))
argparser.add_argument("--num_sentence_words",
type=int,
default=25,
help=("The maximum length of a sentence. Longer "
"sentences will be truncated, and shorter "
"ones will be padded."))
argparser.add_argument("--word_embedding_dim",
type=int,
default=300,
help="Dimensionality of the word embedding layer")
argparser.add_argument("--pretrained_embeddings_file_path",
type=str,
help="Path to a file with pretrained embeddings.",
default=os.path.join(
project_dir, "data/external/",
"word2vec_finance_train_use.txt"))
argparser.add_argument("--fine_tune_embeddings",
action="store_true",
help=("Whether to train the embedding layer "
"(if True), or keep it fixed (False)."))
argparser.add_argument("--rnn_hidden_size",
type=int,
default=256,
help=("The output dimension of the RNN."))
argparser.add_argument("--share_encoder_weights",
action="store_true",
help=("Whether to use the same encoder on both "
"input sentences (thus sharing weights), "
"or a different one for each sentence"))
argparser.add_argument("--output_keep_prob",
type=float,
default=0.9,
help=("The proportion of RNN outputs to keep, "
"where the rest are dropped out."))
argparser.add_argument("--log_period",
type=int,
default=10,
help=("Number of steps between each summary "
"op evaluation."))
argparser.add_argument("--val_period",
type=int,
default=250,
help=("Number of steps between each evaluation of "
"validation performance."))
argparser.add_argument("--log_dir",
type=str,
default=os.path.join(project_dir, "logs/"),
help=("Directory to save logs to."))
argparser.add_argument("--save_period",
type=int,
default=250,
help=("Number of steps between each "
"model checkpoint"))
argparser.add_argument("--save_dir",
type=str,
default=os.path.join(project_dir, "models/"),
help=("Directory to save model checkpoints to."))
argparser.add_argument("--run_id",
type=str,
required=True,
help=("Identifying run ID for this run. If "
"predicting, you probably want this "
"to be the same as the train run_id"))
argparser.add_argument("--model_name",
type=str,
required=True,
help=("Identifying model name for this run. If"
"predicting, you probably want this "
"to be the same as the train run_id"))
argparser.add_argument("--reweight_predictions_for_kaggle",
action="store_true",
help=("Only relevant when predicting. Whether to "
"reweight the prediction probabilities to "
"account for class proportion discrepancy "
"between train and test."))
config = argparser.parse_args()
model_name = config.model_name
run_id = config.run_id
mode = config.mode
print("======================", config.share_encoder_weights)
# Get the data.
batch_size = config.batch_size
if mode == "train":
# Read the train data from a file, and use it to index the
# validation data
data_manager = DataManager(WeizhongInstance)
num_sentence_words = config.num_sentence_words
get_train_data_gen, train_data_size = data_manager.get_train_data_from_file(
[config.train_file],
max_lengths={"num_sentence_words": num_sentence_words})
get_val_data_gen, val_data_size = data_manager.get_validation_data_from_file(
[config.val_file],
max_lengths={"num_sentence_words": num_sentence_words})
else:
# Load the fitted DataManager, and use it to index the test data
logger.info("Loading pickled DataManager from {}".format(
config.dataindexer_load_path))
config.dataindexer_load_path = os.path.join(
config.dataindexer_load_path, config.model_name,
config.run_id.zfill(2), config.model_name + "-" +
config.run_id.zfill(2) + "-" + "DataManager.pkl")
data_manager = pickle.load(open(config.dataindexer_load_path, "rb"))
get_test_data_gen, test_data_size = data_manager.get_test_data_from_file(
[config.test_file])
vars(config)["word_vocab_size"] = data_manager.data_indexer.get_vocab_size(
)
# Log the run parameters.
log_dir = config.log_dir
log_path = os.path.join(log_dir, model_name, run_id.zfill(2))
logger.info("Writing logs to {}".format(log_path))
if not os.path.exists(log_path):
logger.info("log path {} does not exist, "
"creating it".format(log_path))
os.makedirs(log_path)
params_path = os.path.join(log_path, mode + "params.json")
logger.info("Writing params to {}".format(params_path))
with open(params_path, 'w') as params_file:
json.dump(vars(config), params_file, indent=4)
# Get the embeddings.
embedding_manager = EmbeddingManager(data_manager.data_indexer)
embedding_matrix = embedding_manager.get_embedding_matrix(
config.word_embedding_dim, config.pretrained_embeddings_file_path)
vars(config)["word_embedding_matrix"] = embedding_matrix
# Initialize the model.
model = SiameseMatchingBiLSTM(vars(config))
model.build_graph()
if mode == "train":
# Train the model.
num_epochs = config.num_epochs
num_train_steps_per_epoch = int(math.ceil(train_data_size /
batch_size))
num_val_steps = int(math.ceil(val_data_size / batch_size))
log_period = config.log_period
val_period = config.val_period
save_period = config.save_period
save_dir = os.path.join(config.save_dir, model_name,
run_id.zfill(2) + "/")
save_path = os.path.join(save_dir, model_name + "-" + run_id.zfill(2))
logger.info("Checkpoints will be written to {}".format(save_dir))
if not os.path.exists(save_dir):
logger.info("save path {} does not exist, "
"creating it".format(save_dir))
os.makedirs(save_dir)
logger.info("Saving fitted DataManager to {}".format(save_dir))
data_manager_pickle_name = "{}-{}-DataManager.pkl".format(
model_name, run_id.zfill(2))
pickle.dump(
data_manager,
open(os.path.join(save_dir, data_manager_pickle_name), "wb"))
patience = config.early_stopping_patience
model.train(get_train_instance_generator=get_train_data_gen,
get_val_instance_generator=get_val_data_gen,
batch_size=batch_size,
num_train_steps_per_epoch=num_train_steps_per_epoch,
num_epochs=num_epochs,
num_val_steps=num_val_steps,
save_path=save_path,
log_path=log_path,
log_period=log_period,
val_period=val_period,
save_period=save_period,
patience=patience)
else:
# Predict with the model
model_load_dir = os.path.join(config.model_load_dir, config.model_name,
config.run_id.zfill(2))
num_test_steps = int(math.ceil(test_data_size / batch_size))
# Numpy array of shape (num_test_examples, 2)
raw_predictions, lineids = model.predict(
get_test_instance_generator=get_test_data_gen,
model_load_dir=model_load_dir,
batch_size=batch_size,
num_test_steps=num_test_steps)
# Write the predictions to an output submission file
output_predictions_path = os.path.join(
log_path,
model_name + "-" + run_id.zfill(2) + "-output_predictions.csv")
with open(output_predictions_path, "w") as output_file:
output_file.write("test_id,result\n")
for index, lineid in enumerate(lineids):
output_file.write(
str(lineid) + "," + str(raw_predictions[index]) + "\n")
logger.info(
"Writing predictions to {}".format(output_predictions_path))