def load_or_create_vocab_and_embs(arguments):
vocab = Vocabulary(arguments.only_words)
# load or create vocabulary
try:
vocab.load(arguments.vocab_file)
except:
if arguments.embs == None:
vocab = vocab.fit(arguments.train)
else:
vocab = vocab.fit(arguments.train, arguments.embs)
# save vocab for reproducability later
logging.info("> saving vocab to %s" % (arguments.vocab_file))
vocab.save(arguments.vocab_file)
if arguments.embs == None:
embs = None
else:
embs = vocab.load_embedding()
logging.info('shape %s' % (embs.shape))
return vocab, embs
def load_or_create_vocab(arguments):
vocab = Vocabulary(arguments.only_words)
# load or create vocabulary
try:
vocab.load(arguments.vocab_file)
except:
vocab = vocab.fit(arguments.train)
# save vocab for reproducability later
logging.info("> saving vocab to %s" % (arguments.vocab_file))
vocab.save(arguments.vocab_file)
return vocab
parser = argparse.ArgumentParser()
parser.add_argument("--train", required=True)
parser.add_argument("--dev", required=True)
parser.add_argument("--test", required=True)
parser.add_argument("--model", required=True)
arguments, unknown = parser.parse_known_args()
TRAIN_FILE = arguments.train
DEV_FILE = arguments.dev
TEST_FILE = arguments.test
MODEL_FILE = arguments.model
n_epochs = 5
vocab = Vocabulary()
vocab.fit(TRAIN_FILE)
print(">> Loading in data")
TRAIN = vocab.tokenize_conll(arguments.train)
DEV = vocab.tokenize_conll(arguments.dev)
TEST = vocab.tokenize_conll(arguments.test)
encoder = BetaEncodeHandler()
print("> pre-encoding edges")
s = time.time()
TRAIN = pre_encode(encoder, TRAIN, accumulate_vocab=True)
DEV = pre_encode(encoder, DEV)
TEST = pre_encode(encoder, TEST)
print(">> done pre-encoding", time.time() - s)
# 5m is completely arbitrary
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--train",
dest="train",
help="Annotated CONLL train file",
metavar="FILE",
required=True,
)
parser.add_argument(
"--dev",
dest="dev",
help="Annotated CONLL dev file",
metavar="FILE",
required=True,
)
parser.add_argument(
"--test",
dest="test",
help="Annotated CONLL dev test",
metavar="FILE",
required=True,
)
parser.add_argument("--epochs", dest="epochs", type=int, default=30)
parser.add_argument("--tb_dest", dest="tb_dest")
parser.add_argument("--vocab_dest", dest="vocab_dest")
parser.add_argument("--model_dest", dest="model_dest", required=True)
parser.add_argument(
"--embs", dest="embs", help="pre-trained embeddings file name", required=False
)
parser.add_argument(
"--no_update_pretrained_emb",
dest="no_update_pretrained_emb",
help="don't update the pretrained embeddings during training",
default=False,
action="store_true",
)
parser.add_argument("--patience", dest="patience", type=int, default=-1)
arguments, unknown = parser.parse_known_args()
n_epochs = arguments.epochs
vocab = Vocabulary()
if arguments.embs:
vocab = vocab.fit(arguments.train, arguments.embs)
embs = vocab.load_embedding()
print("shape", embs.shape)
else:
vocab = vocab.fit(arguments.train)
embs = None
# save vocab for reproducability later
if arguments.vocab_dest:
print("> saving vocab to", arguments.vocab_dest)
vocab.save(arguments.vocab_dest)
# prep data
print(">> Loading in data")
training_data = vocab.tokenize_conll(arguments.train)
dev_data = vocab.tokenize_conll(arguments.dev)
test_data = vocab.tokenize_conll(arguments.test)
# instantiate model
model = DependencyParser(vocab, embs)
callbacks = []
tensorboard_logger = None
if arguments.tb_dest:
tensorboard_logger = TensorboardLoggerCallback(arguments.tb_dest)
callbacks.append(tensorboard_logger)
save_callback = ModelSaveCallback(arguments.model_dest)
callbacks.append(save_callback)
# prep params
parser = Model(
model,
decoder="eisner",
loss="kiperwasser",
optimizer="adam",
strategy="bucket",
vocab=vocab,
)
parser.train(
training_data,
arguments.dev,
dev_data,
epochs=n_epochs,
batch_size=32,
callbacks=callbacks,
patience=arguments.patience,
)
parser.load_from_file(arguments.model_dest)
metrics = parser.evaluate(arguments.test, test_data, batch_size=32)
test_UAS = metrics["nopunct_uas"]
test_LAS = metrics["nopunct_las"]
print(metrics)
if arguments.tb_dest and tensorboard_logger:
tensorboard_logger.raw_write("test_UAS", test_UAS)
tensorboard_logger.raw_write("test_LAS", test_LAS)
print()
print(">>> Model maxed on dev at epoch", save_callback.best_epoch)
print(">>> Test score:", test_UAS, test_LAS)
parser.add_argument("--epochs", dest="epochs", type=int, default=30)
parser.add_argument("--tb_dest", dest="tb_dest")
parser.add_argument("--vocab_dest", dest="vocab_dest")
parser.add_argument("--model_dest", dest="model_dest", required=True)
parser.add_argument("--embs", dest="embs", help="pre-trained embeddings file name", required=False)
parser.add_argument("--no_update_pretrained_emb", dest="no_update_pretrained_emb", help="don't update the pretrained embeddings during training", default=False, action='store_true')
parser.add_argument("--patience", dest='patience', type=int, default=-1)
parser.add_argument("--dev_mode", dest='dev_mode', default=False, help='small subset of training examples, for code testing', action='store_true')
arguments, unknown = parser.parse_known_args()
n_epochs = arguments.epochs
vocab = Vocabulary()
if arguments.embs == None:
vocab = vocab.fit(arguments.train)
embs = None
else:
vocab = vocab.fit(arguments.train, arguments.embs)
embs = vocab.load_embedding()
print('shape',embs.shape)
# save vocab for reproducability later
if arguments.vocab_dest:
print("> saving vocab to", arguments.vocab_dest)
vocab.save(arguments.vocab_dest)
# prep data
print(">> Loading in data")
training_data = vocab.tokenize_conll(arguments.train)
if arguments.dev_mode:
ARGPARSER = argparse.ArgumentParser()
ARGPARSER.add_argument("--train", required=True)
ARGPARSER.add_argument("--dev", required=True)
ARGPARSER.add_argument("--test", required=True)
ARGPARSER.add_argument("--model", required=True)
ARGUMENTS, UNK = ARGPARSER.parse_known_args()
TRAIN_FILE = ARGUMENTS.train
DEV_FILE = ARGUMENTS.dev
TEST_FILE = ARGUMENTS.test
MODEL_FILE = ARGUMENTS.model
N_EPOCHS = 5
VOCAB = Vocabulary()
VOCAB.fit(TRAIN_FILE)
print("> Loading in data")
TRAIN = VOCAB.tokenize_conll(ARGUMENTS.train)
DEV = VOCAB.tokenize_conll(ARGUMENTS.dev)
TEST = VOCAB.tokenize_conll(ARGUMENTS.test)
ENCODER = BetaEncodeHandler()
print("> Pre-encoding edges")
START_TIME = time.time()
TRAIN = pre_encode(ENCODER, TRAIN, accumulate_vocab=True)
DEV = pre_encode(ENCODER, DEV)
TEST = pre_encode(ENCODER, TEST)
print(">> Done pre-encoding edges", time.time() - START_TIME)
# 5m is completely arbitrary but fits all features for PTB.
def main():
"""Main function."""
argparser = argparse.ArgumentParser()
argparser.add_argument("--train", required=True)
argparser.add_argument("--dev", required=True)
argparser.add_argument("--test", required=True)
argparser.add_argument("--emb", dest="emb")
argparser.add_argument("--epochs", dest="epochs", type=int, default=283)
argparser.add_argument("--vocab_dest", dest="vocab_dest", required=True)
argparser.add_argument("--model_dest", dest="model_dest", required=True)
argparser.add_argument("--lstm_layers", dest="lstm_layers", type=int, default=3)
argparser.add_argument("--dropout", type=int, default=0.33)
arguments, _ = argparser.parse_known_args()
# [Data]
min_occur_count = 2
train_file = arguments.train
dev_file = arguments.dev
vocab_destination = arguments.vocab_dest
model_destination = arguments.model_dest
# [Network]
word_dims = 100
tag_dims = 100
lstm_hiddens = 400
mlp_arc_size = 500
mlp_rel_size = 100
lstm_layers = arguments.lstm_layers
dropout_emb = arguments.dropout
dropout_lstm_input = arguments.dropout
dropout_lstm_hidden = arguments.dropout
dropout_mlp = arguments.dropout
# [Hyperparameters for optimizer]
learning_rate = 2e-3
decay = 0.75
decay_steps = 5000
beta_1 = 0.9
beta_2 = 0.9
epsilon = 1e-12
# [Run]
batch_scale = 5000 # for scaled batching
n_epochs = arguments.epochs
vocab = Vocabulary()
vocab = vocab.fit(train_file, arguments.emb, min_occur_count)
embs = vocab.load_embedding(True) if arguments.emb else None
vocab.save(vocab_destination)
model = DozatManning(
vocab,
word_dims,
tag_dims,
dropout_emb,
lstm_layers,
lstm_hiddens,
dropout_lstm_input,
dropout_lstm_hidden,
mlp_arc_size,
mlp_rel_size,
dropout_mlp,
pretrained_embeddings=embs,
)
optimizer = dy.AdamTrainer(
model.parameter_collection, learning_rate, beta_1, beta_2, epsilon
)
# Callbacks
custom_learning_update_callback = UpdateParamsCallback(
optimizer, learning_rate, decay, decay_steps
)
save_callback = ModelSaveCallback(model_destination)
callbacks = [custom_learning_update_callback, save_callback]
parser = Model(
model,
decoder="cle",
loss="crossentropy",
optimizer=optimizer,
strategy="scaled_batch",
vocab=vocab,
)
# Prep data
training_data = vocab.tokenize_conll(arguments.train)
dev_data = vocab.tokenize_conll(arguments.dev)
test_data = vocab.tokenize_conll(arguments.test)
parser.train(
training_data,
dev_file,
dev_data,
epochs=n_epochs,
batch_size=batch_scale,
callbacks=callbacks,
)
parser.load_from_file(model_destination)
metrics = parser.evaluate(arguments.test, test_data, batch_size=batch_scale)
test_uas = metrics["nopunct_uas"]
test_las = metrics["nopunct_las"]
print()
print(metrics)
print(">> Test score:", test_uas, test_las)
if arguments.lstm_layers != 3:
print(
">> WARNING: running with more or less bilstm layers than origin (%d)"
% arguments.lstm_layers)
if arguments.embedding_file is None:
print(">> WARNING: Running without pretrained embeddings.")
if arguments.no_orth_init:
print(">> Warning: running without orthogonal initilization on parameters")
if arguments.dropout < 0.33:
print(">> Warning: running model with less dropout")
vocab = Vocabulary()
vocab = vocab.fit(train_file, pretrained_embeddings_file, min_occur_count)
embs = vocab.load_embedding(
variance_normalize=True) if arguments.embedding_file else None
# save vocab for reproducing later
if vocab_destination:
vocab.save(vocab_destination)
print("> saving vocab to", vocab_destination)
""" """
model = BaseParser(vocab, word_dims, tag_dims, dropout_emb, lstm_layers,
lstm_hiddens, dropout_lstm_input, dropout_lstm_hidden,
mlp_arc_size, mlp_rel_size, dropout_mlp, embs, orth_init)
""" Instantiate custom optimizer """
optimizer = dy.AdamTrainer(model.parameter_collection, learning_rate, beta_1,
beta_2, epsilon)
""" Callbacks """