def train_tagger(args):
config = configuration.get_config(args.config)
numpy.random.seed(666)
torch.manual_seed(666)
torch.set_printoptions(precision=20)
### gpu
gpu = torch.cuda.is_available()
if args.gpu and gpu:
print("GPU available: {}\t GPU ID: {}".format(gpu, args.gpu))
torch.cuda.manual_seed(666)
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
i = 0
global_step = 0
epoch = 0
train_loss = 0.0
with Timer('Data loading'):
vocab_path = args.vocab if args.vocab != '' else None
label_path = args.labels if args.labels != '' else None
gold_props_path = args.gold if args.gold != '' else None
print('Task: {}'.format(args.task))
if args.task == 'srl':
# Data and evaluator for SRL.
data = TaggerData(
config,
*reader.get_srl_data(config, args.train, args.dev, vocab_path,
label_path))
evaluator = SRLEvaluator(data.get_development_data(),
data.label_dict,
gold_props_file=gold_props_path,
use_se_marker=config.use_se_marker,
pred_props_file=None,
word_dict=data.word_dict)
else:
print "Not implemented yet!"
exit()
# Data and evaluator for PropId.
data = TaggerData(
config,
*reader.get_postag_data(config, args.train, args.dev,
vocab_path, label_path))
evaluator = PropIdEvaluator(data.get_development_data(),
data.label_dict)
batched_dev_data = data.get_development_data(
batch_size=config.dev_batch_size)
print('Dev data has {} batches.'.format(len(batched_dev_data)))
with Timer('Syntactic Information Extracting'
): # extract the syntactic information from file
train_dep_trees = SyntacticCONLL()
dev_dep_trees = SyntacticCONLL()
train_dep_trees.read_from_file(args.train_dep_trees)
dev_dep_trees.read_from_file(args.dev_dep_trees)
# generate the syntactic label dict in training corpus
data.syntactic_dict = train_dep_trees.get_syntactic_label_dict()
data.syntactic_dict.accept_new = False
dev_dep_trees.get_syntactic_label_dict(data.syntactic_dict)
with Timer('Preparation'):
if not os.path.isdir(args.model):
print('Directory {} does not exist. Creating new.'.format(
args.model))
os.makedirs(args.model)
else:
if len(os.listdir(args.model)) > 0:
print ('[WARNING] Log directory {} is not empty, previous checkpoints might be overwritten' \
.format(args.model))
shutil.copyfile(args.config, os.path.join(args.model, 'config'))
# Save word and label dict to model directory.
data.word_dict.save(os.path.join(args.model, 'word_dict'))
data.label_dict.save(os.path.join(args.model, 'label_dict'))
data.syntactic_dict.save(os.path.join(args.model, 'syn_label_dict'))
writer = open(os.path.join(args.model, 'checkpoints.tsv'), 'w')
writer.write(
'step\tdatetime\tdev_loss\tdev_accuracy\tbest_dev_accuracy\n')
with Timer('Building model'):
model = BiLSTMTaggerModel(data, config=config, gpu_id=args.gpu)
if args.gpu:
print "BiLSTMTaggerModel initialize with Cuda!"
model = model.cuda()
if args.gpu != "" and not torch.cuda.is_available():
raise Exception("No GPU Found!")
exit()
for param in model.parameters():
print param.size()
optimizer = torch.optim.Adadelta(
model.parameters(), lr=1.0,
rho=0.95) # initialize the optimizer outside the epoch
batch_position_encoding = position_encoding_init(
200, 100) # 0: root, 1, ...n, n+1: padding | max length 200
while epoch < config.max_epochs:
with Timer("Epoch%d" % epoch) as timer:
model.train()
train_data = data.get_training_data(include_last_batch=True)
for batched_tensor in train_data: # for each batch in the training corpus
x, y, lengths, weights = batched_tensor
word_inputs_seqs, predicate_inputs_seqs, syn_label_inputs_seqs, pes, _, input_lengths, masks, padding_answers = \
batch_data_variable(train_dep_trees, batch_position_encoding, x, y, lengths, weights)
if args.gpu:
word_inputs_seqs, predicate_inputs_seqs, syn_label_inputs_seqs, input_lengths, masks, \
padding_answers = \
word_inputs_seqs.cuda(), predicate_inputs_seqs.cuda(), syn_label_inputs_seqs.cuda(), \
input_lengths.cuda(), masks.cuda(), padding_answers.cuda()
optimizer.zero_grad()
output = model.forward(word_inputs_seqs, predicate_inputs_seqs,
syn_label_inputs_seqs, pes,
input_lengths)
loss = model.compute_loss(output, padding_answers, masks)
loss.backward()
# gradient clipping
torch.nn.utils.clip_grad_norm(model.parameters(), 1.0)
optimizer.step()
train_loss += loss.data # should be tensor not Variable, avoiding the graph accumulates
i += 1
global_step += 1
if i % 400 == 0:
timer.tick("{} training steps, loss={:.3f}".format(
i, float(train_loss / i)))
train_loss = train_loss / i
print("Epoch {}, steps={}, loss={:.3f}".format(
epoch, i, float(train_loss)))
i = 0
epoch += 1
train_loss = 0.0
if epoch % config.checkpoint_every_x_epochs == 0:
with Timer('Evaluation'):
evaluate_tagger(model, batch_position_encoding,
batched_dev_data, dev_dep_trees, evaluator,
writer, global_step)
# Done. :)
writer.close()
def get_scores(config, task, model_path, word_dict_path, label_dict_path, syntactic_dict_path, input_path):
with Timer('Data loading'):
print ('Task: {}'.format(task))
allow_new_words = True
print ('Allow new words in test data: {}'.format(allow_new_words))
# Load word and tag dictionary
word_dict = Dictionary(padding_token=PADDING_TOKEN, unknown_token=UNKNOWN_TOKEN) # word tokens to Dict
label_dict, syntactic_dict = Dictionary(), Dictionary()
word_dict.load(word_dict_path)
label_dict.load(label_dict_path)
syntactic_dict.load(syntactic_dict_path)
data = TaggerData(config, [], [], word_dict, label_dict, None, None)
data.syntactic_dict = syntactic_dict
# Load test data.
if task == 'srl':
test_sentences, emb_inits, emb_shapes = reader.get_srl_test_data(
input_path,
config,
data.word_dict,
data.label_dict,
allow_new_words)
else:
test_sentences, emb_inits, emb_shapes = reader.get_postag_test_data(
input_path,
config,
data.word_dict,
data.label_dict,
allow_new_words)
print ('Read {} sentences.'.format(len(test_sentences)))
# Add pre-trained embeddings for new words in the test data.
# if allow_new_words:
data.embedding_shapes = emb_shapes
data.embeddings = emb_inits
# Batching.
test_data = data.get_test_data(test_sentences, batch_size=config.dev_batch_size)
with Timer('Syntactic Information Extracting'): # extract the syntactic information from file
test_dep_trees = SyntacticCONLL()
test_dep_trees.read_from_file(args.input_dep_trees)
# generate the syntactic label dict in training corpus
data.syntactic_dict.accept_new = False
test_dep_trees.get_syntactic_label_dict(data.syntactic_dict)
with Timer('Model building and loading'):
model = BiLSTMTaggerModel(data, config=config, gpu_id=args.gpu)
model.load(model_path)
for param in model.parameters():
print param.size()
if args.gpu:
print("Initialize the model with GPU!")
model = model.cuda()
with Timer('Running model'):
scores = []
model.eval()
for i, batched_tensor in enumerate(test_data):
x, y, lengths, weights = batched_tensor
word_inputs_seqs, predicate_inputs_seqs, syn_label_inputs_seqs, pes, answers, input_lengths, masks, padding_answers = \
batch_data_variable(test_dep_trees, None, x, y, lengths, weights)
if args.gpu:
word_inputs_seqs, predicate_inputs_seqs, syn_label_inputs_seqs, input_lengths, masks, \
padding_answers = \
word_inputs_seqs.cuda(), predicate_inputs_seqs.cuda(), syn_label_inputs_seqs.cuda(), \
input_lengths.cuda(), masks.cuda(), padding_answers.cuda()
sc = model.forward(word_inputs_seqs, predicate_inputs_seqs, syn_label_inputs_seqs, pes, input_lengths)
sc = sc.data.cpu().numpy() if args.gpu else sc.data.numpy()
sc = [sc[j] for j in range(sc.shape[0])]
scores.extend(sc)
return scores, data, test_sentences, test_data
def get_scores(config, task, model_path, word_dict_path, label_dict_path, syn_label_dict_path, pattern_dict_path, input_path):
with Timer('Data loading'):
print ('Task: {}'.format(task))
allow_new_words = True
print ('Allow new words in test data: {}'.format(allow_new_words))
# Load word and tag dictionary
word_dict = Dictionary(padding_token=PADDING_TOKEN, unknown_token=UNKNOWN_TOKEN) # word tokens to Dict
label_dict = Dictionary()
syn_label_dict = Dictionary()
pattern_dict = Dictionary()
word_dict.load(word_dict_path)
label_dict.load(label_dict_path)
syn_label_dict.load(syn_label_dict_path)
pattern_dict.load(pattern_dict_path)
data = TaggerData(config, [], [], word_dict, label_dict, None, None)
data.syntactic_dict = syn_label_dict
data.pattern_dict = pattern_dict
# Load test data.
if task == 'srl':
test_sentences, emb_inits, emb_shapes = reader.get_srl_test_data(
input_path,
config,
data.word_dict,
data.label_dict,
allow_new_words)
else:
test_sentences, emb_inits, emb_shapes = reader.get_postag_test_data(
input_path,
config,
data.word_dict,
data.label_dict,
allow_new_words)
print ('Read {} sentences.'.format(len(test_sentences)))
# Add pre-trained embeddings for new words in the test data.
# if allow_new_words:
data.embedding_shapes = emb_shapes
data.embeddings = emb_inits
# Batching.
test_data = data.get_test_data(test_sentences, batch_size=config.dev_batch_size)
with Timer('Syntactic Information Extracting'): # extract the syntactic information from file
test_dep_trees = SyntacticCONLL()
test_dep_trees.read_from_file(args.input_dep_trees)
with Timer("Pattern generating..."):
# generate the tree-based position features according the Dependency Tree.
data.pattern_dict.accept_new = False
data.syntactic_dict.accept_new = False
test_dep_trees.get_syntactic_label_dict(data.syntactic_dict)
test_pattern = test_dep_trees.get_tpf2_dict(data.test_tensors, data.pattern_dict)
print("Extract {} test Pattern features".format(len(test_pattern)))
assert len(test_pattern) == len(data.test_tensors)
with Timer("Get test sentences dict"):
test_sentences_w_id = []
for sen in get_srl_sentences(args.input):
test_sentences_w_id.append(' '.join(sen[1]))
test_sentences_ids = [int(sen[0][0]) for sen in test_sentences]
temp = {}
assert len(test_sentences_w_id) == len(test_sentences_ids)
for idx, sen in zip(test_sentences_ids, test_sentences_w_id):
temp[idx] = sen
test_sentences_w_id = temp
with Timer("Loading ELMO"):
test_elmo_hdf5 = hdf5_reader()
test_elmo_hdf5.read_from_file(args.input_elmo, test_sentences_w_id)
with Timer('Model building and loading'):
model = BiLSTMTaggerModel(data, config=config, gpu_id=args.gpu)
model.load(model_path)
for param in model.parameters():
print(param.size())
if args.gpu:
print("Initialize the model with GPU!")
model = model.cuda()
with Timer('Running model'):
scores = []
model.eval()
for i, batched_tensor in enumerate(test_data):
x, y, lengths, weights = batched_tensor
word_inputs_seqs, predicate_inputs_seqs, pattern_inputs_seqs, pattern_li_seqs, pattern_la_seqs, pattern_lp_seqs, \
sentences_ids, answers, input_lengths, masks, padding_answers = \
batch_data_variable(test_pattern, x, y, lengths, weights)
elmo_representations = test_elmo_hdf5.forward(sentences_ids, word_inputs_seqs.size()[-1],
[len(ans) for ans in answers])
if args.gpu:
word_inputs_seqs, predicate_inputs_seqs, pattern_inputs_seqs, pattern_li_seqs, pattern_la_seqs, pattern_lp_seqs, \
input_lengths, masks, padding_answers = \
word_inputs_seqs.cuda(), predicate_inputs_seqs.cuda(), pattern_inputs_seqs.cuda(), pattern_li_seqs.cuda(), \
pattern_la_seqs.cuda(), pattern_lp_seqs.cuda(), input_lengths.cuda(), masks.cuda(), padding_answers.cuda()
elmo_representations = elmo_representations.cuda()
sc = model.forward(word_inputs_seqs, predicate_inputs_seqs, pattern_inputs_seqs,
pattern_li_seqs, pattern_la_seqs, pattern_lp_seqs, elmo_representations,
input_lengths)
sc = sc.data.cpu().numpy() if args.gpu else sc.data.numpy()
sc = [sc[j] for j in range(sc.shape[0])]
scores.extend(sc)
return scores, data, test_sentences, test_data