def train_gemb(args):
config = configuration.get_config(args.config)
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 for eval script
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:
# 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_gemb_development_data(
batch_size=config.dev_batch_size) # data for NN
print('Dev data has {} batches.'.format(len(batched_dev_data)))
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'))
writer = open(os.path.join(args.model, 'gemb_checkpoints.tsv'), 'w')
writer.write(
'step\tdatetime\tdev_loss\tdev_accuracy\tbest_dev_accuracy\n')
with Timer('Building model'):
model = BiLSTMTaggerModel(data, config=config)
model.load(os.path.join(args.model, 'model.npz'))
model.add_gemb()
for param in model.gemb.params:
print param, param.name, param.shape.eval()
gemb_loss_function = model.get_gemb_loss_function()
gemb_eval_function = model.get_eval_with_gemb_function()
ctx_emb_function = model.get_ctx_emb_function()
while epoch < config.max_epochs:
with Timer("Epoch%d" % epoch) as timer:
train_data = data.get_gemb_training_data(include_last_batch=True)
for batched_tensor in train_data:
x, y, oov_pos, _, weights = batched_tensor
loss = gemb_loss_function(x, weights, oov_pos)
train_loss += loss
i += 1
global_step += 1
if i % 400 == 0:
timer.tick("{} training steps, loss={:.3f}".format(
i, train_loss / i))
train_loss = train_loss / i
print("Epoch {}, steps={}, loss={:.3f}".format(epoch, i, train_loss))
i = 0
epoch += 1
train_loss = 0.0
if epoch % config.checkpoint_every_x_epochs == 0:
with Timer('Evaluation'):
evaluate_tagger(model, ctx_emb_function, gemb_eval_function,
batched_dev_data, evaluator, writer,
global_step)
# Done. :)
writer.close()
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("Get training and devlel sentences dict"):
training_sentences = []
for sen in get_srl_sentences(args.train):
if len(sen[1]) <= config.max_train_length:
training_sentences.append(' '.join(sen[1]))
training_ids = [int(sen[0][0]) for sen in data.train_sents]
temp = {}
assert len(training_sentences) == len(training_ids)
for idx, sen in zip(training_ids, training_sentences):
temp[idx] = sen
training_sentences = temp
devel_sentences = [
' '.join(sen[1]) for sen in get_srl_sentences(args.dev)
]
devel_ids = [int(sen[0][0]) for sen in data.dev_sents]
temp = {}
assert len(devel_sentences) == len(devel_ids)
for idx, sen in zip(devel_ids, devel_sentences):
temp[idx] = sen
devel_sentences = temp
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)
with Timer("TPF2 generating..."):
# generate the tree-based position features according the Dependency Tree.
train_tpf2, data.tpf2_dict = train_dep_trees.get_tpf2_dict(
data.train_tensors)
print("Extract {} training TPF2 features".format(len(train_tpf2)))
assert len(train_tpf2) == len(data.train_tensors)
data.tpf2_dict.accept_new = False
dev_tpf2 = dev_dep_trees.get_tpf2_dict(data.dev_tensors,
data.tpf2_dict)
print("Extract {} dev TPF2 features".format(len(dev_tpf2)))
assert len(dev_tpf2) == len(data.dev_tensors)
with Timer("Loading ELMO"):
train_elmo_hdf5 = hdf5_reader()
train_elmo_hdf5.read_from_file(args.train_elmo, training_sentences)
dev_elmo_hdf5 = hdf5_reader()
dev_elmo_hdf5.read_from_file(args.dev_elmo, devel_sentences)
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'))
data.tpf2_dict.save(os.path.join(args.model, 'tpf2_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
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, tpf_ids, sentences_ids, answers, input_lengths, masks, padding_answers = \
batch_data_variable(train_tpf2, x, y, lengths, weights)
elmo_representations = train_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, tpf_ids, input_lengths, masks, padding_answers = \
word_inputs_seqs.cuda(), predicate_inputs_seqs.cuda(), tpf_ids.cuda(), input_lengths.cuda(), masks.cuda(), padding_answers.cuda()
elmo_representations = elmo_representations.cuda()
optimizer.zero_grad()
output = model.forward(word_inputs_seqs, predicate_inputs_seqs,
tpf_ids, elmo_representations,
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)))
sys.stdout.flush()
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, batched_dev_data, dev_tpf2,
dev_elmo_hdf5, evaluator, writer,
global_step)
# Done. :)
writer.close()
# Getting evaluator
gold_props_file = args.gold if args.gold != '' else None
pred_props_file = args.inputprops if args.inputprops != '' else None
if args.task == 'srl':
evaluator = SRLEvaluator(data.get_test_data(test_sentences,
batch_size=None),
data.label_dict,
gold_props_file,
use_se_marker=config.use_se_marker,
pred_props_file=pred_props_file,
word_dict=data.word_dict)
else:
evaluator = PropIdEvaluator(
data.get_test_data(test_sentences, batch_size=None),
data.label_dict)
if args.proto != '':
print 'Writing to proto {}'.format(args.proto)
pb_file = open(args.proto, 'wb')
else:
pb_file = None
with Timer("Decoding"):
transition_params = get_transition_params(data.label_dict.idx2str)
num_tokens = None
# Collect sentence length information
for (i, batched_tensor) in enumerate(test_data):
_, _, nt, _ = batched_tensor
def train_tagger(args):
config = configuration.get_config(args.config)
numpy.random.seed(666)
torch.manual_seed(666)
torch.cuda.manual_seed(666)
### gpu
gpu = torch.cuda.is_available()
print("GPU available: ", 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:
# 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('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'))
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 "Use 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()
"""for param in model.params:
print param, param.name, param.shape.eval()
loss_function = model.get_loss_function()
eval_function = model.get_eval_function()"""
while epoch < config.max_epochs:
with Timer("Epoch%d" % epoch) as timer:
train_data = data.get_training_data(include_last_batch=True)
model.bilstm.dropout = 0.1
for batched_tensor in train_data: # for each batch in the training corpus
x, y, _, weights = batched_tensor
batch_input_lengths = ([sentence_x.shape[0] for sentence_x in x])
max_length = max(batch_input_lengths)
# padding
# input = [numpy.pad(sentence_x, (0, max_length - sentence_x.shape[0]), 'constant') for sentence_x in x]
word_input = [numpy.pad(sentence_x[:, 0], (0, max_length - sentence_x.shape[0]), 'constant') \
for sentence_x in x] # padding
predicate_input = [numpy.pad(sentence_x[:, 1], (0, max_length - sentence_x.shape[0]), 'constant') \
for sentence_x in x] # padding
word_input, predicate_input = numpy.vstack(word_input), numpy.vstack(predicate_input)
# numpy batch input to Variable
word_input_seqs = torch.autograd.Variable(torch.from_numpy(word_input.astype('int64')).long())
predicate_input_seqs = torch.autograd.Variable(torch.from_numpy(predicate_input.astype('int64')).long())
# First: order the batch by decreasing sequence length
input_lengths = torch.LongTensor(batch_input_lengths)
input_lengths, perm_idx = input_lengths.sort(0, descending=True)
word_input_seqs = word_input_seqs[perm_idx]
predicate_input_seqs = predicate_input_seqs[perm_idx]
answer = [None] * len(x) # resort the answer according to the input
count = 0
list_y = list(y)
for (i, ans) in zip(perm_idx, list_y):
answer[count] = list_y[i]
count += 1
answer = numpy.concatenate(answer)
answer = torch.autograd.Variable(torch.from_numpy(answer).type(torch.LongTensor))
answer = answer.view(-1)
# print answer, answer.size()
# Then pack the sequences
# packed_input = torch.nn.utils.rnn.pack_padded_sequence(input_seqs, input_lengths.numpy(), batch_first=True)
# packed_input = packed_input.cuda() if args.gpu else packed_input
if args.gpu:
word_input_seqs, predicate_input_seqs, input_lengths, perm_idx =\
word_input_seqs.cuda(), predicate_input_seqs.cuda(), input_lengths.cuda(), perm_idx.cuda()
answer = answer.cuda()
model.zero_grad()
output = model.forward(word_input_seqs, predicate_input_seqs, input_lengths, perm_idx, len(x)) # (batch input, batch size)
loss = model.loss(output, answer)
loss.backward()
# gradient clipping
# torch.nn.utils.clip_grad_norm(model.parameters(), 1.0)
model.optimizer = torch.optim.Adadelta(model.parameters(), rho=0.95)
model.optimizer.step()
train_loss += loss
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, batched_dev_data, evaluator, writer, global_step)
# Done. :)
writer.close()