def main():
args = parse_args()
random.seed(args.seed)
args = vars(args)
print("[Launching identity lemmatizer...]")
if args['mode'] == 'train':
print(
"[No training is required; will only generate evaluation output...]"
)
doc, metasentences = CoNLL.conll2dict(input_file=args['eval_file'])
document = Document(doc, metasentences=metasentences)
batch = DataLoader(document,
args['batch_size'],
args,
evaluation=True,
conll_only=True)
system_pred_file = args['output_file']
gold_file = args['gold_file']
# use identity mapping for prediction
preds = batch.doc.get([TEXT])
# write to file and score
batch.doc.set([LEMMA], preds)
CoNLL.dict2conll(batch.doc.to_dict(), system_pred_file)
if gold_file is not None:
_, _, score = scorer.score(system_pred_file, gold_file)
print("Lemma score:")
print("{} {:.2f}".format(args['lang'], score * 100))
def evaluate(args):
# file paths
model_file = args['save_dir'] + '/' + args['save_name'] if args['save_name'] is not None \
else '{}/{}_nertagger.pt'.format(args['save_dir'], args['shorthand'])
# load model
use_cuda = args['cuda'] and not args['cpu']
trainer = Trainer(model_file=model_file, use_cuda=use_cuda)
loaded_args, vocab = trainer.args, trainer.vocab
# load config
for k in args:
if k.endswith('_dir') or k.endswith('_file') or k in ['shorthand', 'mode', 'scheme']:
loaded_args[k] = args[k]
# load data
print("Loading data with batch size {}...".format(args['batch_size']))
# TODO
doc = Document(json.load(open(args['eval_file'])))
batch = DataLoader(doc, args['batch_size'], loaded_args, vocab=vocab, evaluation=True)
print("Start evaluation...")
preds = []
for i, b in enumerate(batch):
preds += trainer.predict(b)
system_pred_file = args['output_file']
batch.doc.set(['misc'], [y for x in preds for y in x])
CoNLL.dict2conll(batch.doc.to_dict(), system_pred_file)
gold_tags = batch.tags
_, _, score = scorer.score_by_entity(preds, gold_tags)
#print([(a,b) for a,b in zip(preds,gold_tags)])
print("NER tagger score:")
print("{} {:.2f}".format(args['shorthand'], score*100))
def evaluate(args):
# file paths
system_pred_file = args['output_file']
gold_file = args['gold_file']
model_file = args['save_dir'] + '/' + args['save_name'] if args['save_name'] is not None \
else '{}/{}_parser.pt'.format(args['save_dir'], args['shorthand'])
# load pretrain; note that we allow the pretrain_file to be non-existent
pretrain_file = '{}/{}.pretrain.pt'.format(args['save_dir'], args['shorthand']) if args['pretrain_file'] is None \
else args['pretrain_file']
pretrain = Pretrain(pretrain_file)
# load model
print("Loading model from: {}".format(model_file))
use_cuda = args['cuda'] and not args['cpu']
trainer = Trainer(pretrain=pretrain,
model_file=model_file,
use_cuda=use_cuda)
loaded_args, vocab = trainer.args, trainer.vocab
# load config
for k in args:
if k.endswith('_dir') or k.endswith('_file') or k in ['shorthand'
] or k == 'mode':
loaded_args[k] = args[k]
# load data
print("Loading data with batch size {}...".format(args['batch_size']))
doc, metasentences = CoNLL.conll2dict(input_file=args['eval_file'])
doc = Document(doc, metasentences=metasentences)
batch = DataLoader(doc,
args['batch_size'],
loaded_args,
pretrain,
vocab=vocab,
evaluation=True,
sort_during_eval=True)
if len(batch) > 0:
print("Start evaluation...")
preds = []
for i, b in enumerate(batch):
preds += trainer.predict(b)
else:
# skip eval if dev data does not exist
preds = []
preds = utils.unsort(preds, batch.data_orig_idx)
# write to file and score
batch.doc.set([HEAD, DEPREL], [y for x in preds for y in x])
CoNLL.dict2conll(batch.doc.to_dict(), system_pred_file)
if gold_file is not None:
_, _, score = scorer.score(system_pred_file, gold_file)
print("Parser score:")
print("{} {:.2f}".format(args['shorthand'], score * 100))
def test_depparse_with_pretagged_doc():
nlp = classla.Pipeline(**{'processors': 'depparse', 'dir': TEST_MODELS_DIR, 'lang': 'en',
'depparse_pretagged': True})
doc, metasentences = CoNLL.conll2dict(input_file=EN_DOC_CONLLU_PRETAGGED)
doc = classla.Document(doc, metasentences=metasentences)
processed_doc = nlp(doc)
assert EN_DOC_DEPENDENCY_PARSES_GOLD == '\n\n'.join(
[sent.dependencies_string() for sent in processed_doc.sentences])
def test_conllu(processed_doc):
assert CoNLL.conll_as_string(CoNLL.convert_dict(
processed_doc.to_dict())) == EN_DOC_CONLLU_GOLD
def train(args):
utils.ensure_dir(args['save_dir'])
model_file = args['save_dir'] + '/' + args['save_name'] if args['save_name'] is not None \
else '{}/{}_tagger.pt'.format(args['save_dir'], args['shorthand'])
# load pretrained vectors
vec_file = args['wordvec_file']
pretrain_file = '{}/{}.pretrain.pt'.format(args['save_dir'], args['shorthand']) if args['pretrain_file'] is None \
else args['pretrain_file']
pretrain = Pretrain(pretrain_file, vec_file, args['pretrain_max_vocab'])
# load data
print("Loading data with batch size {}...".format(args['batch_size']))
doc, metasentences = CoNLL.conll2dict(input_file=args['train_file'])
train_doc = Document(doc, metasentences=metasentences)
train_batch = DataLoader(train_doc, args['batch_size'], args, pretrain, evaluation=False)
vocab = train_batch.vocab
doc, metasentences = CoNLL.conll2dict(input_file=args['eval_file'])
dev_doc = Document(doc, metasentences=metasentences)
dev_batch = DataLoader(dev_doc, args['batch_size'], args, pretrain, vocab=vocab, evaluation=True, sort_during_eval=True)
# pred and gold path
system_pred_file = args['output_file']
gold_file = args['gold_file']
# skip training if the language does not have training or dev data
if len(train_batch) == 0 or len(dev_batch) == 0:
print("Skip training because no data available...")
sys.exit(0)
print("Training tagger...")
trainer = Trainer(args=args, vocab=vocab, pretrain=pretrain, use_cuda=args['cuda'])
global_step = 0
max_steps = args['max_steps']
dev_score_history = []
best_dev_preds = []
current_lr = args['lr']
global_start_time = time.time()
format_str = '{}: step {}/{}, loss = {:.6f} ({:.3f} sec/batch), lr: {:.6f}'
if args['adapt_eval_interval']:
args['eval_interval'] = utils.get_adaptive_eval_interval(dev_batch.num_examples, 2000, args['eval_interval'])
print("Evaluating the model every {} steps...".format(args['eval_interval']))
using_amsgrad = False
last_best_step = 0
# start training
train_loss = 0
while True:
do_break = False
for i, batch in enumerate(train_batch):
start_time = time.time()
global_step += 1
loss = trainer.update(batch, eval=False) # update step
train_loss += loss
if global_step % args['log_step'] == 0:
duration = time.time() - start_time
print(format_str.format(datetime.now().strftime("%Y-%m-%d %H:%M:%S"), global_step,\
max_steps, loss, duration, current_lr))
if global_step % args['eval_interval'] == 0:
# eval on dev
print("Evaluating on dev set...")
dev_preds = []
for batch in dev_batch:
preds = trainer.predict(batch)
dev_preds += preds
dev_preds = utils.unsort(dev_preds, dev_batch.data_orig_idx)
dev_batch.doc.set([UPOS, XPOS, FEATS], [y for x in dev_preds for y in x])
CoNLL.dict2conll(dev_batch.doc.to_dict(), system_pred_file)
_, _, dev_score = scorer.score(system_pred_file, gold_file)
train_loss = train_loss / args['eval_interval'] # avg loss per batch
print("step {}: train_loss = {:.6f}, dev_score = {:.4f}".format(global_step, train_loss, dev_score))
train_loss = 0
# save best model
if len(dev_score_history) == 0 or dev_score > max(dev_score_history):
last_best_step = global_step
trainer.save(model_file)
print("new best model saved.")
best_dev_preds = dev_preds
dev_score_history += [dev_score]
print("")
if global_step - last_best_step >= args['max_steps_before_stop']:
if not using_amsgrad:
print("Switching to AMSGrad")
last_best_step = global_step
using_amsgrad = True
trainer.optimizer = optim.Adam(trainer.model.parameters(), amsgrad=True, lr=args['lr'], betas=(.9, args['beta2']), eps=1e-6)
else:
do_break = True
break
if global_step >= args['max_steps']:
do_break = True
break
if do_break: break
train_batch.reshuffle()
print("Training ended with {} steps.".format(global_step))
best_f, best_eval = max(dev_score_history)*100, np.argmax(dev_score_history)+1
print("Best dev F1 = {:.2f}, at iteration = {}".format(best_f, best_eval * args['eval_interval']))
def output_predictions(output_file, trainer, data_generator, vocab, mwt_dict, max_seqlen=1000, orig_text=None, no_ssplit=False):
paragraphs = []
for i, p in enumerate(data_generator.sentences):
start = 0 if i == 0 else paragraphs[-1][2]
length = sum([len(x) for x in p])
paragraphs += [(i, start, start+length, length+1)] # para idx, start idx, end idx, length
paragraphs = list(sorted(paragraphs, key=lambda x: x[3], reverse=True))
all_preds = [None] * len(paragraphs)
all_raw = [None] * len(paragraphs)
eval_limit = max(3000, max_seqlen)
batch_size = trainer.args['batch_size']
batches = int((len(paragraphs) + batch_size - 1) / batch_size)
t = 0
for i in range(batches):
batchparas = paragraphs[i * batch_size : (i + 1) * batch_size]
offsets = [x[1] for x in batchparas]
t += sum([x[3] for x in batchparas])
batch = data_generator.next(eval_offsets=offsets)
raw = batch[3]
N = len(batch[3][0])
if N <= eval_limit:
pred = np.argmax(trainer.predict(batch), axis=2)
else:
idx = [0] * len(batchparas)
Ns = [p[3] for p in batchparas]
pred = [[] for _ in batchparas]
while True:
ens = [min(N - idx1, eval_limit) for idx1, N in zip(idx, Ns)]
en = max(ens)
batch1 = batch[0][:, :en], batch[1][:, :en], batch[2][:, :en], [x[:en] for x in batch[3]]
pred1 = np.argmax(trainer.predict(batch1), axis=2)
for j in range(len(batchparas)):
sentbreaks = np.where((pred1[j] == 2) + (pred1[j] == 4))[0]
if len(sentbreaks) <= 0 or idx[j] >= Ns[j] - eval_limit:
advance = ens[j]
else:
advance = np.max(sentbreaks) + 1
pred[j] += [pred1[j, :advance]]
idx[j] += advance
if all([idx1 >= N for idx1, N in zip(idx, Ns)]):
break
batch = data_generator.next(eval_offsets=[x+y for x, y in zip(idx, offsets)])
pred = [np.concatenate(p, 0) for p in pred]
for j, p in enumerate(batchparas):
len1 = len([1 for x in raw[j] if x != '<PAD>'])
if pred[j][len1-1] < 2:
pred[j][len1-1] = 2
elif pred[j][len1-1] > 2:
pred[j][len1-1] = 4
all_preds[p[0]] = pred[j][:len1]
all_raw[p[0]] = raw[j]
offset = 0
oov_count = 0
doc = []
text = SPACE_RE.sub(' ', orig_text) if orig_text is not None else None
char_offset = 0
use_la_ittb_shorthand = trainer.args['shorthand'] == 'la_ittb'
for j in range(len(paragraphs)):
raw = all_raw[j]
pred = all_preds[j]
current_tok = ''
current_sent = []
for t, p in zip(raw, pred):
if t == '<PAD>':
break
# hack la_ittb
if use_la_ittb_shorthand and t in (":", ";"):
p = 2
offset += 1
if vocab.unit2id(t) == vocab.unit2id('<UNK>'):
oov_count += 1
current_tok += t
if p >= 1:
tok = vocab.normalize_token(current_tok)
assert '\t' not in tok, tok
if len(tok) <= 0:
current_tok = ''
continue
if orig_text is not None:
st = -1
tok_len = 0
for part in SPACE_SPLIT_RE.split(current_tok):
if len(part) == 0: continue
st0 = text.index(part, char_offset) - char_offset
lstripped = part.lstrip()
if st < 0:
st = char_offset + st0 + (len(part) - len(lstripped))
char_offset += st0 + len(part)
additional_info = {START_CHAR: st, END_CHAR: char_offset}
else:
additional_info = dict()
current_sent.append((tok, p, additional_info))
current_tok = ''
if (p == 2 or p == 4) and not no_ssplit:
doc.append(process_sentence(current_sent, mwt_dict))
current_sent = []
assert(len(current_tok) == 0)
if len(current_sent):
doc.append(process_sentence(current_sent, mwt_dict))
if output_file: CoNLL.dict2conll(doc, output_file)
return oov_count, offset, all_preds, doc
def process_pre_tokenized_conllu_text(self, input_src):
"""
Pretokenized text in this case is provided in conllu format.
"""
return CoNLL.conll2dict(input_str=input_src, generate_raw_text=True)
# achieved by enumerating different options of separators that different treebanks might
# use, and comparing that to treating the XPOS tags as separate categories (using a
# WordVocab).
mapping = defaultdict(list)
for sh, fn in zip(shorthands, fullnames):
print('Resolving vocab option for {}...'.format(sh))
if not os.path.exists('data/pos/{}.train.in.conllu'.format(sh)):
raise UserWarning('Training data for {} not found in the data directory, falling back to using WordVocab. To generate the '
'XPOS vocabulary for this treebank properly, please run the following command first:\n'
'\tbash scripts/prep_pos_data.sh {}'.format(fn, fn))
# without the training file, there's not much we can do
key = 'WordVocab(data, shorthand, idx=2)'
mapping[key].append(sh)
continue
doc, metasentences = CoNLL.conll2dict(input_file='data/pos/{}.train.in.conllu'.format(sh))
doc = Document(doc, metasentences=metasentences)
data = doc.get([TEXT, UPOS, XPOS, FEATS], as_sentences=True)
print(f'Original length = {len(data)}')
data = filter_data(data, idx=2)
print(f'Filtered length = {len(data)}')
vocab = WordVocab(data, sh, idx=2, ignore=["_"])
key = 'WordVocab(data, shorthand, idx=2, ignore=["_"])'
best_size = len(vocab) - len(VOCAB_PREFIX)
if best_size > 20:
for sep in ['', '-', '+', '|', ',', ':']: # separators
vocab = XPOSVocab(data, sh, idx=2, sep=sep)
length = sum(len(x) - len(VOCAB_PREFIX) for x in vocab._id2unit.values())
if length < best_size:
key = 'XPOSVocab(data, shorthand, idx=2, sep="{}")'.format(sep)
best_size = length
def test_dict_to_conll():
conll = CoNLL.convert_dict(DICT)
assert conll == CONLL
def test_conll_to_dict():
dicts = CoNLL.convert_conll(CONLL)
assert dicts == DICT
def evaluate(args):
# file paths
system_pred_file = args['output_file']
gold_file = args['gold_file']
model_file = '{}/{}.pt'.format(args['model_dir'], args['model_file'])
# load model
use_cuda = args['cuda'] and not args['cpu']
trainer = Trainer(model_file=model_file, use_cuda=use_cuda)
loaded_args, vocab = trainer.args, trainer.vocab
for k in args:
if k.endswith('_dir') or k.endswith('_file') or k in ['shorthand']:
loaded_args[k] = args[k]
# laod data
print("Loading data with batch size {}...".format(args['batch_size']))
doc, metasentences = CoNLL.conll2dict(input_file=args['eval_file'])
doc = Document(doc, metasentences=metasentences)
batch = DataLoader(doc,
args['batch_size'],
loaded_args,
vocab=vocab,
evaluation=True)
# skip eval if dev data does not exist
if len(batch) == 0:
print("Skip evaluation because no dev data is available...")
print("Lemma score:")
print("{} ".format(args['lang']))
sys.exit(0)
dict_preds = trainer.predict_dict([(e[0].lower(), e[1])
for e in batch.doc.get([TEXT, XPOS])])
if loaded_args.get('dict_only', False):
preds = dict_preds
else:
if loaded_args.get('ensemble_dict', False):
skip = trainer.skip_seq2seq([(e[0].lower(), e[1])
for e in batch.doc.get([TEXT, XPOS])])
doc, metasentences = CoNLL.conll2dict(input_file=args['eval_file'])
dev_doc = Document(doc, metasentences=metasentences)
seq2seq_batch = DataLoader(dev_doc,
args['batch_size'],
loaded_args,
vocab=vocab,
evaluation=True,
skip=skip)
else:
seq2seq_batch = batch
print("Running the seq2seq model...")
preds = []
edits = []
for i, b in enumerate(seq2seq_batch):
ps, es = trainer.predict(b, args['beam_size'])
preds += ps
if es is not None:
edits += es
# preds = trainer.postprocess(batch.doc.get([TEXT]), preds, edits=edits)
if loaded_args.get('ensemble_dict', False):
preds = trainer.postprocess(
[x for x, y in zip(batch.doc.get([TEXT]), skip) if not y],
preds,
edits=edits)
print("[Ensembling dict with seq2seq lemmatizer...]")
i = 0
preds1 = []
for s in skip:
if s:
preds1.append('')
else:
preds1.append(preds[i])
i += 1
preds = trainer.ensemble([(e[0].lower(), e[1])
for e in batch.doc.get([TEXT, XPOS])],
preds1)
else:
preds = trainer.postprocess(batch.doc.get([TEXT]),
preds,
edits=edits)
print("[Ensembling dict with seq2seq lemmatizer...]")
# preds = trainer.ensemble(batch.doc.get([TEXT, UPOS]), preds)
# write to file and score
batch.doc.set([LEMMA], preds)
CoNLL.dict2conll(batch.doc.to_dict(), system_pred_file)
if gold_file is not None:
_, _, score = scorer.score(system_pred_file, gold_file)
print("Lemma score:")
print("{} {:.2f}".format(args['lang'], score * 100))
def train(args):
# load data
print("[Loading data with batch size {}...]".format(args['batch_size']))
doc, metasentences = CoNLL.conll2dict(input_file=args['train_file'])
train_doc = Document(doc, metasentences=metasentences)
train_batch = DataLoader(train_doc,
args['batch_size'],
args,
evaluation=False)
vocab = train_batch.vocab
args['vocab_size'] = vocab['char'].size
args['pos_vocab_size'] = vocab['pos'].size
doc, metasentences = CoNLL.conll2dict(input_file=args['eval_file'])
dev_doc = Document(doc, metasentences=metasentences)
dev_batch = DataLoader(dev_doc,
args['batch_size'],
args,
vocab=vocab,
evaluation=True)
utils.ensure_dir(args['model_dir'])
model_file = '{}/{}_lemmatizer.pt'.format(args['model_dir'], args['lang'])
# pred and gold path
system_pred_file = args['output_file']
gold_file = args['gold_file']
utils.print_config(args)
# skip training if the language does not have training or dev data
if len(train_batch) == 0 or len(dev_batch) == 0:
print("[Skip training because no data available...]")
sys.exit(0)
# start training
# train a dictionary-based lemmatizer
trainer = Trainer(args=args, vocab=vocab, use_cuda=args['cuda'])
print("[Training dictionary-based lemmatizer...]")
dict = train_batch.doc.get([TEXT, XPOS, LEMMA])
dict = [(e[0].lower(), e[1], e[2]) for e in dict]
if args.get('external_dict', None) is not None:
extra_dict = []
for line in open(args['external_dict']):
word, lemma, xpos = line.rstrip('\r\n').split('\t')
extra_dict.append((word.lower(), xpos, lemma))
dict = extra_dict + dict
trainer.train_dict(dict)
print("Evaluating on dev set...")
dev_preds = trainer.predict_dict([
(e[0].lower(), e[1]) for e in dev_batch.doc.get([TEXT, XPOS])
])
dev_batch.doc.set([LEMMA], dev_preds)
CoNLL.dict2conll(dev_batch.doc.to_dict(), system_pred_file)
_, _, dev_f = scorer.score(system_pred_file, gold_file)
print("Dev F1 = {:.2f}".format(dev_f * 100))
if args.get('dict_only', False):
# save dictionaries
trainer.save(model_file)
else:
# train a seq2seq model
print("[Training seq2seq-based lemmatizer...]")
global_step = 0
max_steps = len(train_batch) * args['num_epoch']
dev_score_history = []
best_dev_preds = []
current_lr = args['lr']
global_start_time = time.time()
format_str = '{}: step {}/{} (epoch {}/{}), loss = {:.6f} ({:.3f} sec/batch), lr: {:.6f}'
# start training
for epoch in range(1, args['num_epoch'] + 1):
train_loss = 0
for i, batch in enumerate(train_batch):
start_time = time.time()
global_step += 1
loss = trainer.update(batch, eval=False) # update step
train_loss += loss
if global_step % args['log_step'] == 0:
duration = time.time() - start_time
print(format_str.format(datetime.now().strftime("%Y-%m-%d %H:%M:%S"), global_step,\
max_steps, epoch, args['num_epoch'], loss, duration, current_lr))
# eval on dev
print("Evaluating on dev set...")
dev_preds = []
dev_edits = []
dict_preds = trainer.predict_dict([
(e[0].lower(), e[1]) for e in dev_batch.doc.get([TEXT, XPOS])
])
# for i, batch in enumerate(dev_batch):
# preds, edits = trainer.predict(batch, args['beam_size'])
# dev_preds += preds
# if edits is not None:
# dev_edits += edits
# dev_preds = trainer.postprocess(dev_batch.doc.get([TEXT]), dev_preds, edits=dev_edits)
# try ensembling with dict if necessary
if args.get('ensemble_dict', False):
skip = trainer.skip_seq2seq([
(e[0].lower(), e[1])
for e in dev_batch.doc.get([TEXT, XPOS])
])
doc, metasentences = CoNLL.conll2dict(
input_file=args['eval_file'])
dev_doc = Document(doc, metasentences=metasentences)
seq2seq_batch = DataLoader(dev_doc,
args['batch_size'],
args,
vocab=vocab,
evaluation=True,
skip=skip)
# print("[Ensembling dict with seq2seq model...]")
# dev_preds = trainer.ensemble(dev_batch.doc.get([TEXT, UPOS]), dev_preds)
else:
seq2seq_batch = dev_batch
if args.get('ensemble_dict', False):
dev_preds = trainer.postprocess([
x for x, y in zip(dev_batch.doc.get([TEXT]), skip) if not y
],
dev_preds,
edits=dev_edits)
print("[Ensembling dict with seq2seq lemmatizer...]")
i = 0
preds1 = []
for s in skip:
if s:
preds1.append('')
else:
preds1.append(dev_preds[i])
i += 1
dev_preds = trainer.ensemble(
[(e[0].lower(), e[1])
for e in dev_batch.doc.get([TEXT, XPOS])], preds1)
else:
dev_preds = trainer.postprocess(dev_batch.doc.get([TEXT]),
dev_preds,
edits=dev_edits)
dev_batch.doc.set([LEMMA], dev_preds)
CoNLL.dict2conll(dev_batch.doc.to_dict(), system_pred_file)
_, _, dev_score = scorer.score(system_pred_file, gold_file)
train_loss = train_loss / train_batch.num_examples * args[
'batch_size'] # avg loss per batch
print("epoch {}: train_loss = {:.6f}, dev_score = {:.4f}".format(
epoch, train_loss, dev_score))
# save best model
if epoch == 1 or dev_score > max(dev_score_history):
trainer.save(model_file)
print("new best model saved.")
best_dev_preds = dev_preds
# lr schedule
if epoch > args['decay_epoch'] and dev_score <= dev_score_history[-1] and \
args['optim'] in ['sgd', 'adagrad']:
current_lr *= args['lr_decay']
trainer.update_lr(current_lr)
dev_score_history += [dev_score]
print("")
print("Training ended with {} epochs.".format(epoch))
best_f, best_epoch = max(dev_score_history) * 100, np.argmax(
dev_score_history) + 1
print("Best dev F1 = {:.2f}, at epoch = {}".format(best_f, best_epoch))
def to_conll(self):
""" Produces string output of CoNLLu type.
"""
return CoNLL.conll_as_string(CoNLL.convert_dict(self.to_dict()))