def build_tasks_from_file(conf_path, options=None):
if type(conf_path) is str:
conf = Config.read(conf_path)
elif type(conf_path) is Config:
conf = conf_path
else:
raise TypeError('Unknown configuration type. Expect str or Config.')
if options:
for k, v in options:
conf.update_value(k, v)
# Create data sets
logger.info('Loading data sets')
datasets = {}
lang_datasets = defaultdict(list)
task_datasets = defaultdict(list)
for dataset in conf.datasets:
parser = create_parser(dataset.parser.format, dataset.parser)
(train_conf, dev_conf,
test_conf) = dataset.clone(), dataset.clone(), dataset.clone()
train_conf.update({'path': dataset.files.train, 'parser': parser})
dev_conf.update({
'path': dataset.files.dev,
'parser': parser,
'sample': None
})
train_dataset = create_dataset(dataset.type, train_conf)
dev_dataset = create_dataset(dataset.type, dev_conf)
if hasattr(dataset.files, 'test'):
test_conf.update({
'path': dataset.files.test,
'parser': parser,
'sample': None
})
test_dataset = create_dataset(dataset.type, test_conf)
datasets[dataset.name] = {
'train': train_dataset,
'dev': dev_dataset,
'test': test_dataset,
'language': dataset.language,
'task': dataset.task
}
lang_datasets[dataset.language].append(dataset.name)
task_datasets[dataset.task].append(dataset.name)
# Create vocabs
# I only keep words in the data sets to save memory
# If the model will be applied to an unknown test set, it is better to keep
# all words in pre-trained embeddings.
logger.info('Creating vocabularies')
dataset_counts = {}
lang_token_vocabs = {}
task_label_vocabs = {}
for name, ds in datasets.items():
dataset_counts[name] = compute_metadata(
[ds['train'], ds['dev'], ds['test']])
for lang, ds in lang_datasets.items():
counts = [dataset_counts[d][0] for d in ds]
lang_token_vocabs[lang] = count2vocab(counts,
ignore_case=True,
start_idx=2)
for task, ds in task_datasets.items():
counts = [dataset_counts[d][1] for d in ds]
task_label_vocabs[task] = count2vocab(counts,
ignore_case=False,
start_idx=0,
sort=True)
char_vocab = count2vocab([c[2] for c in dataset_counts.values()],
ignore_case=False,
start_idx=1)
# Report stats
for lang, vocab in lang_token_vocabs.items():
logger.info('#{} token: {}'.format(lang, len(vocab)))
for task, vocab in task_label_vocabs.items():
logger.info('#{} label: {}'.format(task, len(vocab)))
logger.info(vocab)
# Numberize datasets
logger.info('Numberizing data sets')
numberize_conf = []
for ds in datasets.values():
numberize_conf.append((ds['train'], lang_token_vocabs[ds['language']],
task_label_vocabs[ds['task']], char_vocab))
numberize_conf.append((ds['dev'], lang_token_vocabs[ds['language']],
task_label_vocabs[ds['task']], char_vocab))
numberize_conf.append((ds['test'], lang_token_vocabs[ds['language']],
task_label_vocabs[ds['task']], char_vocab))
numberize_datasets(numberize_conf,
token_ignore_case=True,
label_ignore_case=False,
char_ignore_case=False)
# Initialize component confs
logger.info('Initializing component configurations')
word_embed_dim = char_embed_dim = lstm_output_dim = 0
cpnt_confs = {}
for cpnt in conf.components:
if cpnt.model == 'embedding':
cpnt.embedding_dim = cpnt.dimension
word_embed_dim = cpnt.dimension
elif cpnt.model == 'char_cnn':
cpnt.vocab_size = len(char_vocab)
char_embed_dim = sum([x[1] for x in cpnt.filters])
elif cpnt.model == 'lstm':
lstm_output_dim = cpnt.hidden_size * (2
if cpnt.bidirectional else 1)
cpnt_confs[cpnt.name] = cpnt.clone()
# Update component configurations
target_task = ''
target_lang = ''
for task_conf in conf.tasks:
language = task_conf.language
task = task_conf.task
if task_conf.get('ref', False):
target_lang = language
target_task = task
model_conf = task_conf.model
if model_conf.model != 'lstm_crf':
continue
# Update word embedding configuration
cpnt_confs[model_conf.word_embed].num_embeddings = len(
lang_token_vocabs[language])
cpnt_confs[model_conf.word_embed].vocab = lang_token_vocabs[language]
# Update output layer configuration
cpnt_confs[model_conf.univ_layer].out_features = len(
task_label_vocabs[task])
if hasattr(model_conf, 'spec_layer'):
cpnt_confs[model_conf.spec_layer].out_features = len(
task_label_vocabs[task])
# Update CRF configuration
cpnt_confs[model_conf.crf].label_vocab = task_label_vocabs[task]
for _, cpnt_conf in cpnt_confs.items():
if cpnt_conf.model == 'linear' and cpnt_conf.position == 'output':
cpnt_conf.in_features = lstm_output_dim
if cpnt_conf.model == 'lstm':
cpnt_conf.input_size = char_embed_dim + word_embed_dim
if cpnt_conf.model == 'highway' and cpnt_conf.position == 'char':
cpnt_conf.size = char_embed_dim
# Create components
logger.info('Creating components')
components = {k: create_module(v.model, v) for k, v in cpnt_confs.items()}
# Construct models
tasks = []
for task_conf in conf.tasks:
model_conf = task_conf.model
language = task_conf.language
task = task_conf.task
if model_conf.model == 'lstm_crf':
model = LstmCrf(
lang_token_vocabs[language],
task_label_vocabs[task],
char_vocab,
word_embedding=components[model_conf.word_embed],
char_embedding=components[model_conf.char_embed] if hasattr(
model_conf, 'char_embed') else None,
crf=components[model_conf.crf],
lstm=components[model_conf.lstm],
input_layer=None,
univ_fc_layer=components[model_conf.univ_layer],
spec_fc_layer=components[model_conf.spec_layer] if hasattr(
model_conf, 'spec_linear') else None,
embed_dropout_prob=model_conf.embed_dropout,
lstm_dropout_prob=model_conf.lstm_dropout,
linear_dropout_prob=model_conf.linear_dropout,
char_highway=components[model_conf.char_highway] if hasattr(
model_conf, 'char_highway') else None,
use_char_embedding=model_conf.use_char_embedding if hasattr(
model_conf, 'use_char_embedding') else True,
)
# elif model_conf.model == 'cbow':
# pass
else:
raise ValueError('Unknown model: {}'.format(model_conf.model))
logger.debug(model)
task_classes = {'ner': NameTagging, 'pos': PosTagging}
if task in task_classes:
task_obj = task_classes[task](
task_conf.name,
model,
datasets=datasets[task_conf.dataset],
vocabs={
'token': lang_token_vocabs[language],
'label': task_label_vocabs[task],
'char': char_vocab
},
gpu=task_conf.gpu,
# TODO: 'gpu' -> global config
prob=getattr(task_conf, 'prob', 1.0),
lr=getattr(task_conf, 'learning_rate', .001),
momentum=getattr(task_conf, 'momentum', .9),
decay_rate=getattr(task_conf, 'decay_rate', .9),
decay_step=getattr(task_conf, 'decay_step', 10000),
gradient_clipping=getattr(task_conf, 'gradient_clipping', 5.0),
require_eval=getattr(task_conf, 'require_eval', True),
ref=getattr(task_conf, 'ref', False),
aux_task=task_conf.task != target_task,
aux_lang=task_conf.language != target_lang,
)
else:
raise ValueError('Unknown task {}'.format(task))
tasks.append(task_obj)
return tasks, {
'lang_token_vocabs': lang_token_vocabs,
'task_token_vocabs': task_label_vocabs,
'components': components
}
# Load datasets
logger.info('Loading datasets')
train_set = SequenceDataset(Config({
'path': args.train, 'parser': conll_parser, 'batch_size': args.batch_size}))
dev_set = SequenceDataset(Config({
'path': args.dev, 'parser': conll_parser}))
test_set = SequenceDataset(Config({
'path': args.test, 'parser': conll_parser}))
datasets = {'train': train_set, 'dev': dev_set, 'test': test_set}
# Vocabs
logger.info('Building vocabularies')
token_count, label_count, char_count = compute_metadata(
[train_set, dev_set, test_set])
token_vocab = count2vocab([token_count],
start_idx=C.EMBED_START_IDX,
ignore_case=word_ignore_case)
label_vocab = count2vocab([label_count],
start_idx=0,
sort=True,
ignore_case=False)
char_vocab = count2vocab([char_count],
ignore_case=False,
start_idx=C.CHAR_EMBED_START_IDX)
if embed_file:
logger.info('Scaning pre-trained embeddings')
token_vocab = {}
with open(embed_file, 'r', encoding='utf-8') as embed_r:
if args.embed_skip_first:
embed_r.readline()
for line in embed_r:
skip_comment=True)
train_set = SeqLabelDataset(args.train, parser=parser)
dev_set = SeqLabelDataset(args.dev, parser=parser)
test_set = SeqLabelDataset(args.test, parser=parser)
datasets = {'train': train_set, 'dev': dev_set, 'test': test_set}
# Vocabs
logger.info('Building vocabs')
token_count, char_count, label_count = Counter(), Counter(), Counter()
for _, ds in datasets.items():
tc, cc, lc = ds.stats()
token_count.update(tc)
char_count.update(cc)
label_count.update(lc)
token_vocab = count2vocab(token_count,
offset=len(C.TOKEN_PADS),
pads=C.TOKEN_PADS)
char_vocab = count2vocab(char_count, offset=len(C.CHAR_PADS), pads=C.CHAR_PADS)
label_vocab = count2vocab(label_count, offset=1, pads=[(C.PAD, C.PAD_INDEX)])
# print("label_vocab: ", label_vocab) # DEBUG
# idx_token = {v: k for k, v in token_vocab.items()} # not debug
idx_token = {v: k for k, v in token_vocab.items() if k != ''} # DEBUG
# print(idx_token) # DEBUG
# print(idx_token.get(243, "not found")) # DEBUG
# print([str(k) + " " + str(v) for k, v in idx_token.items() # DEBUG
# if (k == 1 or v == 1 or k == '1' or v == '1')]) # DEBUG
idx_label = {v: k for k, v in label_vocab.items()}
train_set.numberize(token_vocab, label_vocab, char_vocab)
dev_set.numberize(token_vocab, label_vocab, char_vocab)
test_set.numberize(token_vocab, label_vocab, char_vocab)
# print("numberized train set:")
'dev': test_set_clct,
'test': test_set_clct
}
# Vocabs
logger.info('Building vocabularies')
token_count_tgt, label_count_tgt, char_count_tgt = compute_metadata(
[train_set_tgt, dev_set_tgt, test_set_tgt])
token_count_cl, label_count_cl, char_count_cl = compute_metadata(
[train_set_cl, dev_set_cl, test_set_cl])
token_count_ct, label_count_ct, char_count_ct = compute_metadata(
[train_set_ct, dev_set_ct, test_set_ct])
token_count_clct, label_count_clct, char_count_clct = compute_metadata(
[train_set_clct, dev_set_clct, test_set_clct])
token_vocab_1 = count2vocab([token_count_tgt, token_count_ct],
start_idx=C.EMBED_START_IDX,
ignore_case=word_ignore_case)
token_vocab_2 = count2vocab([token_count_cl, token_count_clct],
start_idx=C.EMBED_START_IDX,
ignore_case=word_ignore_case)
label_vocab_1 = count2vocab([label_count_tgt, label_count_cl], start_idx=0)
label_vocab_2 = count2vocab([label_count_ct, label_count_clct], start_idx=0)
char_vocab = count2vocab(
[char_count_tgt, char_count_cl, char_count_ct, char_count_clct],
start_idx=C.CHAR_EMBED_START_IDX)
# Scan embedding file
if embed_file_1:
logger.info('Scaning pre-trained embeddings for language 1')
token_vocab_1 = {}
with open(embed_file_1, 'r', encoding='utf-8') as embed_r: