def run_inference(model_dir: str, epoch: Optional[int], device: str, metric: str):
chainer.config.train = False
if device >= 0:
cuda.get_device(device).use()
set_seed()
configs = json.load(open(os.path.join(model_dir, "args")))
snapshot_file, prediction_path = select_snapshot(epoch, metric, model_dir)
logger.debug(f"creat prediction into {prediction_path}")
vocab = Vocabulary.prepare(configs)
num_word_vocab = configs["num_word_vocab"]
num_char_vocab = configs["num_char_vocab"]
num_tag_vocab = configs["num_tag_vocab"]
model = BiLSTM_CRF(configs, num_word_vocab, num_char_vocab, num_tag_vocab)
model_path = os.path.join(model_dir, snapshot_file)
chainer.serializers.load_npz(model_path, model)
logger.debug(f"load {snapshot_file}")
if device >= 0:
model.to_gpu(device)
transformer = DatasetTransformer(vocab)
transform = transformer.transform
test_iterator = create_iterator(
vocab, configs, "test", transform, return_original_sentence=True
)
with open(prediction_path, "w", encoding="utf-8") as file:
for batch in test_iterator:
batch, original_sentences = list(zip(*batch))
in_arrays, t_arrays = converter(batch, device)
p_arrays = model.predict(in_arrays)
word_sentences, t_tag_sentences = list(
zip(*transformer.itransform(in_arrays[0], t_arrays))
)
_, p_tag_sentences = list(
zip(*transformer.itransform(in_arrays[0], p_arrays))
)
sentence_gen = zip(
word_sentences,
t_tag_sentences,
p_tag_sentences,
original_sentences,
) # NOQA
for ws, ts, ps, _os in sentence_gen:
for w, t, p, o in zip(ws, ts, ps, _os):
w = w.replace(" ", "<WHITESPACE>")
o = o.replace(" ", "<WHITESPACE>")
if w != o:
w = f"{w}({o})"
print(f"{w} {t} {p}", file=file)
print(file=file)
def run_inference(
model_dir: str,
epoch: Optional[int],
device: str,
metric: str,
tokenizer: str):
chainer.config.train = False
if device >= 0:
chainer.get_device(device).use()
set_seed()
config = json.load(open(os.path.join(model_dir, "args")))
snapshot_file, prediction_path = select_snapshot(epoch, metric, model_dir)
logger.debug(f"creat prediction into {prediction_path}")
vocab = Vocabulary.prepare(config)
num_word_vocab = config["num_word_vocab"]
num_char_vocab = config["num_char_vocab"]
num_tag_vocab = config["num_tag_vocab"]
model = BiLSTM_CRF(config, num_word_vocab, num_char_vocab, num_tag_vocab)
model_path = os.path.join(model_dir, snapshot_file)
logger.debug(f"load {snapshot_file}")
chainer.serializers.load_npz(model_path, model)
if device >= 0:
model.to_gpu(device)
transformer = DatasetTransformer(vocab)
word_tokenizer = WordTokenizer(tokenizer=tokenizer)
for line in sys.stdin:
input_sentence = [str(t) for t in word_tokenizer.tokenize(line)]
batch = transformer.transform(input_sentence, None)
in_arr, _ = converter([batch])
pd_arr = model.predict(in_arr)
(_, tag_sequence), = transformer.itransform(in_arr[0], pd_arr)
print(' '.join(f"{word}/{tag}" for word, tag in zip(input_sentence, tag_sequence))) # NOQA
def run_training(config: str, device: int, seed: int):
configs = ConfigParser.parse(config)
params = yaml.load(open(config, encoding="utf-8"))
if device >= 0:
cuda.get_device(device).use()
set_seed(seed, device)
vocab = Vocabulary.prepare(configs)
num_word_vocab = max(vocab.dictionaries["word2idx"].values()) + 1
num_char_vocab = max(vocab.dictionaries["char2idx"].values()) + 1
num_tag_vocab = max(vocab.dictionaries["tag2idx"].values()) + 1
model = BiLSTM_CRF(configs, num_word_vocab, num_char_vocab, num_tag_vocab)
transformer = DatasetTransformer(vocab)
transform = transformer.transform
external_configs = configs["external"]
if "word_vector" in external_configs:
syn0 = model.embed_word.W.data
_, word_dim = syn0.shape
pre_word_dim = vocab.gensim_model.vector_size
if word_dim != pre_word_dim:
msg = "Mismatch vector size between model and pre-trained word vectors" # NOQA
msg += f"(model: \x1b[31m{word_dim}\x1b[0m"
msg += f", pre-trained word vector: \x1b[31m{pre_word_dim}\x1b[0m"
raise Exception(msg)
word2idx = vocab.dictionaries["word2idx"]
syn0 = prepare_pretrained_word_vector(word2idx, vocab.gensim_model,
syn0, num_word_vocab)
model.set_pretrained_word_vectors(syn0)
train_iterator = create_iterator(vocab, configs, "train", transform)
valid_iterator = create_iterator(vocab, configs, "valid", transform)
test_iterator = create_iterator(vocab, configs, "test", transform)
if device >= 0:
model.to_gpu(device)
optimizer = create_optimizer(configs)
optimizer.setup(model)
optimizer = add_hooks(optimizer, configs)
updater = T.StandardUpdater(train_iterator,
optimizer,
converter=converter,
device=device)
params = configs.export()
params["num_word_vocab"] = num_word_vocab
params["num_char_vocab"] = num_char_vocab
params["num_tag_vocab"] = num_tag_vocab
epoch = configs["iteration"]["epoch"]
trigger = (epoch, "epoch")
model_path = configs["output"]
timestamp = datetime.datetime.now()
timestamp_str = timestamp.isoformat()
output_path = Path(f"{model_path}.{timestamp_str}")
trainer = T.Trainer(updater, trigger, out=output_path)
save_args(params, output_path)
msg = f"Create \x1b[31m{output_path}\x1b[0m for saving model snapshots"
logging.debug(msg)
entries = ["epoch", "iteration", "elapsed_time", "lr", "main/loss"]
entries += ["validation/main/loss", "validation/main/fscore"]
entries += ["validation_1/main/loss", "validation_1/main/fscore"]
valid_evaluator = NamedEntityEvaluator(valid_iterator,
model,
transformer.itransform,
converter,
device=device)
test_evaluator = NamedEntityEvaluator(test_iterator,
model,
transformer.itransform,
converter,
device=device)
epoch_trigger = (1, "epoch")
snapshot_filename = "snapshot_epoch_{.updater.epoch:04d}"
trainer.extend(valid_evaluator, trigger=epoch_trigger)
trainer.extend(test_evaluator, trigger=epoch_trigger)
trainer.extend(E.observe_lr(), trigger=epoch_trigger)
trainer.extend(E.LogReport(trigger=epoch_trigger))
trainer.extend(E.PrintReport(entries=entries), trigger=epoch_trigger)
trainer.extend(E.ProgressBar(update_interval=20))
trainer.extend(E.snapshot_object(model, filename=snapshot_filename),
trigger=(1, "epoch"))
if "learning_rate_decay" in params:
logger.debug("Enable Learning Rate decay")
trainer.extend(
LearningRateDecay("lr", params["learning_rate"],
params["learning_rate_decay"]),
trigger=epoch_trigger,
)
trainer.run()
import chainer
import pathlib
import logging
import json
if __name__ == '__main__':
logger = logging.getLogger(__name__)
fmt = '%(asctime)s : %(threadName)s : %(levelname)s : %(message)s'
logging.basicConfig(level=logging.DEBUG, format=fmt)
args = parse_inference_args()
chainer.config.train = False
if args.device >= 0:
chainer.cuda.get_device(args.device).use()
set_seed()
model_dir = pathlib.Path(args.model)
configs = json.load(open(model_dir / 'args'))
metric = args.metric.replace('/', '.')
snapshot_file, prediction_path = select_snapshot(args, model_dir)
logger.debug(f'creat prediction into {prediction_path}')
vocab = Vocabulary.prepare(configs)
num_word_vocab = configs['num_word_vocab']
num_char_vocab = configs['num_char_vocab']
num_tag_vocab = configs['num_tag_vocab']
model = BiLSTM_CRF(configs, num_word_vocab, num_char_vocab, num_tag_vocab)
f'- match1: \x1b[31m{match1}\x1b[0m, match2: \x1b[31m{match2}\x1b[0m'
) # NOQA
return syn0
if __name__ == '__main__':
logger = logging.getLogger(__name__)
fmt = '[%(name)s] %(asctime)s : %(threadName)s : %(levelname)s : %(message)s' # NOQA
logging.basicConfig(level=logging.DEBUG, format=fmt)
args = parse_train_args()
params = yaml.load(open(args.config, encoding='utf-8'))
if args.device >= 0:
chainer.cuda.get_device(args.device).use()
set_seed(args.seed, args.device)
configs = ConfigParser.parse(args.config)
config_path = Path(args.config)
vocab = Vocabulary.prepare(configs)
num_word_vocab = max(vocab.dictionaries['word2idx'].values()) + 1
num_char_vocab = max(vocab.dictionaries['char2idx'].values()) + 1
num_tag_vocab = max(vocab.dictionaries['tag2idx'].values()) + 1
model = BiLSTM_CRF(configs, num_word_vocab, num_char_vocab, num_tag_vocab)
transformer = DatasetTransformer(vocab)
transform = transformer.transform
external_configs = configs['external']