def load_lm(save_path, checkpoint_name, bpe_code):
lm = TransformerLanguageModel.from_pretrained(save_path,
checkpoint_name,
tokenizer='moses',
bpe='fastbpe',
bpe_codes=bpe_code)
lm.eval()
lm.cuda()
return lm
def load_lm(lm_path, model_type, dict_path):
path, checkpoint = os.path.split(lm_path)
if model_type == "convlm":
model_handle = FConvLanguageModel.from_pretrained(
path, checkpoint,
os.path.split(dict_path)[0])
elif model_type == "transformer":
model_handle = TransformerLanguageModel.from_pretrained(
path, checkpoint,
os.path.split(dict_path)[0])
else:
raise Exception(
"Unsupported language model type: use 'convlm' or 'transformer' models"
)
model = model_handle.models[0].decoder.cuda()
model.eval()
print(model)
return model
def main():
args = parse_args()
logging.basicConfig(level=logging.DEBUG)
logging.debug(args)
tokens = args.prefix.split(" ")
num_tokens = len(tokens)
assert args.sample_length >= num_tokens, "--sample-length (%d) must be equal to or higher than length of --prefix (%d)" % (
args.sample_length, num_tokens)
actual_length = args.sample_length - num_tokens
custom_lm = TransformerLanguageModel.from_pretrained(
args.model_dir,
'checkpoint_best.pt',
verbose=args.verbose,
max_len_b=actual_length)
print(custom_lm.sample(args.prefix, beam=5))
def main():
args = parse_args()
logging.basicConfig(level=logging.DEBUG)
logging.debug(args)
lm = TransformerLanguageModel.from_pretrained(args.model, 'checkpoint_best.pt')
# disable dropout
lm.eval()
if args.cuda:
lm.cuda()
with open(args.input, "r") as infile:
num_lines = sum(1 for line in infile)
logging.debug("Number of lines in input file: %d" % num_lines)
seen = 0
with open(args.input, "r") as infile, open(args.output, "w") as outfile:
for line in infile:
line = line.strip()
score = lm.score(line)['positional_scores'].mean()
if args.score_type in [SCORE_TYPE_PPL, SCORE_TYPE_NEGLOGPROB]:
score = score.neg()
if args.score_type == SCORE_TYPE_PPL:
score = score.exp()
outfile.write("%f\n" % score)
seen += 1
if seen % LOG_INTERVAL == 0:
logging.debug("Processed lines: %d / %d" % (seen, num_lines))
print(output)
# output = ru_lm.sample('Дмитрий Карпов - это ', beam=1, sampling=True, sampling_topk=10, temperature=0.6)
# print(output)
# output = ru_lm.sample('Михаил Бурцев - это ', beam=1, sampling=True, sampling_topk=10, temperature=0.7)
# print(output)
# output = ru_lm.sample('Диляра Баймурзина - это ', beam=1, sampling=True, sampling_topk=10, temperature=0.8)
# print(output)
import ipdb
ipdb.set_trace()
print(output)
#
# ################################################################
from fairseq.models.transformer_lm import TransformerLanguageModel
custom_lm = TransformerLanguageModel.from_pretrained(
'/home/alx/Cloud/spell_corr/py__spelling_corrector/language_models/fairseq_transformer_lm',
'checkpoint100.pt',
tokenizer='moses',
bpe='fastbpe')
custom_lm.sample('Barack Obama', beam=5)
# Sample from the language model
# ru_lm.sample('Barack Obama', beam=1, sampling=True, sampling_topk=10, temperature=0.8)
# "Barack Obama is coming to Sydney and New Zealand (...)"
# #
# # # The same interface can be used with custom models as well
# # from fairseq.models.transformer_lm import TransformerLanguageModel
# # custom_lm = TransformerLanguageModel.from_pretrained('/path/to/model/dir', 'checkpoint100.pt', tokenizer='moses', bpe='fastbpe')
# # custom_lm.sample('Barack Obama', beam=5)
from fairseq.models.transformer_lm import TransformerLanguageModel
lg = "it"
model = TransformerLanguageModel.from_pretrained(
f"models/word/{lg}/transformer",
checkpoint_file="checkpoint_best.pt",
data_name_or_path=f"models/word/{lg}/transformer/bin/",
)
model.eval()
def __init__(self, lm_model_path):
self.lm = TransformerLanguageModel.from_pretrained(
lm_model_path, 'checkpoint_best.pt', tokenizer='moses')
self.lm.eval()
if torch.cuda.is_available():
self.lm.cuda()