def __init__(self, args, task, models):
self.args = args
self.task = task
self.models = models
self.src_dict = task.source_dictionary
self.tgt_dict = task.target_dictionary
self.use_cuda = torch.cuda.is_available() and not getattr(
args, 'cpu', False)
if self.use_cuda:
if getattr(args, 'fp16', False):
self.half()
self.cuda()
# optimize model for generation
for model in self.models:
model.make_generation_fast_(
beamable_mm_beam_size=(None if getattr(args, 'no_beamable_mm',
False) else getattr(
args, 'beam', 5)),
need_attn=getattr(args, 'print_alignment', False),
)
self.generator = self.task.build_generator(args)
# Load alignment dictionary for unknown word replacement
# (None if no unknown word replacement, empty if no path to align dictionary)
self.align_dict = utils.load_align_dict(
getattr(args, 'replace_unk', None))
self.tokenizer = encoders.build_tokenizer(args)
self.bpe = encoders.build_bpe(args)
def __init__(self, args, task, model):
super().__init__()
self.args = args
self.task = task
self.model = model
self.bpe = encoders.build_bpe(args)
# this is useful for determining the device
self.register_buffer('_float_tensor',
torch.tensor([0], dtype=torch.float))
def main(args):
utils.import_user_module(args)
if args.buffer_size < 1:
args.buffer_size = 1
if args.max_tokens is None and args.max_sentences is None:
args.max_sentences = 1
assert not args.sampling or args.nbest == args.beam, \
'--sampling requires --nbest to be equal to --beam'
assert not args.max_sentences or args.max_sentences <= args.buffer_size, \
'--max-sentences/--batch-size cannot be larger than --buffer-size'
print(args)
use_cuda = torch.cuda.is_available() and not args.cpu
# Setup task, e.g., translation
task = tasks.setup_task(args)
# Load ensemble
print('| loading model(s) from {}'.format(args.path))
models, _model_args = checkpoint_utils.load_model_ensemble(
args.path.split(':'),
arg_overrides=eval(args.model_overrides),
task=task,
)
# Set dictionaries
src_dict = task.source_dictionary
tgt_dict = task.target_dictionary
# Optimize ensemble for generation
for model in models:
model.make_generation_fast_(
beamable_mm_beam_size=None if args.no_beamable_mm else args.beam,
need_attn=args.print_alignment,
)
if args.fp16:
model.half()
if use_cuda:
model.cuda()
# Initialize generator
generator = task.build_generator(args)
# Handle tokenization and BPE
tokenizer = encoders.build_tokenizer(args)
bpe = encoders.build_bpe(args)
def encode_fn(x):
if tokenizer is not None:
x = tokenizer.encode(x)
if bpe is not None:
x = bpe.encode(x)
return x
def decode_fn(x):
if bpe is not None:
x = bpe.decode(x)
if tokenizer is not None:
x = tokenizer.decode(x)
return x
# Load alignment dictionary for unknown word replacement
# (None if no unknown word replacement, empty if no path to align dictionary)
align_dict = utils.load_align_dict(args.replace_unk)
max_positions = utils.resolve_max_positions(
task.max_positions(),
*[model.max_positions() for model in models]
)
if args.buffer_size > 1:
print('| Sentence buffer size:', args.buffer_size)
print('| Type the input sentence and press return:')
start_id = 0
for inputs in buffered_read(args.input, args.buffer_size):
results = []
for batch in make_batches(inputs, args, task, max_positions, encode_fn):
src_tokens = batch.src_tokens
src_lengths = batch.src_lengths
if use_cuda:
src_tokens = src_tokens.cuda()
src_lengths = src_lengths.cuda()
sample = {
'net_input': {
'src_tokens': src_tokens,
'src_lengths': src_lengths,
},
}
translations = task.inference_step(generator, models, sample)
for i, (id, hypos) in enumerate(zip(batch.ids.tolist(), translations)):
src_tokens_i = utils.strip_pad(src_tokens[i], tgt_dict.pad())
results.append((start_id + id, src_tokens_i, hypos))
# sort output to match input order
for id, src_tokens, hypos in sorted(results, key=lambda x: x[0]):
if src_dict is not None:
src_str = src_dict.string(src_tokens, args.remove_bpe)
print('S-{}\t{}'.format(id, src_str))
# Process top predictions
print("nbest:",args.nbest)
print("len(hypos)",len(hypos))
for hypo in hypos[:min(len(hypos), args.nbest)]:
hypo_tokens, hypo_str, alignment = utils.post_process_prediction(
hypo_tokens=hypo['tokens'].int().cpu(),
src_str=src_str,
alignment=hypo['alignment'].int().cpu() if hypo['alignment'] is not None else None,
align_dict=align_dict,
tgt_dict=tgt_dict,
remove_bpe=args.remove_bpe,
)
hypo_str = decode_fn(hypo_str)
print('H-{}\t{}\t{}'.format(id, hypo['score'], hypo_str))
print('P-{}\t{}'.format(
id,
' '.join(map(lambda x: '{:.4f}'.format(x), hypo['positional_scores'].tolist()))
))
if args.print_alignment:
print('A-{}\t{}'.format(
id,
' '.join(map(lambda x: str(utils.item(x)), alignment))
))
# update running id counter
start_id += len(inputs)
def _run(self, sink_embed, sink_token, *receivers):
# Windows does not support logger in MP environment, thus get a new logger
# inside the process for better compatibility
logger = set_logger(colored('WORKER-%d' % self.worker_id, 'yellow'), self.verbose)
for sock, addr in zip(receivers, self.worker_address):
sock.connect(addr)
sink_embed.connect(self.sink_address)
sink_token.connect(self.sink_address)
# 下面是pytorch 代码
if self.args.max_tokens is None and self.args.max_sentences is None:
self.args.max_sentences = 1
assert not self.args.sampling or self.args.nbest == self.args.beam, \
'--sampling requires --nbest to be equal to --beam'
use_cuda = torch.cuda.is_available() and not self.args.cpu
# Setup task, e.g., translation
task = tasks.setup_task(self.args)
# Load ensemble
print('| loading model(s) from {}'.format(self.args.path))
models, _model_args = checkpoint_utils.load_model_ensemble(
self.args.path.split(':'),
arg_overrides=eval(self.args.model_overrides),
task=task,
)
# Set dictionaries
src_dict = task.source_dictionary
tgt_dict = task.target_dictionary
# Optimize ensemble for generation
for model in models:
model.make_generation_fast_(
beamable_mm_beam_size=None if self.args.no_beamable_mm else self.args.beam,
need_attn=self.args.print_alignment,
)
if self.args.fp16:
model.half()
if use_cuda:
model.cuda()
# Initialize generator
generator = task.build_generator(self.args)
# Handle tokenization and BPE
tokenizer = encoders.build_tokenizer(self.args)
bpe = encoders.build_bpe(self.args)
def encode_fn(x):
if tokenizer is not None:
x = tokenizer.encode(x)
if bpe is not None:
x = bpe.encode(x)
return x
def decode_fn(x):
if bpe is not None:
x = bpe.decode(x)
if tokenizer is not None:
x = tokenizer.decode(x)
return x
# Load alignment dictionary for unknown word replacement
# (None if no unknown word replacement, empty if no path to align dictionary)
align_dict = utils.load_align_dict(self.args.replace_unk)
max_positions = utils.resolve_max_positions(
task.max_positions(),
*[model.max_positions() for model in models]
)
start_id = 0
for token in self.input_fn_builder(receivers, sink_token)():
inputs = token["inputs"]
results = []
for batch in self.make_batches(inputs, self.args, task, max_positions, encode_fn):
src_tokens = batch.src_tokens
src_lengths = batch.src_lengths
if use_cuda:
src_tokens = src_tokens.cuda()
src_lengths = src_lengths.cuda()
sample = {
'net_input': {
'src_tokens': src_tokens,
'src_lengths': src_lengths,
},
}
translations = task.inference_step(generator, models, sample)
for i, (id, hypos) in enumerate(zip(batch.ids.tolist(), translations)):
src_tokens_i = utils.strip_pad(src_tokens[i], tgt_dict.pad())
results.append((start_id + id, src_tokens_i, hypos))
res = []
# sort output to match input order
for id, src_tokens, hypos in sorted(results, key=lambda x: x[0]):
if src_dict is not None:
src_str = src_dict.string(src_tokens, self.args.remove_bpe)
res_each = []
for hypo in hypos[:min(len(hypos), self.args.nbest)]:
hypo_tokens, hypo_str, alignment = utils.post_process_prediction(
hypo_tokens=hypo['tokens'].int().cpu(),
src_str=src_str,
alignment=hypo['alignment'].int().cpu() if hypo['alignment'] is not None else None,
align_dict=align_dict,
tgt_dict=tgt_dict,
remove_bpe=self.args.remove_bpe,
)
hypo_str = decode_fn(hypo_str)
res_each.append(hypo_str)
res.append(res_each)
start_id += len(inputs)
send_PAGE(src=sink_embed, dest=token['client_id'], res=res, req_id=ServerCmd.data_restokens)
logger.info('job done\t client: %s' % (token['client_id']))
def load_dataset(self, split, epoch=0, combine=False):
"""Load a given dataset split.
Args:
split (str): name of the split (e.g., train, valid, test)
"""
paths = self.args.data.split(':')
assert len(paths) > 0
data_path = paths[epoch % len(paths)]
split_path = os.path.join(data_path, split)
dataset = data_utils.load_indexed_dataset(
split_path,
self.source_dictionary,
self.args.dataset_impl,
combine=combine,
)
if dataset is None:
raise FileNotFoundError('Dataset not found: {} ({})'.format(split, split_path))
# create continuous blocks of tokens
dataset = TokenBlockDataset(
dataset,
dataset.sizes,
self.args.tokens_per_sample - 1, # one less for <s>
pad=self.source_dictionary.pad(),
eos=self.source_dictionary.eos(),
break_mode=self.args.sample_break_mode,
)
# prepend beginning-of-sentence token (<s>, equiv. to [CLS] in BERT)
dataset = PrependTokenDataset(dataset, self.source_dictionary.bos())
# create masked input and targets
if self.args.mask_whole_words:
bpe = encoders.build_bpe(self.args)
if bpe is not None:
def is_beginning_of_word(i):
if i < self.source_dictionary.nspecial:
# special elements are always considered beginnings
return True
tok = self.source_dictionary[i]
if tok.startswith('madeupword'):
return True
try:
return bpe.is_beginning_of_word(tok)
except ValueError:
return True
mask_whole_words = torch.ByteTensor(list(
map(is_beginning_of_word, range(len(self.source_dictionary)))
))
else:
mask_whole_words = None
src_dataset, tgt_dataset = MaskTokensDataset.apply_mask(
dataset,
self.source_dictionary,
pad_idx=self.source_dictionary.pad(),
mask_idx=self.mask_idx,
seed=self.args.seed,
mask_prob=self.args.mask_prob,
leave_unmasked_prob=self.args.leave_unmasked_prob,
random_token_prob=self.args.random_token_prob,
freq_weighted_replacement=self.args.freq_weighted_replacement,
mask_whole_words=mask_whole_words,
)
with data_utils.numpy_seed(self.args.seed + epoch):
shuffle = np.random.permutation(len(src_dataset))
self.datasets[split] = SortDataset(
NestedDictionaryDataset(
{
'id': IdDataset(),
'net_input': {
'src_tokens': PadDataset(
src_dataset,
pad_idx=self.source_dictionary.pad(),
left_pad=False,
),
'src_lengths': NumelDataset(src_dataset, reduce=False),
},
'target': PadDataset(
tgt_dataset,
pad_idx=self.source_dictionary.pad(),
left_pad=False,
),
'nsentences': NumSamplesDataset(),
'ntokens': NumelDataset(src_dataset, reduce=True),
},
sizes=[src_dataset.sizes],
),
sort_order=[
shuffle,
src_dataset.sizes,
],
)