Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[WIP] Safe bpe dropout for LM + joiner disjoin dropout #2009

Open
wants to merge 5 commits into
base: master
Choose a base branch
from
Open

[WIP] Safe bpe dropout for LM + joiner disjoin dropout #2009

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
799783e
implement safe bpe dropout for LM + joiner disjoin dropout
Feb 17, 2021
4965928
fix missing model_task in opts
Feb 17, 2021
14cdcf4
disable sp sampling
Feb 17, 2021
98f8189
fix dropped element in JoinerDropout
Feb 17, 2021
003fe4a
fix joiner dropout
Feb 19, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
77 changes: 77 additions & 0 deletions onmt/transforms/misc.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,8 @@
from onmt.utils.logging import logger
from onmt.transforms import register_transform
from .transform import Transform
import random
from onmt.constants import ModelTask, SubwordMarker


@register_transform(name='filtertoolong')
Expand Down Expand Up @@ -41,6 +43,81 @@ def _repr_args(self):
)


@register_transform(name="joiner_dropout")
class JoinerDropoutTransform(Transform):
"""Disjoin joiner with probability dropout"""

def __init__(self, opts):
super().__init__(opts)

@classmethod
def add_options(cls, parser):
"""Avalilable options relate to this Transform."""
group = parser.add_argument_group("Transform/JoinerDropout")
group.add(
"--src_joiner_dropout",
"-src_joiner_dropout",
type=float,
default=0.0,
help="Source dropout probability.",
)
group.add(
"--tgt_joiner_dropout",
"-tgt_joiner_dropout",
type=float,
default=0.0,
help="Target dropout probability.",
)

def _parse_opts(self):
self.src_joiner_dropout = self.opts.src_joiner_dropout
self.tgt_joiner_dropout = self.opts.tgt_joiner_dropout
self.model_task = getattr(self.opts, "model_task", None)

def dropout_separate_joiner(self, seq, side="src"):
out_seq = []
dropout = (
self.src_joiner_dropout
if side == "src"
else self.tgt_joiner_dropout
)
for elem in seq:
if len(elem) > 1 and elem.startswith(SubwordMarker.JOINER):
if random.random() < dropout:
Copy link
Collaborator Author

@funboarder13920 funboarder13920 Feb 17, 2021 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not sure about the necessity of doing right and left token sides
It might create a difficulty to retrieve the initial token when detokenizing.
I can use a special token do distinguish left and right or remove the left side disjoin which doesn't occur much (mainly in punctuation)

out_seq.append(SubwordMarker.JOINER)
elem = elem[1:]
if len(elem) > 1 and elem.endswith(SubwordMarker.JOINER):
if random.random() < dropout:
out_seq.append(elem[:-1])
elem = elem[-1:]
out_seq.append(elem)

return out_seq

def apply(self, example, is_train=False, stats=None, **kwargs):
"""Return None if too long else return as is."""
if not is_train:
return example
else:
src_out = self.dropout_separate_joiner(example["src"], "src")
example["src"] = src_out
if self.model_task == ModelTask.LANGUAGE_MODEL:
example["tgt"] = src_out
else:
tgt_out = self.dropout_separate_joiner(example["tgt"], "tgt")
example["tgt"] = tgt_out
return example

def _repr_args(self):
"""Return str represent key arguments for class."""
return "{}={}, {}={}".format(
"src_joiner_dropout",
self.src_joiner_dropout,
"tgt_joiner_dropout",
self.tgt_joiner_dropout,
)


@register_transform(name='prefix')
class PrefixTransform(Transform):
"""Add Prefix to src (& tgt) sentence."""
Expand Down
71 changes: 53 additions & 18 deletions onmt/transforms/tokenize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,7 @@
from onmt.utils.logging import logger
from onmt.transforms import register_transform
from .transform import Transform
from onmt.constants import ModelTask


class TokenizerTransform(Transform):
Expand Down Expand Up @@ -90,6 +91,7 @@ def _parse_opts(self):
self.tgt_subword_vocab = self.opts.tgt_subword_vocab
self.src_vocab_threshold = self.opts.src_vocab_threshold
self.tgt_vocab_threshold = self.opts.tgt_vocab_threshold
self.model_task = getattr(self.opts, "model_task", None)

def _repr_args(self):
"""Return str represent key arguments for TokenizerTransform."""
Expand Down Expand Up @@ -169,7 +171,10 @@ def _tokenize(self, tokens, side='src', is_train=False):
def apply(self, example, is_train=False, stats=None, **kwargs):
"""Apply sentencepiece subword encode to src & tgt."""
src_out = self._tokenize(example['src'], 'src', is_train)
tgt_out = self._tokenize(example['tgt'], 'tgt', is_train)
if self.model_task == ModelTask.LANGUAGE_MODEL:
tgt_out = src_out
else:
tgt_out = self._tokenize(example['tgt'], 'tgt', is_train)
if stats is not None:
n_words = len(example['src']) + len(example['tgt'])
n_subwords = len(src_out) + len(tgt_out)
Expand Down Expand Up @@ -243,7 +248,10 @@ def _tokenize(self, tokens, side='src', is_train=False):
def apply(self, example, is_train=False, stats=None, **kwargs):
"""Apply bpe subword encode to src & tgt."""
src_out = self._tokenize(example['src'], 'src', is_train)
tgt_out = self._tokenize(example['tgt'], 'tgt', is_train)
if self.model_task == ModelTask.LANGUAGE_MODEL:
tgt_out = src_out
else:
tgt_out = self._tokenize(example['tgt'], 'tgt', is_train)
if stats is not None:
n_words = len(example['src']) + len(example['tgt'])
n_subwords = len(src_out) + len(tgt_out)
Expand Down Expand Up @@ -327,7 +335,7 @@ def get_specials(cls, opts):
tgt_specials.update(_case_specials)
return (set(), set())

def _get_subword_kwargs(self, side='src'):
def _get_subword_kwargs(self, side='src', is_train=False):
"""Return a dict containing kwargs relate to `side` subwords."""
subword_type = self.tgt_subword_type if side == 'tgt' \
else self.src_subword_type
Expand All @@ -338,6 +346,10 @@ def _get_subword_kwargs(self, side='src'):
subword_alpha = self.tgt_subword_alpha if side == 'tgt' \
else self.src_subword_alpha
kwopts = dict()
if not is_train:
# disable random aspects during validation
subword_alpha = 0
subword_nbest = 1
if subword_type == 'bpe':
kwopts['bpe_model_path'] = subword_model
kwopts['bpe_dropout'] = subword_alpha
Expand All @@ -360,42 +372,65 @@ def warm_up(self, vocabs=None):
"""Initialize Tokenizer models."""
super().warm_up(None)
import pyonmttok
src_subword_kwargs = self._get_subword_kwargs(side='src')

src_subword_kwargs = self._get_subword_kwargs(
side="src", is_train=True
)
valid_src_subword_kwargs = self._get_subword_kwargs(
side="src", is_train=False
)
src_tokenizer = pyonmttok.Tokenizer(
**src_subword_kwargs, **self.src_other_kwargs
)
tgt_subword_kwargs = self._get_subword_kwargs(side='tgt')
_diff_vocab = (
src_subword_kwargs.get('vocabulary_path', '') !=
tgt_subword_kwargs.get('vocabulary_path', '') or
src_subword_kwargs.get('vocabulary_threshold', 0) !=
tgt_subword_kwargs.get('vocabulary_threshold', 0))
Comment on lines -368 to -372
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I personally prefer current one which seems more readable

valid_src_tokenizer = pyonmttok.Tokenizer(
**valid_src_subword_kwargs, **self.src_other_kwargs
)
tgt_subword_kwargs = self._get_subword_kwargs(
side="tgt", is_train=True
)
_diff_vocab = src_subword_kwargs.get(
"vocabulary_path", ""
) != tgt_subword_kwargs.get(
"vocabulary_path", ""
) or src_subword_kwargs.get(
"vocabulary_threshold", 0
) != tgt_subword_kwargs.get(
"vocabulary_threshold", 0
)
if self.share_vocab and not _diff_vocab:
self.load_models = {
'src': src_tokenizer,
'tgt': src_tokenizer
"src": {"train": src_tokenizer, "valid": valid_src_tokenizer},
"tgt": {"train": src_tokenizer, "valid": valid_src_tokenizer},
}
else:
tgt_subword_kwargs = self._get_subword_kwargs(side='tgt')
tgt_tokenizer = pyonmttok.Tokenizer(
**tgt_subword_kwargs, **self.tgt_other_kwargs
)
valid_tgt_subword_kwargs = self._get_subword_kwargs(
side="tgt", is_train=False
)
valid_tgt_tokenizer = pyonmttok.Tokenizer(
**valid_tgt_subword_kwargs, **self.tgt_other_kwargs
)
self.load_models = {
'src': src_tokenizer,
'tgt': tgt_tokenizer
"src": {"train": src_tokenizer, "valid": valid_src_tokenizer},
"tgt": {"train": tgt_tokenizer, "valid": valid_tgt_tokenizer},
}

def _tokenize(self, tokens, side='src', is_train=False):
"""Do OpenNMT Tokenizer's tokenize."""
tokenizer = self.load_models[side]
tokenizer = self.load_models[side]['train' if is_train else 'valid']
sentence = ' '.join(tokens)
segmented, _ = tokenizer.tokenize(sentence)
return segmented

def apply(self, example, is_train=False, stats=None, **kwargs):
"""Apply OpenNMT Tokenizer to src & tgt."""
src_out = self._tokenize(example['src'], 'src')
tgt_out = self._tokenize(example['tgt'], 'tgt')
src_out = self._tokenize(example['src'], 'src', is_train)
if self.model_task == ModelTask.LANGUAGE_MODEL:
tgt_out = src_out
else:
tgt_out = self._tokenize(example['tgt'], 'tgt', is_train)
if stats is not None:
n_words = len(example['src']) + len(example['tgt'])
n_subwords = len(src_out) + len(tgt_out)
Expand Down