def get_recurrent_tokenizer(vocab,
max_context_tokens,
unk_token,
pad_token,
device="cpu"):
"""
Return a tokenizer to be used with recurrent-based models
"""
question_tokenizer = Tokenizer(WordLevel(vocab, unk_token=unk_token))
question_tokenizer.normalizer = Sequence(
[StripAccents(), Lowercase(), Strip()])
question_tokenizer.pre_tokenizer = PreSequence(
[Whitespace(), Punctuation()])
question_tokenizer.enable_padding(direction="right",
pad_id=vocab[pad_token],
pad_type_id=1,
pad_token=pad_token)
context_tokenizer = Tokenizer(WordLevel(vocab, unk_token=unk_token))
context_tokenizer.normalizer = Sequence(
[StripAccents(), Lowercase(), Strip()])
context_tokenizer.pre_tokenizer = PreSequence(
[Whitespace(), Punctuation()])
context_tokenizer.enable_padding(
direction="right",
pad_id=vocab[pad_token],
pad_type_id=1,
pad_token=pad_token,
)
context_tokenizer.enable_truncation(max_context_tokens)
return RecurrentSquadTokenizer(question_tokenizer,
context_tokenizer,
device=device)
def test_truncation(self):
tokenizer = Tokenizer(BPE())
tokenizer.add_tokens(["my", "name", "is", "john", "pair"])
tokenizer.enable_truncation(2)
# Can truncate single sequences
output = tokenizer.encode("my name is john")
assert output.tokens == ["my", "name"]
# Can truncate pair sequences as well
output = tokenizer.encode("my name is john", "pair")
assert output.tokens == ["my", "pair"]
def wordpiece_tokenize(line):
tokenizer = Tokenizer(WordPiece(wordpiece_dict3))
tokenizer.enable_padding(length=200)
tokenizer.enable_truncation(max_length=200)
tokenizer.pre_tokenizer = WhitespaceSplit()
tokenizer.post_processor = TemplateProcessing(
single="[CLS] $A [SEP]",
pair="[CLS] $A [SEP] $B:1 [SEP]:1",
special_tokens=[
("[CLS]", 1),
("[SEP]", 2),
],
)
output = tokenizer.encode(line)
return (output.ids)
def test_post_process(self):
tokenizer = Tokenizer(BPE())
tokenizer.add_tokens(["my", "name", "is", "john", "pair"])
tokenizer.enable_truncation(2)
tokenizer.enable_padding(length=4)
encoding = tokenizer.encode("my name is john")
pair_encoding = tokenizer.encode("pair")
# Can post process a single encoding
output = tokenizer.post_process(encoding)
assert output.tokens == ["my", "name", "[PAD]", "[PAD]"]
# Can post process a pair of encodings
output = tokenizer.post_process(encoding, pair_encoding)
assert output.tokens == ["my", "pair", "[PAD]", "[PAD]"]
def __init__(
self,
inputs: List[str],
targets: List[Tensor],
key: Tensor,
tokenizer: Tokenizer,
max_len: int = 200,
batch_size: int = 32,
) -> None:
self.inputs = inputs
self.targets = np.array(onp.array(targets))
tokenizer.enable_truncation(max_len)
tokenizer.enable_padding()
self.tokenizer = tokenizer
self.max_len = max_len
self.batch_size = batch_size
self.key = key
self.len = len(self.inputs)
self.num_batches = math.ceil(self.len / batch_size)
def __init__(
self,
vocab_file,
sep_token="<sep>",
cls_token="<cls>",
pad_token="<pad>",
mask_token="<mask>",
lowercase: bool = True,
):
tokenizer = Tokenizer(WordLevel(vocab_file, unk_token=unk_token))
tokenizer.normalizer = Strip()
tokenizer.pre_tokenizer = CharDelimiterSplit(" ")
tokenizer.post_processor = BertProcessing(
("</s>", tokenizer.token_to_id("</s>")),
("<s>", tokenizer.token_to_id("<s>")),
)
tokenizer.enable_truncation(max_length=512)
# Let the tokenizer know about special tokens if they are part of the vocab
if tokenizer.token_to_id(str(unk_token)) is not None:
tokenizer.add_special_tokens([str(unk_token)])
if tokenizer.token_to_id(str(sep_token)) is not None:
tokenizer.add_special_tokens([str(sep_token)])
if tokenizer.token_to_id(str(cls_token)) is not None:
tokenizer.add_special_tokens([str(cls_token)])
if tokenizer.token_to_id(str(pad_token)) is not None:
tokenizer.add_special_tokens([str(pad_token)])
if tokenizer.token_to_id(str(mask_token)) is not None:
tokenizer.add_special_tokens([str(mask_token)])
parameters = {
"model": "WordLevel",
"unk_token": unk_token,
"sep_token": sep_token,
"cls_token": cls_token,
"pad_token": pad_token,
"mask_token": mask_token,
"lowercase": lowercase,
}
super().__init__(tokenizer, parameters)
def main(args):
if args.do_train:
# Initialize a tokenizer
files = get_smi_files(args.training_files)
print("Training BPE tokenizer using the following files:{}".format(
files))
tokenizer = Tokenizer(models.BPE(unk_token="<unk>"))
tokenizer.enable_padding(pad_id=args.vocab_size + 2,
pad_token="<pad>",
length=args.pad_len)
tokenizer.enable_truncation(max_length=args.pad_len,
strategy='only_first')
tokenizer.normalizer = Sequence([NFKC()])
tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(
add_prefix_space=False)
tokenizer.decoder = decoders.ByteLevel()
tokenizer.post_processor = processors.ByteLevel(trim_offsets=True)
# Train the tokenizer
trainer = trainers.BpeTrainer(show_progress=True,
vocab_size=args.vocab_size,
min_frequency=args.min_frequency)
tokenizer.train(files, trainer=trainer)
tokenizer.add_tokens(["<start>", "<end>"])
tokenizer.save(os.path.join('tokenizers', args.tokenizer_name),
pretty=True)
print("Trained vocab size: {}".format(tokenizer.get_vocab_size()))
if args.do_test:
# Test the tokenizer
tokenizer = Tokenizer.from_file(
os.path.join('tokenizers', args.tokenizer_name))
print("Testing with SMILES String: {}".format(args.test_string))
encoding = tokenizer.encode(args.test_string)
print("Encoded string: {}".format(encoding.tokens))
print(encoding.ids)
decoded = tokenizer.decode(encoding.ids)
print("Decoded string: {}".format(decoded))
def test_truncation(self):
tokenizer = Tokenizer(BPE())
tokenizer.add_tokens(["my", "name", "is", "john", "pair"])
tokenizer.enable_truncation(2)
# Can truncate single sequences
output = tokenizer.encode("my name is john")
assert output.tokens == ["my", "name"]
# Can truncate pair sequences as well
output = tokenizer.encode("my name is john", "pair")
assert output.tokens == ["my", "pair"]
# Can get the params and give them to enable_truncation
trunc = tokenizer.truncation
tokenizer.enable_truncation(**trunc)
# Left truncation direction
tokenizer.enable_truncation(2, direction="left")
output = tokenizer.encode("my name is john")
assert output.tokens == ["is", "john"]
output = tokenizer.encode("my name is john", "pair")
assert output.tokens == ["john", "pair"]
def preprocess_data(args):
label_counter = Counter([])
examples_per_file = Counter()
print("Reading all files for labels.")
for input_file in args.input_files:
with xopen(input_file, "rt") as f:
for example, labels in input_readers[args.task](f):
examples_per_file[input_file] += 1
label_counter.update(labels)
if args.top_n_labels > 0:
mlb_full = MultiLabelBinarizer(sparse_output=True)
mlb_full = mlb_full.fit(label_counter.keys())
label_counter = dict(label_counter.most_common(args.top_n_labels))
mlb = MultiLabelBinarizer(sparse_output=True)
# Passing a list in a list because that's what the function wants.
mlb = mlb.fit([[pair for pair in label_counter]])
# Save list of partial -> full mapping if doing top N labels.
if args.top_n_labels > 0:
label_mapping = np.where(np.in1d(mlb_full.classes_,
mlb.classes_))[0].tolist()
with xopen(args.label_mapping, "wt") as f:
f.write(json.dumps(label_mapping))
# Also save the full labels.
with xopen(args.full_labels, "wt") as f:
f.write(json.dumps(list(mlb_full.classes_)))
# Save list of labels.
with xopen(args.labels_out, "wt") as f:
f.write(json.dumps(list(mlb.classes_)))
# Set parallel tokenization thread count.
os.environ["RAYON_NUM_THREADS"] = str(args.processes)
from tokenizers import Tokenizer, decoders, trainers
from tokenizers.models import WordPiece
from tokenizers.normalizers import BertNormalizer
from tokenizers.pre_tokenizers import BertPreTokenizer
from tokenizers.processors import BertProcessing
if args.task == 'cafa':
# Define our custom tokenizer.
# It is exactly the same as the default BERT tokenizer, except for max_input_chars_per_word
# being 20000 instead of 100. This tokenizer is very slow on the long protein sequences.
tokenizer = WordPiece.from_files(args.vocab,
unk_token="[UNK]",
max_input_chars_per_word=20000)
tokenizer = Tokenizer(tokenizer)
tokenizer.add_special_tokens(["[UNK]", "[SEP]", "[CLS]"])
tokenizer.normalizer = BertNormalizer(lowercase=args.do_lower_case)
tokenizer.pre_tokenizer = BertPreTokenizer()
tokenizer.post_processor = BertProcessing(
("[SEP]", tokenizer.token_to_id("[SEP]")),
("[CLS]", tokenizer.token_to_id("[CLS]")))
tokenizer.decoder = decoders.WordPiece(prefix='##')
else:
tokenizer = BertWordPieceTokenizer(args.vocab,
lowercase=args.do_lower_case)
tokenizer.enable_padding(max_length=args.seq_len)
tokenizer.enable_truncation(max_length=args.seq_len)
for input_file in args.input_files:
with xopen(input_file, 'rt') as in_f:
file_name = generate_out_filename(input_file, args)
with xopen(file_name, "wt") as out_f:
print("Processing to: ", file_name)
# Write the shape as the first row, useful for the finetuning.
out_f.write(
json.dumps((examples_per_file[input_file],
len(label_counter))) + '\n')
batch_size = min(examples_per_file[input_file],
args.processes * 100)
example_batch = []
labels_batch = []
with ParallelGenerator(input_readers[args.task](in_f),
max_lookahead=batch_size) as g:
for example, labels in g:
example_batch.append(example)
labels_batch.append(labels)
if len(example_batch) == batch_size:
example_batch = tokenizer.encode_batch(
example_batch)
labels_batch = mlb.transform(labels_batch)
for example, labels in zip(example_batch,
labels_batch):
# Convert sparse arrays to python lists for json dumping.
# print(labels);input()
labels = labels.nonzero()[1].tolist()
out_f.write(
json.dumps([example.ids, labels]) + '\n')
example_batch = []
labels_batch = []
# Write out whatever is left in the last smaller batch.
example_batch = tokenizer.encode_batch(example_batch)
labels_batch = mlb.transform(labels_batch)
for example, labels in zip(example_batch, labels_batch):
# Convert sparse arrays to python lists for json dumping.
# print(labels);input()
labels = labels.nonzero()[1].tolist()
out_f.write(json.dumps([example.ids, labels]) + '\n')
class SentencePieceBPETokenizer:
"""Custom SentencePiece tokenizer"""
unk_token = '<unk>'
pad_token = '<pad>'
def __init__(self,
vocab: Dict[str, int] = None,
merges: List[Tuple[str, str]] = None,
dropout: float = None,
max_length: Optional[int] = 64) -> None:
"""Constructor
Args:
vocab (Dict[str, int]): A dictionary of string keys and their ids.
merges (List[Tuple[str, str]]): A list of pairs of tokens.
dropout (float): BPE dropout
max_length (int, optional): The max length at which to truncate.
Defaults to `64`.
"""
self.tokenizer = Tokenizer(
BPE(vocab, merges, dropout=dropout, unk_token=self.unk_token))
self.tokenizer.normalizer = BertNormalizer() # noqa
self.tokenizer.pre_tokenizer = pre_tokenizers.Metaspace() # noqa
self.tokenizer.decoder = decoders.Metaspace() # noqa
self.tokenizer.add_special_tokens([self.pad_token, self.unk_token])
self.tokenizer.enable_padding(pad_token=self.pad_token)
self.tokenizer.enable_truncation(max_length)
@classmethod
def train(cls,
dataset: Sequence[str],
vocab_size: int = 1000,
min_frequency: int = 2,
dropout: float = 0.0,
max_length: Optional[int] = 64) -> 'SentencePieceBPETokenizer':
instance = cls(dropout=dropout, max_length=max_length)
trainer = trainers.BpeTrainer(
vocab_size=vocab_size,
min_frequency=min_frequency,
special_tokens=[cls.pad_token, cls.unk_token])
instance.tokenizer.train_from_iterator(dataset, trainer=trainer)
instance.tokenizer.model.dropout = None
return instance
@property
def vocab_size(self):
return len(self.tokenizer.get_vocab())
def serialize(self):
return self.tokenizer.to_str()
@classmethod
def deserialize(cls, s: str) -> 'SentencePieceBPETokenizer':
tokenizer = cls()
tokenizer.tokenizer = Tokenizer.from_str(s)
return tokenizer
def encode(self, text: str) -> Dict[str, Any]:
encoding = self.tokenizer.encode(text)
outputs = {
'ids': torch.tensor(encoding.ids),
'mask': torch.tensor(encoding.attention_mask),
'spans': encoding.offsets,
}
return outputs
def encode_batch(self, batch: List[str]):
encodings = self.tokenizer.encode_batch(batch)
outputs = {
'ids': torch.tensor([e.ids for e in encodings]),
'mask': torch.tensor([e.attention_mask for e in encodings]),
'spans': [e.offsets for e in encodings],
}
return outputs
class LitTokenizer:
def __init__(self,
padding=False,
truncation=False,
max_length=None,
lower=False,
lang=None):
super().__init__()
self.UNK_WORD = '[UNK]'
self.PAD_WORD = '[PAD]'
self.MASK_WORD = '[MASK]'
self.SOS_WORD = '[SOS]'
self.EOS_WORD = '[EOS]'
self.special_tokens = [
self.UNK_WORD, self.PAD_WORD, self.MASK_WORD, self.SOS_WORD,
self.EOS_WORD
]
# Define tokenizer
self.tokenizer = None
self.configure_tokenizers(padding, truncation, max_length, lower)
# Other
self.lang = lang
def get_vocab_size(self):
return self.tokenizer.get_vocab_size()
def configure_tokenizers(self, padding, truncation, max_length, lower):
# Settings
pad_length = None
if padding in {True, "longest"}:
pass
elif padding in {"max_length"}:
pad_length = max_length
elif padding in {False, "do_not_pad"}:
pass
else:
raise ValueError("Unknown padding type")
# SRC tokenizer
tok_normalizers = [NFD(), Strip()]
if lower:
tok_normalizers += [Lowercase()]
self.tokenizer = Tokenizer(tok_model()) # unk_token=... not working
self.tokenizer.add_special_tokens(self.special_tokens)
self.tokenizer.pre_tokenizer = pre_tokenizers.Sequence(
[WhitespaceSplit()])
self.tokenizer.normalizer = normalizers.Sequence(
tok_normalizers) # StripAccents requires NFD
self.tokenizer.decoder = tok_decoder()
# Define template (Needed for the sos/eos tokens)
basic_template = TemplateProcessing(
single=f"{self.SOS_WORD} $A {self.EOS_WORD}",
pair=
f"{self.SOS_WORD} $A {self.EOS_WORD} {self.SOS_WORD} $B {self.EOS_WORD}",
special_tokens=[
(self.SOS_WORD, self.tokenizer.token_to_id(self.SOS_WORD)),
(self.EOS_WORD, self.tokenizer.token_to_id(self.EOS_WORD))
],
)
self.tokenizer.post_processor = basic_template
if padding:
self.tokenizer.enable_padding(pad_id=self.tokenizer.token_to_id(
self.PAD_WORD),
pad_token=self.PAD_WORD,
length=pad_length)
if truncation:
self.tokenizer.enable_truncation(max_length,
stride=0,
strategy='longest_first')
def load_vocab(self, vocab, merges):
vocab, merges = tok_model.read_file(vocab, merges)
self.tokenizer.model = tok_model(vocab, merges)
def train_vocab(self, files, vocab_size=32000, min_frequency=3):
# Train trainer
trainer = tok_trainer(vocab_size=vocab_size,
min_frequency=min_frequency)
self.tokenizer.train(files, trainer)
def save_vocab(self, output_dir, prefix):
self.tokenizer.model.save(output_dir, prefix)
def pad(self, examples, keys=None):
pad_idx = self.special_tokens.index(self.PAD_WORD)
# Keys to modify
if not keys:
keys = list(examples[0].keys())
d = {}
for k in keys:
# Collect same-type items (list of IDs, list of masks,...)
d[k] = [x[k] for x in examples]
# Get max length (value to pad)
max_length = max([x.shape[-1] for x in d[k]])
# Apply padding
for i, x in enumerate(examples):
unpadded_t = x[k]
if k == "ids":
tmp = torch.full((max_length, ),
fill_value=pad_idx,
device=unpadded_t.device) # All padding
elif k == "attention_mask":
tmp = torch.full(
(max_length, ), fill_value=0,
device=unpadded_t.device) # No attention mask
else:
raise TypeError("Unknown key")
tmp[:unpadded_t.shape[-1]] = unpadded_t
d[k][i] = tmp
return d
def encode(self, x):
return self.tokenizer.encode(x)
def decode(self, x):
if isinstance(x, torch.Tensor):
assert len(x.shape) == 2
x = x.detach().cpu().numpy()
return [self.tokenizer.decode(x_i) for x_i in x]
def preprocess_data(args):
label_counter = Counter([])
examples_per_file = Counter()
print("Reading all files for labels.")
for input_file in args.input_files:
with xopen(input_file, "rt") as f:
for example, labels in input_readers[args.task](f):
examples_per_file[input_file] += 1
label_counter.update(labels)
if args.top_n_labels > 0:
mlb_full = MultiLabelBinarizer(sparse_output=True)
mlb_full = mlb_full.fit(label_counter.keys())
label_counter = dict(label_counter.most_common(args.top_n_labels))
mlb = MultiLabelBinarizer(sparse_output=True)
# Passing a list in a list because that's what the function wants.
if args.labels_in:
labels = json.load(open(args.labels_in))
mlb = mlb.fit([labels])
else:
mlb = mlb.fit([[pair for pair in label_counter]])
# Save list of partial -> full mapping if doing top N labels.
if args.top_n_labels > 0:
label_mapping = np.where(np.in1d(mlb_full.classes_,
mlb.classes_))[0].tolist()
with xopen(args.label_mapping, "wt") as f:
f.write(json.dumps(label_mapping))
# Also save the full labels.
with xopen(args.full_labels, "wt") as f:
f.write(json.dumps(list(mlb_full.classes_)))
# Save list of labels.
with xopen(args.labels_out, "wt") as f:
f.write(json.dumps(list(mlb.classes_)))
# Set parallel tokenization thread count.
os.environ["RAYON_NUM_THREADS"] = str(args.processes)
from tokenizers import Tokenizer, decoders, trainers
from tokenizers.models import WordPiece
from tokenizers.normalizers import BertNormalizer
from tokenizers.pre_tokenizers import BertPreTokenizer
from tokenizers.processors import BertProcessing
if args.task == 'cafa':
# Define our custom tokenizer.
# It is exactly the same as the default BERT tokenizer, except for max_input_chars_per_word
# being 20000 instead of 100. This tokenizer is very slow on the long protein sequences.
tokenizer = WordPiece.from_files(args.vocab,
unk_token="[UNK]",
max_input_chars_per_word=20000)
tokenizer = Tokenizer(tokenizer)
tokenizer.add_special_tokens(["[UNK]", "[SEP]", "[CLS]"])
tokenizer.normalizer = BertNormalizer(lowercase=args.do_lower_case)
tokenizer.pre_tokenizer = BertPreTokenizer()
tokenizer.post_processor = BertProcessing(
("[SEP]", tokenizer.token_to_id("[SEP]")),
("[CLS]", tokenizer.token_to_id("[CLS]")))
tokenizer.decoder = decoders.WordPiece(prefix='##')
else:
tokenizer = BertWordPieceTokenizer(args.vocab,
lowercase=args.do_lower_case)
tokenizer.enable_padding(max_length=args.seq_len)
tokenizer.enable_truncation(max_length=args.seq_len)
for input_file in args.input_files:
with xopen(input_file, 'rt') as in_f:
file_name = generate_out_filename(input_file, args)
with xopen(file_name, "wt") as out_f:
print("Processing to: ", file_name)
# Write the shape as the first row, useful for the finetuning.
if args.labels_in:
n_labels = len(json.load(open(args.labels_in)))
else:
n_labels = len(label_counter)
out_f.write(
json.dumps((examples_per_file[input_file], n_labels)) +
'\n')
batch_size = min(examples_per_file[input_file],
args.processes * 100)
example_batch = []
labels_batch = []
doc_idx_batch = []
with ParallelGenerator(input_readers[args.task](in_f),
max_lookahead=batch_size) as g:
START_POS = int(args.window_start) / 100
for doc_idx, (example, labels) in enumerate(g):
#example = ' '.join(example.split(' ')[-510:])
example_batch.append(example)
labels_batch.append(labels)
doc_idx_batch.append(doc_idx)
if len(example_batch) == batch_size:
example_batch = tokenizer.encode_batch(
example_batch)
labels_batch = mlb.transform(labels_batch)
for example, labels, doc_idx in zip(
example_batch, labels_batch,
doc_idx_batch):
# Convert sparse arrays to python lists for json dumping.
# print(labels);input()
labels = labels.nonzero()[1].tolist()
"""try:
[][0]
print("DOC_LEN:",len(example.overflowing)+1)
mid = len(example.overflowing)//2
out_f.write(json.dumps( [example.overflowing[mid].ids, labels, len(example.overflowing)+1] ) + '\n')
except IndexError:
out_f.write(json.dumps( [example.ids, labels, len(example.overflowing)+1] ) + '\n')"""
if args.all_blocks or args.n_blocks > 0:
blocks = [example.ids] + [
blk.ids for blk in example.overflowing
]
#print("BLOCKS:%d,TOKENS:%d" % (len(list(blocks)), sum([len(list(tokens)) for tokens in blocks])))
for b, block in enumerate(blocks, 2):
if b > args.n_blocks and args.n_blocks > 0:
break
out_f.write(
json.dumps(
[block, labels, doc_idx]) +
'\n')
else:
window = get_window(example, START_POS)
assert len(window) == 512
assert all(
[type(y) is int for y in window])
out_f.write(
json.dumps([window, labels]) + '\n')
example_batch = []
labels_batch = []
# Write out whatever is left in the last smaller batch.
example_batch = tokenizer.encode_batch(example_batch)
labels_batch = mlb.transform(labels_batch)
for example, labels, doc_idx in zip(
example_batch, labels_batch, doc_idx_batch):
# Convert sparse arrays to python lists for json dumping.
# print(labels);input()
labels = labels.nonzero()[1].tolist()
"""try:
[][0]
print("DOC_LEN:",len(example.overflowing)+1)
mid = len(example.overflowing)//2
out_f.write(json.dumps( [example.overflowing[mid].ids, labels, len(example.overflowing)+1] ) + '\n')
except IndexError:
out_f.write(json.dumps( [example.ids, labels, len(example.overflowing)+1] ) + '\n')"""
if args.all_blocks or args.n_blocks > 0:
blocks = [example.ids] + [
blk.ids for blk in example.overflowing
]
#print("BLOCKS:%d,TOKENS:%d" % (len(list(blocks)), sum([len(list(tokens)) for tokens in blocks])))
for b, block in enumerate(blocks, 2):
if b > args.n_blocks and args.n_blocks > 0:
break
out_f.write(
json.dumps([block, labels, doc_idx]) +
'\n')
else:
out_f.write(
json.dumps(
[get_window(example, START_POS), labels]) +
'\n')