def test_encode_decode(self):
corpus = (
'This is a corpus of text that provides a bunch of tokens from which '
'to build a vocabulary. It will be used when strings are encoded '
'with a SubwordTextTokenizer subclass. The encoder was coded by a coder.'
)
alphabet = set(corpus) ^ {' '}
original = 'This is a coded sentence encoded by the SubwordTextTokenizer.'
encoder = SubwordTextTokenizer.build_to_target_size_from_corpus(
[corpus, original], target_size=100, min_val=2, max_val=10)
# Encoding should be reversible.
encoded = encoder.encode(original)
decoded = encoder.decode(encoded)
self.assertEqual(original, decoded)
# The substrings coded and coder are frequent enough in the corpus that
# they should appear in the vocabulary even though they are substrings
# of other included strings.
subtoken_strings = encoded
self.assertIn('encoded_', subtoken_strings)
self.assertIn('coded_', subtoken_strings)
self.assertIn('SubwordTextTokenizer_', encoder._all_subtoken_strings)
self.assertIn('coder_', encoder._all_subtoken_strings)
# Every character in the corpus should be in the encoder's alphabet and
# its subtoken vocabulary.
self.assertTrue(alphabet.issubset(encoder._alphabet))
for a in alphabet:
self.assertIn(a, encoder._all_subtoken_strings)
def test_unicode(self):
corpus = 'Cat emoticons. \U0001F638 \U0001F639 \U0001F63A \U0001F63B'
token_counts = collections.Counter(corpus.split(' '))
encoder = SubwordTextTokenizer.build_to_target_size_from_token_counts(
100, token_counts, 2, 10)
self.assertIn('\U0001F638', encoder._alphabet)
self.assertIn('\U0001F63B', encoder._all_subtoken_strings)
def __init__(self,
sample,
append_sos=False,
append_eos=False,
target_vocab_size=None,
min_occurrences=1,
max_occurrences=1e3,
reserved_tokens=DEFAULT_RESERVED_TOKENS,
sos_index=DEFAULT_SOS_INDEX,
eos_index=DEFAULT_EOS_INDEX,
unknown_index=DEFAULT_UNKNOWN_INDEX,
padding_index=DEFAULT_PADDING_INDEX,
**kwargs):
super().__init__(**kwargs)
self.append_sos = append_sos
self.append_eos = append_eos
self.sos_index = sos_index
self.eos_index = eos_index
self.unknown_index = unknown_index
self.reserved_tokens = reserved_tokens
self.padding_index = padding_index
if target_vocab_size is None:
self.tokenizer = SubwordTextTokenizer()
self.tokenizer.build_from_corpus(sample, min_count=min_occurrences)
else:
target_vocab_size -= len(reserved_tokens)
self.tokenizer = SubwordTextTokenizer.build_to_target_size_from_corpus(
sample,
target_size=target_vocab_size,
min_val=min_occurrences,
max_val=max_occurrences)
self.index_to_token = reserved_tokens.copy()
self.token_to_index = {token: index for index, token in enumerate(reserved_tokens)}
for token in self.tokenizer.vocab:
self.index_to_token.append(token)
self.token_to_index[token] = len(self.index_to_token) - 1
def test_raises_exception_when_not_encodable(self):
corpus = 'the quick brown fox jumps over the lazy dog'
token_counts = collections.Counter(corpus.split(' '))
# Deliberately exclude some required encoding chars from the alphabet
# and token list, making some strings unencodable.
encoder = SubwordTextTokenizer.build_to_target_size_from_token_counts(
100, token_counts, 2, 10)
original = 'This has UPPER CASE letters that are out of alphabet'
# Previously there was a bug which produced an infinite loop in this case.
with self.assertRaises(AssertionError):
encoder.encode(original)
def test_small_vocab(self):
corpus = 'The quick brown fox jumps over the lazy dog'
token_counts = collections.Counter(corpus.split(' '))
alphabet = set(corpus) ^ {' '}
encoder = SubwordTextTokenizer.build_to_target_size_from_token_counts(
10, token_counts, 2, 10)
# All vocabulary elements are in the alphabet and subtoken strings even
# if we requested a smaller vocabulary to assure all expected strings
# are encodable.
self.assertTrue(alphabet.issubset(encoder._alphabet))
for a in alphabet:
self.assertIn(a, encoder._all_subtoken_strings)
def test_encodable_when_not_in_alphabet(self):
corpus = 'the quick brown fox jumps over the lazy dog'
token_counts = collections.Counter(corpus.split(' '))
encoder = SubwordTextTokenizer.build_to_target_size_from_token_counts(
100, token_counts, 2, 10)
original = 'This has UPPER CASE letters that are out of alphabet'
# Early versions could have an infinite loop when breaking into subtokens
# if there was any out-of-alphabet characters in the encoded string.
encoded = encoder.encode(original)
decoded = encoder.decode(encoded)
self.assertEqual(original, decoded)
encoded_str = ''.join(encoded)
self.assertIn('\\84;', encoded_str)
def test_token_counts(self):
token_counts = SubwordTextTokenizer._count_tokens(self.corpus)
expected = {
u"'": 2,
u".": 2,
u". ": 1,
u"... ": 1,
u"Groucho": 1,
u"Marx": 1,
u"Mitch": 1,
u"Hedberg": 1,
u"I": 3,
u"in": 2,
u"my": 2,
u"pajamas": 2,
}
self.assertDictContainsSubset(expected, token_counts)
class SubwordEncoder(TextEncoder):
""" Invertibly encoding text using a limited vocabulary.
Applies Googles Tensor2Tensor ``SubwordTextTokenizer`` that invertibly encodes a native string
as a
sequence of subtokens from a limited vocabulary. In order to build the vocabulary, it uses
recursive binary search to find a minimum token count `x`
(s.t. `min_occurrences` <= `x` <= `max_occurrences`) that most closely matches the
`target_size`.
**Tokenizer Reference:**
https://github.com/tensorflow/tensor2tensor/blob/8bdecbe434d93cb1e79c0489df20fee2d5a37dc2/tensor2tensor/data_generators/text_encoder.py#L389
Args:
sample (list): Sample of data used to build encoding dictionary.
append_eos (bool, optional): If ``True`` append EOS token onto the end to the encoded
vector.
target_vocab_size (int, optional): Desired size of vocab.
min_occurrences (int, optional): Lower bound for the minimum token count.
max_occurrences (int, optional): Upper bound for the minimum token count.
reserved_tokens (list of str, optional): List of reserved tokens inserted in the beginning
of the dictionary.
eos_index (int, optional): The eos token is used to encode the end of a sequence. This is
the index that token resides at.
unknown_index (int, optional): The unknown token is used to encode unseen tokens. This is
the index that token resides at.
padding_index (int, optional): The padding token is used to encode sequence padding. This is
the index that token resides at.
**kwargs: Keyword arguments passed onto ``TextEncoder.__init__``.
"""
def __init__(self,
sample,
append_eos=False,
target_vocab_size=None,
min_occurrences=1,
max_occurrences=1e3,
reserved_tokens=DEFAULT_RESERVED_TOKENS,
eos_index=DEFAULT_EOS_INDEX,
unknown_index=DEFAULT_UNKNOWN_INDEX,
padding_index=DEFAULT_PADDING_INDEX,
**kwargs):
super().__init__(**kwargs)
self.append_eos = append_eos
self.eos_index = eos_index
self.unknown_index = unknown_index
self.reserved_tokens = reserved_tokens
self.padding_index = padding_index
if target_vocab_size is None:
self.tokenizer = SubwordTextTokenizer()
self.tokenizer.build_from_corpus(sample, min_count=min_occurrences)
else:
target_vocab_size -= len(reserved_tokens)
self.tokenizer = SubwordTextTokenizer.build_to_target_size_from_corpus(
sample,
target_size=target_vocab_size,
min_val=min_occurrences,
max_val=max_occurrences)
self.itos = reserved_tokens.copy()
self.stoi = {
token: index
for index, token in enumerate(reserved_tokens)
}
for token in self.tokenizer.vocab:
self.itos.append(token)
self.stoi[token] = len(self.itos) - 1
@property
def vocab(self):
"""
Returns:
list: List of tokens in the dictionary.
"""
return self.itos
@property
def vocab_size(self):
"""
Returns:
int: Number of tokens in the dictionary.
"""
return len(self.vocab)
def encode(self, sequence):
""" Encodes a ``sequence``.
Args:
sequence (str): String ``sequence`` to encode.
Returns:
torch.Tensor: Encoding of the ``sequence``.
"""
sequence = super().encode(sequence)
sequence = self.tokenizer.encode(sequence)
vector = [
self.stoi.get(token, self.unknown_index) for token in sequence
]
if self.append_eos:
vector.append(self.eos_index)
return torch.tensor(vector)
def decode(self, encoded):
""" Decodes a tensor into a sequence.
Args:
encoded (torch.Tensor): Encoded sequence.
Returns:
str: Sequence decoded from ``encoded``.
"""
encoded = super().decode(encoded)
return self.tokenizer.decode([self.itos[index] for index in encoded])
def test_is_pickleable():
tokenizer = SubwordTextTokenizer()
pickle.dumps(tokenizer)