def tokenize(self, texts):
if self.verbose:
print("Tokenizing {} documents...".format(len(texts)))
tokens = get_texts(texts)
texts_length = [len(t) for t in tokens]
try:
self.vocab
if self.verbose:
print("Using existing vocabulary")
except AttributeError:
if self.verbose:
print("Building Vocab...")
self.vocab = Vocab.create(tokens,
max_vocab=self.max_vocab,
min_freq=self.min_freq)
texts_numz = [self.vocab.numericalize(t) for t in texts]
sorted_texts_length = sorted(texts_length)
self.maxlen = int(np.quantile(sorted_texts_length, q=self.q))
# self.maxlen = sorted_texts_length[int(self.q * len(sorted_texts_length))]
if self.verbose:
print("Padding documents...")
padded_texts = [
pad_sequences(t,
self.maxlen,
pad_first=self.pad_first,
pad_idx=self.pad_idx) for t in texts_numz
]
return np.stack(padded_texts, axis=0)
def test_simple_vocab():
input_vocab = Vocab(['a', f'b{cpe}', 'c', 'd', f'e{cpe}'])
actual_vocab, actual_first_non_term_index = _create_term_vocab(input_vocab)
expected_itos = [f'b{cpe}', f'e{cpe}', 'a', 'c', 'd']
assert expected_itos == actual_vocab.itos
assert 2 == actual_first_non_term_index
def test_only_terminals():
input_vocab = Vocab([f'b{cpe}', f'e{cpe}'])
actual_vocab, actual_first_non_term_index = _create_term_vocab(input_vocab)
expected_itos = [f'b{cpe}', f'e{cpe}']
assert expected_itos == actual_vocab.itos
assert 2 == actual_first_non_term_index
def test_only_non_terminals():
input_vocab = Vocab(['a', 'c', 'd'])
actual_vocab, actual_first_non_term_index = _create_term_vocab(input_vocab)
expected_itos = ['a', 'c', 'd']
assert expected_itos == actual_vocab.itos
assert 0 == actual_first_non_term_index
def test_empty_vocab():
input_vocab = Vocab([])
actual_vocab, actual_first_non_term_index = _create_term_vocab(input_vocab)
expected_itos = []
assert expected_itos == actual_vocab.itos
assert 0 == actual_first_non_term_index
def prepare_text(df: pd.DataFrame, text_col: str, max_vocab: int,
min_freq: int, maxlen: int, word_vectors_path: str):
texts = df[text_col].tolist()
# texts = [t.lower() for t in texts]
tokens = get_texts_gensim(texts)
vocab = Vocab.create(tokens, max_vocab=max_vocab, min_freq=min_freq)
sequences = [vocab.numericalize(t) for t in tokens]
padded_seq = np.array([pad_sequences(s, maxlen=maxlen) for s in sequences])
print("Our vocabulary contains {} words".format(len(vocab.stoi)))
embedding_matrix = build_embeddings_matrix(vocab, word_vectors_path)
return padded_seq, vocab, embedding_matrix
def tokenize(self, texts):
if self.verbose:
print("Running sentence tokenizer for {} documents...".format(
len(texts)))
texts_sents = self._sentencizer(texts)
# from nested to flat list. For speed purposes
all_sents = [s for sents in texts_sents for s in sents]
# saving the lengths of the documents: 1) for padding purposes and 2) to
# compute consecutive ranges so we can "fold" the list again
texts_length = [0] + [len(s) for s in texts_sents]
range_idx = [
sum(texts_length[:i + 1]) for i in range(len(texts_length))
]
if self.verbose:
print("Tokenizing {} sentences...".format(len(all_sents)))
sents_tokens = get_texts(all_sents)
# saving the lengths of sentences for padding purposes
sents_length = [len(s) for s in sents_tokens]
try:
self.vocab
if self.verbose:
print("Using existing vocabulary")
except AttributeError:
if self.verbose:
print("Building Vocab...")
self.vocab = Vocab.create(sents_tokens,
max_vocab=self.max_vocab,
min_freq=self.min_freq)
# 'numericalize' each sentence
sents_numz = [self.vocab.numericalize(s) for s in sents_tokens]
# group the sentences again into documents
texts_numz = [
sents_numz[range_idx[i]:range_idx[i + 1]]
for i in range(len(range_idx[:-1]))
]
# compute max lengths for padding purposes
self.maxlen_sent = int(np.quantile(sents_length, q=self.q))
self.maxlen_doc = int(np.quantile(texts_length[1:], q=self.q))
if self.verbose:
print("Padding sentences and documents...")
padded_texts = [
pad_nested_sequences(
r,
self.maxlen_sent,
self.maxlen_doc,
pad_sent_first=self.pad_sent_first,
pad_doc_first=self.pad_doc_first,
pad_idx=self.pad_idx,
) for r in texts_numz
]
return np.stack(padded_texts, axis=0)
def check_data(data_bunch: DataBunch, vocab: Vocab, verbose: bool,
allow_unks: bool) -> None:
first_batch = data_bunch.one_batch()[0]
if not allow_unks and not contains_no_value(first_batch,
UNKNOWN_TOKEN_INDEX):
raise ValueError(
f"Unknown is found : {[vocab.textify(seq) for seq in first_batch]}"
)
if verbose:
logger.info(f'Displaying the first batch:\n{first_batch}')
token_seqs = [vocab.textify(seq) for seq in first_batch]
logger.info(pformat(token_seqs))
def evaluate_lm(data_path,
model_dir,
tokenizer_lang="xx",
evaluate_custom_perplexity=False):
"""
Evaluate metrics of a trained language model using any dataset of texts from CSV file.
Attributes:
data_path (str): Path to CSV file with texts in the first column.
model_dir (str): Directory with a trained language model.
tokenizer_lang (str): Language setting for tokenizer.
evaluate_custom_perplexity (bool): The perplexity estimated as e^(avg. loss),
but the average loss changes slightly with batch size. To get perplexity computed in
slower but controlled fashion, set `evaluate_custom_perplexity` to True. Discrepancy
between perplexity and custom perplexity is empirically approximately 1%.
"""
model_dir = Path(model_dir)
with open(model_dir / "lm_itos.pkl", "rb") as f:
itos = pickle.load(f)
data_df = pd.read_csv(data_path, header=None)
data = TextLMDataBunch.from_df("",
data_df,
data_df,
text_cols=0,
tokenizer=Tokenizer(lang=tokenizer_lang),
vocab=Vocab(itos))
with open(model_dir / "model_hparams.json", "r") as model_hparams_file:
model_hparams = json.load(model_hparams_file)
learner = lm_learner(data,
AWD_LSTM,
model_dir,
pretrained=True,
config=model_hparams)
loss, acc = learner.validate()
print("Loss: {}, Perplexity: {}, Accuracy: {}".format(
loss, exp(loss), acc))
if evaluate_custom_perplexity:
print(
"Custom perplexity: {}, Fraction OOV: {}, OOV perplexity contribution: {}"
.format(*evaluate_perplexity(learner, data.valid_ds.x)))
def __init__(self, path: str, after_epoch: Optional[int] = None,
force_use_cpu: bool = False, load_only_description: bool = False, device: Optional[int] = None):
if not os.path.exists(path):
raise FileNotFoundError(f'Path does not exist: {path}')
self._force_use_cpu = force_use_cpu
self._id = os.path.basename(path)
path_to_config_file = os.path.join(path, CONFIG_FILE_NAME)
path_to_metrics_file = os.path.join(path, METRICS_FILE_NAME)
path_to_tags_file = os.path.join(path, TAGS_FILE_NAME)
self._metrics = None
self._config = None
self._tags = []
self._context: List[str] = []
try:
self._config: LMTrainingConfig = load_config_or_metrics_from_file(path_to_config_file, LMTrainingConfig)
except FileNotFoundError:
logger.warning(f'Config file not found: {path_to_config_file}')
try:
self._metrics: LMTrainingMetrics = load_config_or_metrics_from_file(os.path.join(path, METRICS_FILE_NAME), LMTrainingMetrics)
except FileNotFoundError:
logger.warning(f'File with metrics not found: {path_to_metrics_file}')
if os.path.exists(path_to_tags_file):
value = read_value_from_file(path_to_tags_file, value_type=str)
if value != '':
self._tags = value.split(',')
self.prep_function = self._config.prep_function
self._load_only_description = load_only_description
if not load_only_description:
# we might want to load only description without loading actual weights when we want
# to save time when loading multiple models to choose one of them to work with
self._original_vocab = Vocab.load(os.path.join(path, VOCAB_FILE_NAME))
term_vocab, self._first_nonterm_token = _create_term_vocab(self._original_vocab)
self._model, self._vocab = self._load_model(path, after_epoch, term_vocab, device=device)
to_test_mode(self._model)
self._initial_snapshot = take_hidden_state_snapshot(self._model)
# last_predicted_token_tensor is a rank-2 tensor!
self._last_predicted_token_tensor = torch.tensor([self._vocab.numericalize([self.STARTING_TOKEN])],
device=get_device(self._force_use_cpu))
def test_simple(open_mock):
# Given
file_mock1 = file_mock_with_lines(['1', 'My Class'])
file_mock2 = file_mock_with_lines(['1', 'hi'])
open_mock.side_effect = [file_mock1, file_mock2]
numericalizer = Numericalizer(Vocab(
['`unk', '`pad', '1', 'My', 'Class', 'hi']),
n_cpus=1)
text_list = TextList(
[Path('/path/to/some/file1'),
Path('/path/to/some/file2')])
# when
numericalizer.process(text_list)
# then
expected = np.array([np.array([2, 3, 4]), np.array([2, 5])])
assert all_trues(np.equal(expected, text_list.items, dtype=np.object))
def build_vocab(docs: t.List[str], max_vocab: int = 10000, min_freq: int = 5) -> Vocab:
return Vocab.create(docs, max_vocab=max_vocab, min_freq=min_freq)
def log_vocab(self, vocab: Vocab) -> None:
logger.info(f"Vocab size: {len(vocab.itos)}")
vocab.save(os.path.join(self.path_to_trained_model, VOCAB_FILE_NAME))
if self.comet_experiment:
self.comet_experiment.log_parameter("vocabulary", len(vocab.itos))
import sys
import dataprep.api.corpus as api
from dataprep.api.corpus import PreprocessedCorpus
from fastai.text import Vocab, TextList, NumericalizeProcessor, OpenFileProcessor
if __name__ == '__main__':
prep_corpus: PreprocessedCorpus = api.bpe('/home/hlib/dev/yahtzee',
'10k',
calc_vocab=True)
vocab = Vocab(list(prep_corpus.load_vocab().keys()))
text_list = TextList.from_folder(prep_corpus.path_to_prep_dataset, vocab=vocab, extensions=['.prep'],
processor=[OpenFileProcessor(),
NumericalizeProcessor(vocab=vocab, max_vocab=sys.maxsize, min_freq=0)])\
.split_by_rand_pct()\
.label_for_lm()\
.databunch()
print(text_list)
def _create_term_vocab(vocab: Vocab) -> Tuple[Vocab, int]:
terminal_token_indices = {i for i, k in enumerate(vocab.itos) if is_terminal_subtoken(k)}
term_vocab_list = [vocab.itos[i] for i in terminal_token_indices]
non_term_vocab_list = [vocab.itos[i] for i in range(len(vocab.itos)) if i not in terminal_token_indices]
term_vocab = Vocab(term_vocab_list + non_term_vocab_list)
return term_vocab, len(term_vocab_list)
def create_vocab_for_lm(prep_corpus: PreprocessedCorpus) -> Vocab:
return Vocab(
['`unk', '`pad'] +
list(map(lambda x: to_literal_str(x),
prep_corpus.load_vocab().keys())))
def numericalize(doc: str, vocab: Vocab) -> t.List[int]:
return vocab.numericalize(doc)
