class ChatDataset(Dataset):
def __init__(self, filepath, tok_vocab, max_seq_len=128) -> None:
self.filepath = filepath
self.data = pd.read_csv(self.filepath)
self.bos_token = '<s>'
self.eos_token = '</s>'
self.max_seq_len = max_seq_len
self.tokenizer = PreTrainedTokenizerFast(tokenizer_file=tok_vocab,
bos_token=self.bos_token,
eos_token=self.eos_token,
unk_token='<unk>',
pad_token='<pad>',
mask_token='<mask>')
def __len__(self):
return len(self.data)
def make_input_id_mask(self, tokens, index):
input_id = self.tokenizer.convert_tokens_to_ids(tokens)
attention_mask = [1] * len(input_id)
if len(input_id) < self.max_seq_len:
while len(input_id) < self.max_seq_len:
input_id += [self.tokenizer.pad_token_id]
attention_mask += [0]
else:
# logging.warning(f'exceed max_seq_len for given article : {index}')
input_id = input_id[:self.max_seq_len -
1] + [self.tokenizer.eos_token_id]
attention_mask = attention_mask[:self.max_seq_len]
return input_id, attention_mask
def __getitem__(self, index):
record = self.data.iloc[index]
q, a = record['Q'], record['A']
q_tokens = [self.bos_token] + \
self.tokenizer.tokenize(q) + [self.eos_token]
a_tokens = [self.bos_token] + \
self.tokenizer.tokenize(a) + [self.eos_token]
encoder_input_id, encoder_attention_mask = self.make_input_id_mask(
q_tokens, index)
decoder_input_id, decoder_attention_mask = self.make_input_id_mask(
a_tokens, index)
labels = self.tokenizer.convert_tokens_to_ids(
a_tokens[1:(self.max_seq_len + 1)])
if len(labels) < self.max_seq_len:
while len(labels) < self.max_seq_len:
# for cross entropy loss masking
labels += [-100]
return {
'input_ids':
np.array(encoder_input_id, dtype=np.int_),
'attention_mask':
np.array(encoder_attention_mask, dtype=np.float_),
'decoder_input_ids':
np.array(decoder_input_id, dtype=np.int_),
'decoder_attention_mask':
np.array(decoder_attention_mask, dtype=np.float_),
'labels':
np.array(labels, dtype=np.int_)
}
# tok.save("THE_TEST.tokenizer.json", pretty=True)
# print(tok.encode("𝕿𝖍𝖊 𝖖𝖚𝖎𝖈𝖐, 𝖇𝖗𝖔𝖜𝖓 🦊 𝖏𝖚𝖒𝖕𝖘 𝖔𝖛𝖊𝖗 𝖙𝖍𝖊 𝖑𝖆𝖟𝖞 🐶").tokens)
#
# tok = Tokenizer.from_file("THE_TEST.tokenizer.json")
# # with open("THE_TEST.tokenizer.json", "r") as f:
# # t = f.read()
# # tok = Tokenizer.from_str(t)
# print(tok.encode("𝕿𝖍𝖊 𝖖𝖚𝖎𝖈𝖐, 𝖇𝖗𝖔𝖜𝖓 🦊 𝖏𝖚𝖒𝖕𝖘 𝖔𝖛𝖊𝖗 𝖙𝖍𝖊 𝖑𝖆𝖟𝖞 🐶").tokens)
from tokenizers import Tokenizer
from tokenizers.implementations import BaseTokenizer
from transformers import PreTrainedTokenizerFast, LineByLineTextDataset
# tokenizer = Tokenizer(
# BPE("../../data/roberta-base-vocab.json", "../../data/roberta-base-merges.txt")
# )
tokenizer = Tokenizer.from_file("../../data/roberta-tok.tokenizer")
print(tokenizer.encode("Hello there!").tokens)
tok_transformers = PreTrainedTokenizerFast(BaseTokenizer(tokenizer))
print(tok_transformers.tokenize("Hello there!"))
dataset = LineByLineTextDataset(tokenizer=tok_transformers,
file_path="../../data/botchan.txt",
block_size=12)
# tokenizer = ByteLevelBPETokenizer.from_files(
# "../../data/roberta-base-vocab.json", "../../data/roberta-base-merges.txt"
# )
# print(tokenizer.encode("Hello there!").tokens)
def convert_examples_to_features(
model_meta: ModelMeta,
input_examples: List[InputExample],
class_list: List[str],
tokenizer: PreTrainedTokenizerFast,
use_iob2_format: bool = False) -> List[InputFeatures]:
max_allowed_tokens = model_meta.max_seq_length - tokenizer.num_special_tokens_to_add(
) + int(model_meta.sep_token_extra)
class_map = {label: i for i, label in enumerate(class_list)}
features = []
for ex_index, example in tqdm(enumerate(input_examples),
total=len(input_examples),
desc='generating features'):
#loop over sentences in an example
context_subsets = []
current_subset = []
last_count = 0
for sentence_meta in example.sentence_meta_list:
tokens = []
label_ids = []
for word, label in zip(sentence_meta.words, sentence_meta.labels):
word_tokens = tokenizer.tokenize(word)
if len(word_tokens) > 0:
tokens.extend(word_tokens)
# Use the real label id for the first token of the word, and padding ids for the remaining tokens
temp_pad_id = model_meta.pad_token_label_id
label_id = class_map[label]
if label != 'O':
if use_iob2_format:
temp_pad_id = class_map['I-' + label[2:]]
else:
temp_pad_id = class_map[label]
label_ids.extend([label_id] + [temp_pad_id] *
(len(word_tokens) - 1))
assert (len(tokens) == len(label_ids))
num_tokens = len(tokens)
sent_feature = SentenceMeta(tokens, label_ids)
if last_count + num_tokens >= max_allowed_tokens:
if len(current_subset) == 0:
#meaning a single sentence is exceeding the max_seq_len of the model
context_subsets.append([sent_feature])
else:
context_subsets.append(current_subset)
current_subset = [sent_feature]
last_count = num_tokens
else:
last_count += num_tokens
current_subset.append(sent_feature)
if len(current_subset) > 0:
context_subsets.append(current_subset)
#finally create concatenated features based on sent features
for context_subset in context_subsets:
feature_tokens = []
feature_labels = []
for sent_feature in context_subset:
feature_tokens.extend(sent_feature.words)
feature_labels.extend(sent_feature.labels)
input_features = convert_tokens_to_features(
tokenizer=tokenizer,
model_meta=model_meta,
feature_tokens=feature_tokens,
feature_labels=feature_labels)
features.append(input_features)
return features
