Spaces:
Sleeping
Sleeping
Upload event_detection_dataset.py
Browse files- event_detection_dataset.py +77 -0
event_detection_dataset.py
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import numpy as np
|
| 2 |
+
import torch
|
| 3 |
+
from torch.utils.data import Dataset
|
| 4 |
+
from collections import defaultdict
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
class event_detection_data(Dataset):
|
| 9 |
+
def __init__(self, raw_data, tokenizer, max_len, domain_adaption=False, wwm_prob=0.1):
|
| 10 |
+
self.len = len(raw_data)
|
| 11 |
+
self.data = raw_data
|
| 12 |
+
self.tokenizer = tokenizer
|
| 13 |
+
self.max_len = max_len
|
| 14 |
+
self.domain_adaption = domain_adaption
|
| 15 |
+
self.wwm_prob = wwm_prob
|
| 16 |
+
|
| 17 |
+
def __getitem__(self, index):
|
| 18 |
+
tokenized_inputs = self.tokenizer(
|
| 19 |
+
self.data[index]["text"],
|
| 20 |
+
add_special_tokens=True,
|
| 21 |
+
max_length=self.max_len,
|
| 22 |
+
padding='max_length',
|
| 23 |
+
return_token_type_ids=True,
|
| 24 |
+
truncation=True,
|
| 25 |
+
is_split_into_words=True
|
| 26 |
+
)
|
| 27 |
+
|
| 28 |
+
ids = tokenized_inputs['input_ids']
|
| 29 |
+
mask = tokenized_inputs['attention_mask']
|
| 30 |
+
|
| 31 |
+
if self.domain_adaption:
|
| 32 |
+
if self.tokenizer.is_fast:
|
| 33 |
+
input_ids, labels = self._whole_word_masking(self.tokenizer, tokenized_inputs, self.wwm_prob)
|
| 34 |
+
return {
|
| 35 |
+
'input_ids': torch.tensor(input_ids),
|
| 36 |
+
'attention_mask': torch.tensor(mask),
|
| 37 |
+
'labels': torch.tensor(labels, dtype=torch.long)
|
| 38 |
+
}
|
| 39 |
+
else:
|
| 40 |
+
print("requires fast tokenizer for word_ids")
|
| 41 |
+
else:
|
| 42 |
+
return {
|
| 43 |
+
'input_ids': torch.tensor(ids),
|
| 44 |
+
'attention_mask': torch.tensor(mask),
|
| 45 |
+
'targets': torch.tensor(self.data[index]["text_tag_id"][0], dtype=torch.long)
|
| 46 |
+
}
|
| 47 |
+
|
| 48 |
+
def __len__(self):
|
| 49 |
+
return self.len
|
| 50 |
+
|
| 51 |
+
def _whole_word_masking(self, tokenizer, tokenized_inputs, wwm_prob):
|
| 52 |
+
word_ids = tokenized_inputs.word_ids(0)
|
| 53 |
+
|
| 54 |
+
# create a map between words_ids and natural id
|
| 55 |
+
mapping = defaultdict(list)
|
| 56 |
+
current_word_index = -1
|
| 57 |
+
current_word = None
|
| 58 |
+
|
| 59 |
+
for idx, word_id in enumerate(word_ids):
|
| 60 |
+
if word_id is not None:
|
| 61 |
+
if word_id != current_word:
|
| 62 |
+
current_word = word_id
|
| 63 |
+
current_word_index += 1
|
| 64 |
+
mapping[current_word_index].append(idx)
|
| 65 |
+
|
| 66 |
+
# randomly mask words
|
| 67 |
+
mask = np.random.binomial(1, wwm_prob, (len(mapping),))
|
| 68 |
+
input_ids = tokenized_inputs["input_ids"]
|
| 69 |
+
|
| 70 |
+
# labels only contains masked words as target
|
| 71 |
+
labels = [-100] * len(input_ids)
|
| 72 |
+
|
| 73 |
+
for word_id in np.where(mask == 1)[0]:
|
| 74 |
+
for idx in mapping[word_id]:
|
| 75 |
+
labels[idx] = tokenized_inputs["input_ids"][idx]
|
| 76 |
+
input_ids[idx] = tokenizer.mask_token_id
|
| 77 |
+
return input_ids, labels
|