You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Describe the bug
When using textattack.augmentation.recipes.CheckListAugmenter on the yahoo_answers_topics dataset in Huggingface, I get an error:
...
File "~\textattack\transformations\word_swaps\word_swap_change_number.py", line 108, in _alter_number
num_list = np.random.randint(max(num - change, 1), num + change, self.n, dtype=np.int64)
File "mtrand.pyx", line 765, in numpy.random.mtrand.RandomState.randint
File "_bounded_integers.pyx", line 1245, in numpy.random._bounded_integers._rand_int64
ValueError: high is out of bounds for int64
Whatever number is being found in the text is too large for np.int32. A partial workaround that solved the problem in ag_news was to change change the dtype=np.int64, but its still happening now for yahoo_answers_topics.
To Reproduce
class AugMapper:
def __init__(self, augmenter):
self.augmenter = augmenter # textattack augmenter recipe
def apply_to_batch(self, batch):
new_texts, new_labels = [], []
for text, label in zip(batch['text'], batch['label']):
augmented_text = self.augmenter.augment(text)
new_texts.extend(augmented_text)
new_labels.extend([label] * len(augmented_text))
return {
"text": new_texts,
"label": new_labels,
"idx": list(range(len(new_labels))),
}
if __name__ == "__main__":
import os
import glob
from datasets import load_dataset, load_from_disk
from textattack.augmentation.recipes import (
EasyDataAugmenter,
CheckListAugmenter,
)
# checklist
checklist_augmenter = CheckListAugmenter(
transformations_per_example=3
)
checklist_aug_mapper = AugMapper(checklist_augmenter)
dataset_paths = ["yahoo_answers_topics"]
for dataset_path in dataset_paths:
print(dataset_path)
dataset = load_dataset(dataset_path, split="train")
# augment + save checklist
checklist_dataset = dataset.map(checklist_aug_mapper.apply_to_batch, batched=True, batch_size=10)
Expected behavior
The augmentation should not error out
Screenshots or Traceback
If applicable, add screenshots to help explain your problem. Also, copy and paste tracebacks produced by the bug.
System Information (please complete the following information):
OS: Windows
Textattack version: 0.3.8
Additional context
Potential fix that is working so far --> cap the values to the max that numpy can handle for int64
def _alter_number(self, num):
"""helper function of _get_new_number, replace a number with another
random number within the range of self.max_change."""
if num not in [0, 2, 4]:
change = int(num * self.max_change) + 1
if num >= 0:
num_list = np.random.randint(
low=max(num - change, 1),
high=min(num + change, np.iinfo(np.int64).max - 1),
size=self.n,
dtype=np.int64)
else:
num_list = np.random.randint(
low=max(num - change, np.iinfo(np.int64).min + 1),
high=min(0, num + change),
size=self.n,
dtype=np.int64)
return num_list
return []
The text was updated successfully, but these errors were encountered:
Describe the bug
When using
textattack.augmentation.recipes.CheckListAugmenter
on theyahoo_answers_topics
dataset in Huggingface, I get an error:Whatever number is being found in the text is too large for np.int32. A partial workaround that solved the problem in
ag_news
was to change change the dtype=np.int64, but its still happening now foryahoo_answers_topics
.To Reproduce
Expected behavior
The augmentation should not error out
Screenshots or Traceback
If applicable, add screenshots to help explain your problem. Also, copy and paste tracebacks produced by the bug.
System Information (please complete the following information):
Additional context
Potential fix that is working so far --> cap the values to the max that numpy can handle for int64
The text was updated successfully, but these errors were encountered: