def __init__(self, model, top_n=1, replace_stopwords=False):
validate_model_gradient_word_swap_compatibility(model)
if not hasattr(model, 'word_embeddings'):
raise ValueError(
'Model needs word embedding matrix for gradient-based word swap'
)
if not hasattr(model, 'lookup_table'):
raise ValueError(
'Model needs lookup table for gradient-based word swap')
if not hasattr(model, 'zero_grad'):
raise ValueError(
'Model needs `zero_grad()` for gradient-based word swap')
if not hasattr(model.tokenizer, 'convert_id_to_word'):
raise ValueError(
'Tokenizer needs `convert_id_to_word()` for gradient-based word swap'
)
if not hasattr(model.tokenizer, 'pad_id'):
raise ValueError(
'Tokenizer needs `pad_id` for gradient-based word swap')
if not hasattr(model.tokenizer, 'oov_id'):
raise ValueError(
'Tokenizer needs `oov_id` for gradient-based word swap')
self.loss = torch.nn.CrossEntropyLoss()
self.model = model
self.pad_id = self.model.tokenizer.pad_id
self.oov_id = self.model.tokenizer.oov_id
self.top_n = top_n
self.replace_stopwords = replace_stopwords
if replace_stopwords:
self.stopwords = set()
else:
from nltk.corpus import stopwords
self.stopwords = set(stopwords.words('english'))
def __init__(self, model, top_n=1):
validate_model_gradient_word_swap_compatibility(model)
if not hasattr(model, 'word_embeddings'):
raise ValueError(
'Model needs word embedding matrix for gradient-based word swap'
)
if not hasattr(model, 'lookup_table'):
raise ValueError(
'Model needs lookup table for gradient-based word swap')
if not hasattr(model, 'zero_grad'):
raise ValueError(
'Model needs `zero_grad()` for gradient-based word swap')
if not hasattr(model.tokenizer, 'convert_id_to_word'):
raise ValueError(
'Tokenizer needs `convert_id_to_word()` for gradient-based word swap'
)
if not hasattr(model.tokenizer, 'pad_id'):
raise ValueError(
'Tokenizer needs `pad_id` for gradient-based word swap')
if not hasattr(model.tokenizer, 'oov_id'):
raise ValueError(
'Tokenizer needs `oov_id` for gradient-based word swap')
self.loss = torch.nn.CrossEntropyLoss()
self.model = model
self.pad_id = self.model.tokenizer.pad_id
self.oov_id = self.model.tokenizer.oov_id
self.top_n = top_n
self.is_black_box = False
def __init__(self, model_wrapper, top_n=1):
# Unwrap model wrappers. Need raw model for gradient.
if not isinstance(model_wrapper, textattack.models.wrappers.ModelWrapper):
raise TypeError(f"Got invalid model wrapper type {type(model_wrapper)}")
self.model = model_wrapper.model
self.model_wrapper = model_wrapper
self.tokenizer = self.model_wrapper.tokenizer
# Make sure we know how to compute the gradient for this model.
validate_model_gradient_word_swap_compatibility(self.model)
# Make sure this model has all of the required properties.
if not hasattr(self.model, "get_input_embeddings"):
raise ValueError(
"Model needs word embedding matrix for gradient-based word swap"
)
if not hasattr(self.tokenizer, "pad_token_id") and self.tokenizer.pad_token_id:
raise ValueError(
"Tokenizer needs to have `pad_token_id` for gradient-based word swap"
)
self.top_n = top_n
self.is_black_box = False
def __init__(self, model_wrapper, top_n=1):
# Unwrap model wrappers. Need raw model for gradient.
if not isinstance(model_wrapper,
textattack.models.wrappers.ModelWrapper):
raise TypeError(
f"Got invalid model wrapper type {type(model_wrapper)}")
self.model = model_wrapper.model
self.model_wrapper = model_wrapper
self.tokenizer = self.model_wrapper.tokenizer
# Make sure we know how to compute the gradient for this model.
validate_model_gradient_word_swap_compatibility(self.model)
# Make sure this model has all of the required properties.
if not hasattr(self.model, "word_embeddings"):
raise ValueError(
"Model needs word embedding matrix for gradient-based word swap"
)
if not hasattr(self.model, "lookup_table"):
raise ValueError(
"Model needs lookup table for gradient-based word swap")
if not hasattr(self.model, "zero_grad"):
raise ValueError(
"Model needs `zero_grad()` for gradient-based word swap")
if not hasattr(self.tokenizer, "convert_id_to_word"):
raise ValueError(
"Tokenizer needs `convert_id_to_word()` for gradient-based word swap"
)
if not hasattr(self.tokenizer, "pad_id"):
raise ValueError(
"Tokenizer needs `pad_id` for gradient-based word swap")
if not hasattr(self.tokenizer, "oov_id"):
raise ValueError(
"Tokenizer needs `oov_id` for gradient-based word swap")
self.loss = torch.nn.CrossEntropyLoss()
self.pad_id = self.model_wrapper.tokenizer.pad_id
self.oov_id = self.model_wrapper.tokenizer.oov_id
self.top_n = top_n
self.is_black_box = False