def get_positions(self,
input_ids: T,
tenzorizer: Tensorizer,
model: torch.nn.Module = None):
if not self.token_id:
self.token_id = tenzorizer.get_token_id(self.token)
token_indexes = (input_ids == self.token_id).nonzero()
# check if all samples in input_ids has index presence and out a default value otherwise
bsz = input_ids.size(0)
if bsz == token_indexes.size(0):
return token_indexes
token_indexes_result = []
found_idx_cnt = 0
for i in range(bsz):
if (found_idx_cnt < token_indexes.size(0)
and token_indexes[found_idx_cnt][0] == i):
# this samples has the special token
token_indexes_result.append(token_indexes[found_idx_cnt])
found_idx_cnt += 1
else:
logger.warning("missing special token %s", input_ids[i])
token_indexes_result.append(
torch.tensor([i, 0]).to(input_ids.device)
) # setting 0-th token, i.e. CLS for BERT as the special one
token_indexes_result = torch.stack(token_indexes_result, dim=0)
return token_indexes_result
def _select_span_with_token(
text: str, tensorizer: Tensorizer, token_str: str = "[START_ENT]"
) -> T:
id = tensorizer.get_token_id(token_str)
query_tensor = tensorizer.text_to_tensor(text)
if id not in query_tensor:
query_tensor_full = tensorizer.text_to_tensor(text, apply_max_len=False)
token_indexes = (query_tensor_full == id).nonzero()
if token_indexes.size(0) > 0:
start_pos = token_indexes[0, 0].item()
# add some randomization to avoid overfitting to a specific token position
left_shit = int(tensorizer.max_length / 2)
rnd_shift = int((rnd.random() - 0.5) * left_shit / 2)
left_shit += rnd_shift
query_tensor = query_tensor_full[start_pos - left_shit :]
cls_id = tensorizer.tokenizer.cls_token_id
if query_tensor[0] != cls_id:
query_tensor = torch.cat([torch.tensor([cls_id]), query_tensor], dim=0)
from dpr.models.reader import _pad_to_len
query_tensor = _pad_to_len(
query_tensor, tensorizer.get_pad_id(), tensorizer.max_length
)
query_tensor[-1] = tensorizer.tokenizer.sep_token_id
# logger.info('aligned query_tensor %s', query_tensor)
assert id in query_tensor, "query_tensor={}".format(query_tensor)
return query_tensor
else:
raise RuntimeError(
"[START_ENT] toke not found for Entity Linking sample query={}".format(
text
)
)
else:
return query_tensor
