OMG / src /prompt_attention /seq_aligner.py
Fucius's picture
Upload 52 files
ad5354d verified
raw
history blame
2.51 kB
import torch
import numpy as np
def get_word_inds(text: str, word_place: int, tokenizer):
split_text = text.split(" ")
if type(word_place) is str:
word_place = [i for i, word in enumerate(split_text) if word_place == word]
elif type(word_place) is int:
word_place = [word_place]
out = []
if len(word_place) > 0:
words_encode = [tokenizer.decode([item]).strip("#") for item in tokenizer.encode(text)][1:-1]
cur_len, ptr = 0, 0
for i in range(len(words_encode)):
cur_len += len(words_encode[i])
if ptr in word_place:
out.append(i + 1)
if cur_len >= len(split_text[ptr]):
ptr += 1
cur_len = 0
return np.array(out)
def get_replacement_mapper_(x: str, y: str, tokenizer, max_len=77):
words_x = x.split(' ')
words_y = y.split(' ')
if len(words_x) != len(words_y):
raise ValueError(f"attention replacement edit can only be applied on prompts with the same length"
f" but prompt A has {len(words_x)} words and prompt B has {len(words_y)} words.")
inds_replace = [i for i in range(len(words_y)) if words_y[i] != words_x[i]]
inds_source = [get_word_inds(x, i, tokenizer) for i in inds_replace]
inds_target = [get_word_inds(y, i, tokenizer) for i in inds_replace]
mapper = np.zeros((max_len, max_len))
i = j = 0
cur_inds = 0
while i < max_len and j < max_len:
if cur_inds < len(inds_source) and inds_source[cur_inds][0] == i:
inds_source_, inds_target_ = inds_source[cur_inds], inds_target[cur_inds]
if len(inds_source_) == len(inds_target_):
mapper[inds_source_, inds_target_] = 1
else:
ratio = 1 / len(inds_target_)
for i_t in inds_target_:
mapper[inds_source_, i_t] = ratio
cur_inds += 1
i += len(inds_source_)
j += len(inds_target_)
elif cur_inds < len(inds_source):
mapper[i, j] = 1
i += 1
j += 1
else:
mapper[j, j] = 1
i += 1
j += 1
return torch.from_numpy(mapper).float()
def get_replacement_mapper(prompts, tokenizer, max_len=77):
x_seq = prompts[0]
mappers = []
for i in range(1, len(prompts)):
mapper = get_replacement_mapper_(x_seq, prompts[i], tokenizer, max_len)
mappers.append(mapper)
return torch.stack(mappers)