Spaces:
Running
on
Zero
Running
on
Zero
File size: 5,603 Bytes
0a63786 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 |
from dataclasses import dataclass
import torch
import numpy as np
@dataclass
class Edit:
old: str
new: str
weight: float = 1.0
@dataclass
class Insert:
text: str
weight: float = 1.0
@property
def old(self):
return ""
@property
def new(self):
return self.text
@dataclass
class Delete:
text: str
weight: float = 1.0
@property
def old(self):
return self.text
@property
def new(self):
return ""
@dataclass
class Text:
text: str
weight: float = 1.0
@property
def old(self):
return self.text
@property
def new(self):
return self.text
@torch.inference_mode()
def get_text_embedding(prompt, tokenizer, text_encoder):
text_input_ids = tokenizer(
prompt,
padding="max_length",
truncation=True,
max_length=tokenizer.model_max_length,
return_tensors="pt",
).input_ids
text_embeddings = text_encoder(text_input_ids.to(text_encoder.device))[0]
return text_embeddings
@torch.inference_mode()
def encode_text(text_pieces, tokenizer, text_encoder):
n_old_tokens = 0
n_new_tokens = 0
new_id_to_old_id = []
weights = []
for piece in text_pieces:
old, new = piece.old, piece.new
old_tokens = tokenizer.tokenize(old)
new_tokens = tokenizer.tokenize(new)
if len(old_tokens) == 0 and len(new_tokens) == 0:
continue
elif old == new:
n_old_tokens += len(old_tokens)
n_new_tokens += len(new_tokens)
new_id_to_old_id.extend(range(n_old_tokens - len(old_tokens), n_old_tokens))
elif len(old_tokens) == 0:
# insert
new_id_to_old_id.extend([-1] * len(new_tokens))
n_new_tokens += len(new_tokens)
elif len(new_tokens) == 0:
# delete
n_old_tokens += len(old_tokens)
else:
# replace
n_old_tokens += len(old_tokens)
n_new_tokens += len(new_tokens)
start = n_old_tokens - len(old_tokens)
end = n_old_tokens
ids = np.linspace(start, end, len(new_tokens), endpoint=False).astype(int)
new_id_to_old_id.extend(list(ids))
weights.extend([piece.weight] * len(new_tokens))
old_prompt = " ".join([piece.old for piece in text_pieces])
new_prompt = " ".join([piece.new for piece in text_pieces])
old_text_input_ids = tokenizer(
old_prompt,
padding="max_length",
truncation=True,
max_length=tokenizer.model_max_length,
return_tensors="pt",
).input_ids
new_text_input_ids = tokenizer(
new_prompt,
padding="max_length",
truncation=True,
max_length=tokenizer.model_max_length,
return_tensors="pt",
).input_ids
old_text_embeddings = text_encoder(old_text_input_ids.to(text_encoder.device))[0]
new_text_embeddings = text_encoder(new_text_input_ids.to(text_encoder.device))[0]
value = new_text_embeddings.clone() # batch (1), seq, dim
key = new_text_embeddings.clone()
for i, (j, weight) in enumerate(zip(new_id_to_old_id, weights)):
if 0 <= j < old_text_embeddings.shape[1]:
key[0, i] = old_text_embeddings[0, j]
value[0, i] *= weight
return key, value
@torch.inference_mode()
def get_text_embedding_openclip(prompt, text_encoder, device='cuda'):
import open_clip
text_input_ids = open_clip.tokenize(prompt)
text_embeddings = text_encoder(text_input_ids.to(device))
return text_embeddings
@torch.inference_mode()
def encode_text_openclip(text_pieces, text_encoder, device='cuda'):
import open_clip
n_old_tokens = 0
n_new_tokens = 0
new_id_to_old_id = []
weights = []
for piece in text_pieces:
old, new = piece.old, piece.new
old_tokens = open_clip.tokenize(old)
new_tokens = open_clip.tokenize(new)
if len(old_tokens) == 0 and len(new_tokens) == 0:
continue
elif old == new:
n_old_tokens += len(old_tokens)
n_new_tokens += len(new_tokens)
new_id_to_old_id.extend(range(n_old_tokens - len(old_tokens), n_old_tokens))
elif len(old_tokens) == 0:
# insert
new_id_to_old_id.extend([-1] * len(new_tokens))
n_new_tokens += len(new_tokens)
elif len(new_tokens) == 0:
# delete
n_old_tokens += len(old_tokens)
else:
# replace
n_old_tokens += len(old_tokens)
n_new_tokens += len(new_tokens)
start = n_old_tokens - len(old_tokens)
end = n_old_tokens
ids = np.linspace(start, end, len(new_tokens), endpoint=False).astype(int)
new_id_to_old_id.extend(list(ids))
weights.extend([piece.weight] * len(new_tokens))
old_prompt = " ".join([piece.old for piece in text_pieces])
new_prompt = " ".join([piece.new for piece in text_pieces])
old_text_input_ids = open_clip.tokenize(old_prompt)
new_text_input_ids = open_clip.tokenize(new_prompt)
old_text_embeddings = text_encoder(old_text_input_ids.to(device))
new_text_embeddings = text_encoder(new_text_input_ids.to(device))
value = new_text_embeddings.clone() # batch (1), seq, dim
key = new_text_embeddings.clone()
for i, (j, weight) in enumerate(zip(new_id_to_old_id, weights)):
if 0 <= j < old_text_embeddings.shape[1]:
key[0, i] = old_text_embeddings[0, j]
value[0, i] *= weight
return key, value |