blob: 7ffa77f8902eebe16cd7f8fe34d07035463eaa16 (
plain) (
blame)
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
|
from pathlib import Path
import torch
from transformers import CLIPTextModel, CLIPTokenizer
def load_text_embedding(embeddings, token_id, file):
data = torch.load(file, map_location="cpu")
assert len(data.keys()) == 1, 'embedding data has multiple terms in it'
emb = next(iter(data.values()))
if len(emb.shape) == 1:
emb = emb.unsqueeze(0)
embeddings[token_id] = emb
def load_text_embeddings(tokenizer: CLIPTokenizer, text_encoder: CLIPTextModel, embeddings_dir: Path):
if not embeddings_dir.exists() or not embeddings_dir.is_dir():
return []
files = [file for file in embeddings_dir.iterdir() if file.is_file()]
tokens = [file.stem for file in files]
added = tokenizer.add_tokens(tokens)
token_ids = tokenizer.convert_tokens_to_ids(tokens)
text_encoder.resize_token_embeddings(len(tokenizer))
token_embeds = text_encoder.get_input_embeddings().weight.data
for (token_id, file) in zip(token_ids, files):
load_text_embedding(token_embeds, token_id, file)
return tokens
|
