Upload 2 files
Browse files
embeddingsXL.unum.eng.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:0e634bd6b3a7f2d02fe70ef93133d46eba58f43e8893fdd569ba99a568a3714b
|
3 |
+
size 104074336
|
generate-dict-embeddingsXL.py
ADDED
@@ -0,0 +1,133 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/env python
|
2 |
+
|
3 |
+
"""
|
4 |
+
(SDXL counterpart of "cliptextmodel-generate-embeddings.py".
|
5 |
+
Not following that name, because we dont use "cliptextmodel")
|
6 |
+
|
7 |
+
Take filenames of an SDXL clip-g type text_encoder2 and config file
|
8 |
+
Read in a wordlist from "dictionary"
|
9 |
+
Generate the official "embedding" tensor for each one.
|
10 |
+
Save the result set to "{outputfile}"
|
11 |
+
|
12 |
+
Defaults to loading openai/clip-vit-large-patch14 from huggingface hub,
|
13 |
+
for purposes of tokenizer, since thats what sdxl does anyway
|
14 |
+
|
15 |
+
RULES of the loader:
|
16 |
+
1. The text_encoder2 model file must appear to be either
|
17 |
+
in current directory or one down. So, do NOT use
|
18 |
+
badpath1=some/directory/tree/file.here
|
19 |
+
badpath2=/absolutepath
|
20 |
+
2. Yes, you MUST have a matching config.json file
|
21 |
+
3. if you have no safetensor alternative, you can get away with using pytorch_model.bin
|
22 |
+
|
23 |
+
Sample location for such things that you can download:
|
24 |
+
https://huggingface.co/stablediffusionapi/edge-of-realism/tree/main/text_encoder/
|
25 |
+
If there is a .safetensors AND a .bin file, ignore the .bin file
|
26 |
+
|
27 |
+
Alternatively, you can also convert a singlefile model, such as is downloaded from civitai,
|
28 |
+
by using the utility at
|
29 |
+
https://github.com/huggingface/diffusers/blob/main/scripts/convert_original_stable_diffusion_to_diffusers.py
|
30 |
+
Args should look like
|
31 |
+
convert_original_stable_diffusion_to_diffusers.py \
|
32 |
+
--checkpoint_file somemodel.safetensors \
|
33 |
+
--dump_path extractdir --to_safetensors --from_safetensors
|
34 |
+
|
35 |
+
"""
|
36 |
+
|
37 |
+
|
38 |
+
outputfile="embeddingsXL.temp.safetensors"
|
39 |
+
|
40 |
+
import sys
|
41 |
+
import torch
|
42 |
+
from safetensors.torch import save_file
|
43 |
+
from transformers import CLIPProcessor, CLIPTextModel, CLIPTextModelWithProjection
|
44 |
+
|
45 |
+
processor=None
|
46 |
+
|
47 |
+
|
48 |
+
tmodel2=None
|
49 |
+
model_path2=None
|
50 |
+
model_config2=None
|
51 |
+
|
52 |
+
if len(sys.argv) == 3:
|
53 |
+
model_path2=sys.argv[1]
|
54 |
+
model_config2=sys.argv[2]
|
55 |
+
else:
|
56 |
+
print("You have to give name of modelfile and config file")
|
57 |
+
sys.exit(1)
|
58 |
+
|
59 |
+
device=torch.device("cuda")
|
60 |
+
|
61 |
+
|
62 |
+
|
63 |
+
def initXLCLIPmodel(model_path,model_config):
|
64 |
+
global tmodel2,processor
|
65 |
+
# yes, oddly they all uses the same one, basically
|
66 |
+
processor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")
|
67 |
+
|
68 |
+
print("loading",model_path)
|
69 |
+
tmodel2 = CLIPTextModelWithProjection.from_pretrained(model_path,config=model_config,local_files_only=True,use_safetensors=True)
|
70 |
+
tmodel2.to(device)
|
71 |
+
|
72 |
+
|
73 |
+
def embed_from_text2(text):
|
74 |
+
global processor,tmodel2
|
75 |
+
inputs = processor(text=text, return_tensors="pt")
|
76 |
+
inputs.to(device)
|
77 |
+
|
78 |
+
with torch.no_grad():
|
79 |
+
outputs = tmodel2(**inputs)
|
80 |
+
embeddings = outputs.text_embeds
|
81 |
+
return embeddings
|
82 |
+
|
83 |
+
|
84 |
+
# "inputs" == magic pre-embedding format
|
85 |
+
def embed_from_inputs(inputs):
|
86 |
+
global processor,tmodel2
|
87 |
+
with torch.no_grad():
|
88 |
+
outputs = tmodel2(**inputs)
|
89 |
+
embedding = outputs.text_embeds
|
90 |
+
|
91 |
+
return embedding
|
92 |
+
|
93 |
+
|
94 |
+
initXLCLIPmodel(model_path2,model_config2)
|
95 |
+
inputs = processor(text="dummy", return_tensors="pt")
|
96 |
+
inputs.to(device)
|
97 |
+
|
98 |
+
|
99 |
+
with open("dictionary","r") as f:
|
100 |
+
tokendict = f.readlines()
|
101 |
+
tokendict = [token.strip() for token in tokendict] # Remove trailing newlines
|
102 |
+
|
103 |
+
|
104 |
+
count=1
|
105 |
+
all_embeddings = []
|
106 |
+
|
107 |
+
for word in tokendict:
|
108 |
+
emb = embed_from_text2(word)
|
109 |
+
emb=emb.unsqueeze(0) # stupid matrix magic to make the cat work
|
110 |
+
all_embeddings.append(emb)
|
111 |
+
count+=1
|
112 |
+
if (count %100) ==0:
|
113 |
+
print(count)
|
114 |
+
|
115 |
+
|
116 |
+
|
117 |
+
"""
|
118 |
+
for id in range(49405):
|
119 |
+
inputs.input_ids[0][1]=id
|
120 |
+
|
121 |
+
emb=embed_from_inputs(inputs)
|
122 |
+
all_embeddings.append(emb)
|
123 |
+
if (id %100) ==0:
|
124 |
+
print(id)
|
125 |
+
"""
|
126 |
+
|
127 |
+
embs = torch.cat(all_embeddings,dim=0)
|
128 |
+
print("Shape of result = ",embs.shape)
|
129 |
+
|
130 |
+
print(f"Saving the calculatiuons to {outputfile}...")
|
131 |
+
save_file({"embeddings": embs}, outputfile)
|
132 |
+
|
133 |
+
|