Upload 2 files
Browse files- compare-allids-embeds.py +56 -0
- generate-allid-embeddings.py +62 -15
compare-allids-embeds.py
ADDED
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/env python
|
2 |
+
|
3 |
+
""" Work in progress
|
4 |
+
temp utility.
|
5 |
+
Load in two pre-calculated embeddings files.
|
6 |
+
(eg: *.allid.*)
|
7 |
+
|
8 |
+
Go through the full range and calculate distances between each.
|
9 |
+
|
10 |
+
Add up and display
|
11 |
+
|
12 |
+
This covers the full official range of tokenids,
|
13 |
+
0-49405
|
14 |
+
|
15 |
+
"""
|
16 |
+
|
17 |
+
|
18 |
+
import sys
|
19 |
+
import torch
|
20 |
+
from safetensors import safe_open
|
21 |
+
|
22 |
+
file1=sys.argv[1]
|
23 |
+
file2=sys.argv[2]
|
24 |
+
|
25 |
+
|
26 |
+
|
27 |
+
device=torch.device("cuda")
|
28 |
+
print(f"reading {file1} embeddings now",file=sys.stderr)
|
29 |
+
model = safe_open(file1,framework="pt",device="cuda")
|
30 |
+
embs1=model.get_tensor("embeddings")
|
31 |
+
embs1.to(device)
|
32 |
+
print("Shape of loaded embeds =",embs1.shape)
|
33 |
+
|
34 |
+
print(f"reading {file2} embeddings now",file=sys.stderr)
|
35 |
+
model = safe_open(file2,framework="pt",device="cuda")
|
36 |
+
embs2=model.get_tensor("embeddings")
|
37 |
+
embs2.to(device)
|
38 |
+
print("Shape of loaded embeds =",embs2.shape)
|
39 |
+
|
40 |
+
if torch.equal(embs1 , embs2):
|
41 |
+
print("HEY! Both files are identical!")
|
42 |
+
exit(0)
|
43 |
+
|
44 |
+
print(f"calculating distances...")
|
45 |
+
|
46 |
+
# This calculates a full cross matrix of ALL distances to ALL other points
|
47 |
+
# in other tensor
|
48 |
+
##targetdistances = torch.cdist( embs1,embs2, p=2)
|
49 |
+
|
50 |
+
|
51 |
+
targetdistances = torch.norm(embs2 - embs1, dim=1)
|
52 |
+
print(targetdistances.shape)
|
53 |
+
tl=targetdistances.tolist()
|
54 |
+
|
55 |
+
print(tl[:10])
|
56 |
+
|
generate-allid-embeddings.py
CHANGED
@@ -1,11 +1,31 @@
|
|
1 |
#!/bin/env python
|
2 |
|
3 |
-
"""
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
"""
|
10 |
|
11 |
|
@@ -13,11 +33,17 @@ import sys
|
|
13 |
import json
|
14 |
import torch
|
15 |
from safetensors.torch import save_file
|
16 |
-
from transformers import CLIPProcessor,CLIPModel
|
17 |
|
18 |
clipsrc="openai/clip-vit-large-patch14"
|
19 |
processor=None
|
20 |
model=None
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
device=torch.device("cuda")
|
23 |
|
@@ -25,22 +51,42 @@ device=torch.device("cuda")
|
|
25 |
def init():
|
26 |
global processor
|
27 |
global model
|
|
|
|
|
|
|
28 |
# Load the processor and model
|
29 |
print("loading processor from "+clipsrc,file=sys.stderr)
|
30 |
processor = CLIPProcessor.from_pretrained(clipsrc)
|
31 |
print("done",file=sys.stderr)
|
32 |
-
|
33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
print("done",file=sys.stderr)
|
35 |
|
36 |
model = model.to(device)
|
37 |
|
38 |
|
39 |
-
|
40 |
def embed_from_inputs(inputs):
|
41 |
with torch.no_grad():
|
42 |
-
|
43 |
-
|
|
|
|
|
|
|
|
|
|
|
44 |
|
45 |
return embedding
|
46 |
|
@@ -55,14 +101,15 @@ for id in range(49405):
|
|
55 |
inputs.input_ids[0][1]=id
|
56 |
|
57 |
emb=embed_from_inputs(inputs)
|
58 |
-
emb=emb.unsqueeze(0) # stupid matrix magic to make the cat work
|
59 |
all_embeddings.append(emb)
|
60 |
if (id %100) ==0:
|
61 |
print(id)
|
62 |
|
63 |
embs = torch.cat(all_embeddings,dim=0)
|
64 |
print("Shape of result = ",embs.shape)
|
65 |
-
|
66 |
-
|
|
|
|
|
67 |
|
68 |
|
|
|
1 |
#!/bin/env python
|
2 |
|
3 |
+
"""
|
4 |
+
Take a CLIPTextModel compatible text encoder.
|
5 |
+
Go through the official range of tokens IDs (0-49405)
|
6 |
+
Generate the official "embedding" tensor for each one.
|
7 |
+
Save the result set to "temp.allids.safetensors"
|
8 |
+
|
9 |
+
Defaults to loading openai/clip-vit-large-patch14 from huggingface hub.
|
10 |
+
However, can take optional pair of arguments to a .safetensor model, and config file
|
11 |
+
RULES of the loader:
|
12 |
+
1. the model file must appear to be either in current directory or one down. So,
|
13 |
+
badpath1=some/directory/tree/file.here
|
14 |
+
badpath2=/absolutepath
|
15 |
+
2. yes, you MUST have a matching config.json file
|
16 |
+
3. if you have no alternative, you can get away with using pytorch_model.bin
|
17 |
+
|
18 |
+
Sample location for such things that you can download:
|
19 |
+
https://huggingface.co/stablediffusionapi/edge-of-realism/tree/main/text_encoder/
|
20 |
+
If there is a .safetensors AND a .bin file, ignore the .bin file
|
21 |
+
|
22 |
+
You can also convert a singlefile model, such as is downloaded from civitai,
|
23 |
+
by using the utility at
|
24 |
+
https://github.com/huggingface/diffusers/blob/main/scripts/convert_original_stable_diffusion_to_diffusers.py
|
25 |
+
Args should look like
|
26 |
+
convert_original_stable_diffusion_to_diffusers.py --checkpoint_file somemodel.safetensors \
|
27 |
+
--dump_path extractdir --to_safetensors --from_safetensors
|
28 |
+
|
29 |
"""
|
30 |
|
31 |
|
|
|
33 |
import json
|
34 |
import torch
|
35 |
from safetensors.torch import save_file
|
36 |
+
from transformers import CLIPProcessor,CLIPModel,CLIPTextModel
|
37 |
|
38 |
clipsrc="openai/clip-vit-large-patch14"
|
39 |
processor=None
|
40 |
model=None
|
41 |
+
encfile=None
|
42 |
+
configfile=None
|
43 |
+
|
44 |
+
if len(sys.argv) == 3:
|
45 |
+
encfile=sys.argv[1]
|
46 |
+
configfile=sys.argv[2]
|
47 |
|
48 |
device=torch.device("cuda")
|
49 |
|
|
|
51 |
def init():
|
52 |
global processor
|
53 |
global model
|
54 |
+
global encfile
|
55 |
+
global configfile
|
56 |
+
|
57 |
# Load the processor and model
|
58 |
print("loading processor from "+clipsrc,file=sys.stderr)
|
59 |
processor = CLIPProcessor.from_pretrained(clipsrc)
|
60 |
print("done",file=sys.stderr)
|
61 |
+
|
62 |
+
# originally done this way. But its not the right one to use
|
63 |
+
#print("loading model from "+clipsrc,file=sys.stderr)
|
64 |
+
#model = CLIPModel.from_pretrained(clipsrc)
|
65 |
+
#print("done",file=sys.stderr)
|
66 |
+
if encfile != None:
|
67 |
+
print("loading model from "+encfile,file=sys.stderr)
|
68 |
+
model = CLIPTextModel.from_pretrained(
|
69 |
+
encfile,config=configfile,local_files_only=True,use_safetensors=True
|
70 |
+
)
|
71 |
+
else:
|
72 |
+
print("loading model from "+clipsrc,file=sys.stderr)
|
73 |
+
model = CLIPTextModel.from_pretrained(clipsrc)
|
74 |
+
|
75 |
print("done",file=sys.stderr)
|
76 |
|
77 |
model = model.to(device)
|
78 |
|
79 |
|
80 |
+
# "inputs" == magic pre-embedding format
|
81 |
def embed_from_inputs(inputs):
|
82 |
with torch.no_grad():
|
83 |
+
# This way is for CLIPModel
|
84 |
+
#text_features = model.get_text_features(**inputs)
|
85 |
+
#embedding = text_features[0]
|
86 |
+
|
87 |
+
outputs = model(**inputs)
|
88 |
+
embeddings = outputs.pooler_output
|
89 |
+
embedding = embeddings
|
90 |
|
91 |
return embedding
|
92 |
|
|
|
101 |
inputs.input_ids[0][1]=id
|
102 |
|
103 |
emb=embed_from_inputs(inputs)
|
|
|
104 |
all_embeddings.append(emb)
|
105 |
if (id %100) ==0:
|
106 |
print(id)
|
107 |
|
108 |
embs = torch.cat(all_embeddings,dim=0)
|
109 |
print("Shape of result = ",embs.shape)
|
110 |
+
|
111 |
+
outputfile="cliptextmodel.temp.allids.safetensors"
|
112 |
+
print(f"Saving the calculatiuons to {outputfile}...")
|
113 |
+
save_file({"embeddings": embs}, outputfile)
|
114 |
|
115 |
|