Util to find top token in each dimention
Browse files- generate-allid-toptokens.py +55 -0
generate-allid-toptokens.py
ADDED
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/env python
|
2 |
+
|
3 |
+
"""
|
4 |
+
CONCEPT:
|
5 |
+
Load in a precalculated embeddings file of all the tokenids (0-49405)
|
6 |
+
(see "generate-allid-embeddings[XL].py")
|
7 |
+
|
8 |
+
For each dimension, calculate which tokenid has the highest value.
|
9 |
+
Print out list, keyed by dimension.
|
10 |
+
|
11 |
+
In theory, this should auto-adjust, whether the embeddings file
|
12 |
+
is SD, or SDXL (clip_l or clip_g)
|
13 |
+
|
14 |
+
"""
|
15 |
+
|
16 |
+
|
17 |
+
import sys
|
18 |
+
import json
|
19 |
+
import torch
|
20 |
+
from safetensors import safe_open
|
21 |
+
|
22 |
+
file1=sys.argv[1]
|
23 |
+
file2=sys.argv[2]
|
24 |
+
|
25 |
+
print(f"reading in json from {file2} now",file=sys.stderr)
|
26 |
+
with open(file2, "r") as file:
|
27 |
+
json_data = json.load(file)
|
28 |
+
|
29 |
+
token_names = {v: k for k, v in json_data.items()}
|
30 |
+
|
31 |
+
#print(token_names)
|
32 |
+
|
33 |
+
device=torch.device("cuda")
|
34 |
+
print(f"reading {file1} embeddings now",file=sys.stderr)
|
35 |
+
model = safe_open(file1,framework="pt",device="cuda")
|
36 |
+
embs1=model.get_tensor("embeddings")
|
37 |
+
embs1.to(device)
|
38 |
+
print("Shape of loaded embeds =",embs1.shape)
|
39 |
+
|
40 |
+
|
41 |
+
print(f"calculating distances...",file=sys.stderr)
|
42 |
+
|
43 |
+
indices = torch.argmax(embs1, dim=0)
|
44 |
+
|
45 |
+
print("Shape of results=",indices.shape,file=sys.stderr)
|
46 |
+
|
47 |
+
indices=indices.tolist()
|
48 |
+
|
49 |
+
counter=0
|
50 |
+
for token_num in indices:
|
51 |
+
#print("num:",token_num)
|
52 |
+
print(counter,token_names.get(token_num))
|
53 |
+
counter+=1
|
54 |
+
|
55 |
+
|