ppbrown commited on
Commit
14b75c0
1 Parent(s): 431066d

Actually works properly

Browse files

prior versions were generated using CLIPModel.
Turns out thats the WRONG model. Now using the same code that SD actually uses: CLIPTextModel

Files changed (1) hide show
  1. generate-embedding.py +12 -9
generate-embedding.py CHANGED
@@ -3,7 +3,6 @@
3
  """ Work in progress
4
  NB: This is COMPLETELY DIFFERENT from "generate-embeddings.py"!!!
5
 
6
-
7
  Plan:
8
  Take input for a single word or phrase.
9
  Generate a embedding file, "generated.safetensors"
@@ -20,7 +19,7 @@ import sys
20
  import json
21
  import torch
22
  from safetensors.torch import save_file
23
- from transformers import CLIPProcessor,CLIPModel
24
 
25
  import logging
26
  # Turn off stupid mesages from CLIPModel.load
@@ -41,18 +40,18 @@ def init():
41
  processor = CLIPProcessor.from_pretrained(clipsrc)
42
  print("done",file=sys.stderr)
43
  print("loading model from "+clipsrc,file=sys.stderr)
44
- model = CLIPModel.from_pretrained(clipsrc)
45
  print("done",file=sys.stderr)
46
 
47
  model = model.to(device)
48
 
49
- def standard_embed_calc(text):
50
  inputs = processor(text=text, return_tensors="pt")
51
  inputs.to(device)
52
  with torch.no_grad():
53
- text_features = model.get_text_features(**inputs)
54
- embedding = text_features[0]
55
- return embedding
56
 
57
 
58
  init()
@@ -60,11 +59,15 @@ init()
60
 
61
  word = input("type a phrase to generate an embedding for: ")
62
 
63
- emb = standard_embed_calc(word)
64
- embs=emb.unsqueeze(0) # stupid matrix magic to make the cat work
 
65
 
66
  print("Shape of result = ",embs.shape)
 
67
  output = "generated.safetensors"
 
 
68
  print(f"Saving to {output}...")
69
  save_file({"emb_params": embs}, output)
70
 
 
3
  """ Work in progress
4
  NB: This is COMPLETELY DIFFERENT from "generate-embeddings.py"!!!
5
 
 
6
  Plan:
7
  Take input for a single word or phrase.
8
  Generate a embedding file, "generated.safetensors"
 
19
  import json
20
  import torch
21
  from safetensors.torch import save_file
22
+ from transformers import CLIPProcessor,CLIPTextModel
23
 
24
  import logging
25
  # Turn off stupid mesages from CLIPModel.load
 
40
  processor = CLIPProcessor.from_pretrained(clipsrc)
41
  print("done",file=sys.stderr)
42
  print("loading model from "+clipsrc,file=sys.stderr)
43
+ model = CLIPTextModel.from_pretrained(clipsrc)
44
  print("done",file=sys.stderr)
45
 
46
  model = model.to(device)
47
 
48
+ def cliptextmodel_embed_calc(text):
49
  inputs = processor(text=text, return_tensors="pt")
50
  inputs.to(device)
51
  with torch.no_grad():
52
+ outputs = model(**inputs)
53
+ embeddings = outputs.pooler_output
54
+ return embeddings
55
 
56
 
57
  init()
 
59
 
60
  word = input("type a phrase to generate an embedding for: ")
61
 
62
+ emb = cliptextmodel_embed_calc(word)
63
+ #embs=emb.unsqueeze(0) # stupid matrix magic to make it the required shape
64
+ embs=emb
65
 
66
  print("Shape of result = ",embs.shape)
67
+
68
  output = "generated.safetensors"
69
+ if all(char.isalpha() for char in word):
70
+ output=f"{word}.safetensors"
71
  print(f"Saving to {output}...")
72
  save_file({"emb_params": embs}, output)
73