ppbrown commited on
Commit
e346676
1 Parent(s): 3523378
Files changed (1) hide show
  1. generate-embeddingXL.py +40 -29
generate-embeddingXL.py CHANGED
@@ -3,7 +3,7 @@
3
  """ Work in progress
4
 
5
  Similar to generate-embedding.py, but outputs in the format
6
- that SDXL models expect. I hope.
7
 
8
  Also tries to load the SDXL base text encoder specifically.
9
  Requires you populate the two paths mentioned immediately below this comment section.
@@ -20,12 +20,12 @@ Plan:
20
 
21
  Note that you can generate an embedding from two words, or even more
22
 
23
- I could also include a "clip_l" key, but..
24
- Meh.
25
  """
26
 
27
- model_path = "text_encoder_xl.safetensors"
28
- model_config = "text_encoder_2_config.json"
 
 
29
 
30
  import sys
31
  import torch
@@ -36,17 +36,25 @@ from safetensors.torch import save_file
36
  # Note that it doesnt like a leading "/" in the name!!
37
 
38
 
39
- model=None
 
40
  processor=None
41
 
42
  device=torch.device("cuda")
43
 
 
 
 
 
 
 
 
44
  # Note the default, required 2 pathnames
45
- def initXLCLIPmodel():
46
- global model
47
  print("loading",model_path)
48
- model = CLIPTextModelWithProjection.from_pretrained(model_path,config=model_config,local_files_only=True,use_safetensors=True)
49
- model.to(device)
50
 
51
  # a bit wierd, but SDXL seems to still use this tokeninzer
52
  def initCLIPprocessor():
@@ -56,17 +64,27 @@ def initCLIPprocessor():
56
  processor = CLIPProcessor.from_pretrained(CLIPname)
57
 
58
  def embed_from_text(text):
59
- global processor,model
60
  if processor == None:
61
  initCLIPprocessor()
62
- initXLCLIPmodel()
63
- print("getting tokens")
 
 
 
 
 
 
 
 
 
 
 
64
  inputs = processor(text=text, return_tensors="pt")
65
  inputs.to(device)
66
 
67
- print("getting embeddings?")
68
- outputs = model(**inputs)
69
- print("finalizing")
70
  embeddings = outputs.text_embeds
71
  return embeddings
72
 
@@ -76,24 +94,17 @@ def embed_from_text(text):
76
 
77
  word = input("type a phrase to generate an embedding for: ")
78
 
79
- emb = embed_from_text(word)
80
- #embs=emb.unsqueeze(0) # stupid matrix magic to make it the required shape
81
- embs=emb
82
 
83
- print("Shape of result = ",embs.shape)
84
- # Note that programs like shapes such as
85
- # torch.Size([1, 768])
86
 
87
- output = "generatedXL.safetensors"
88
  # if single word used, then rename output file
89
  if all(char.isalpha() for char in word):
90
- output=f"{word}XL.safetensors"
91
  print(f"Saving to {output}...")
92
- save_file({"clip_g": embs}, output)
93
 
94
- # technically we are saving a shape ([1][1280])
95
- # whereas official XL embeddings files, are
96
- # (clip_g) shape ([8][1280])
97
- # (clip_l) shape ([8][768])
98
 
99
 
 
3
  """ Work in progress
4
 
5
  Similar to generate-embedding.py, but outputs in the format
6
+ that SDXL models expect.
7
 
8
  Also tries to load the SDXL base text encoder specifically.
9
  Requires you populate the two paths mentioned immediately below this comment section.
 
20
 
21
  Note that you can generate an embedding from two words, or even more
22
 
 
 
23
  """
24
 
25
+ model_path1 = "text_encoder.safetensors"
26
+ model_config1 = "text_encoder_config.json"
27
+ model_path2 = "text_encoder_2.safetensors"
28
+ model_config2 = "text_encoder_2_config.json"
29
 
30
  import sys
31
  import torch
 
36
  # Note that it doesnt like a leading "/" in the name!!
37
 
38
 
39
+ tmodel1=None
40
+ tmodel2=None
41
  processor=None
42
 
43
  device=torch.device("cuda")
44
 
45
+ def initCLIPmodel(model_path,model_config):
46
+ global tmodel1
47
+ print("loading",model_path)
48
+ tmodel1 = CLIPTextModel.from_pretrained(model_path,config=model_config,local_files_only=True,use_safetensors=True)
49
+ tmodel1.to(device)
50
+
51
+ #
52
  # Note the default, required 2 pathnames
53
+ def initXLCLIPmodel(model_path,model_config):
54
+ global tmodel2
55
  print("loading",model_path)
56
+ tmodel2 = CLIPTextModelWithProjection.from_pretrained(model_path,config=model_config,local_files_only=True,use_safetensors=True)
57
+ tmodel2.to(device)
58
 
59
  # a bit wierd, but SDXL seems to still use this tokeninzer
60
  def initCLIPprocessor():
 
64
  processor = CLIPProcessor.from_pretrained(CLIPname)
65
 
66
  def embed_from_text(text):
67
+ global processor,tmodel1
68
  if processor == None:
69
  initCLIPprocessor()
70
+ initCLIPmodel(model_path1,model_config1)
71
+ inputs = processor(text=text, return_tensors="pt")
72
+ inputs.to(device)
73
+
74
+ print("getting embeddings1")
75
+ outputs = tmodel1(**inputs)
76
+ embeddings = outputs.pooler_output
77
+ return embeddings
78
+
79
+ def embed_from_text2(text):
80
+ global processor,tmodel2
81
+ if tmodel2 == None:
82
+ initXLCLIPmodel(model_path2,model_config2)
83
  inputs = processor(text=text, return_tensors="pt")
84
  inputs.to(device)
85
 
86
+ print("getting embeddings2")
87
+ outputs = tmodel2(**inputs)
 
88
  embeddings = outputs.text_embeds
89
  return embeddings
90
 
 
94
 
95
  word = input("type a phrase to generate an embedding for: ")
96
 
97
+ emb1 = embed_from_text(word)
98
+ emb2 = embed_from_text2(word)
 
99
 
100
+ print("Shape of results = ",emb1.shape,emb2.shape)
 
 
101
 
102
+ output = "generated_XL.safetensors"
103
  # if single word used, then rename output file
104
  if all(char.isalpha() for char in word):
105
+ output=f"{word}_XL.safetensors"
106
  print(f"Saving to {output}...")
107
+ save_file({"clip_g": emb2,"clip_l":emb1}, output)
108
 
 
 
 
 
109
 
110