ppbrown commited on
Commit
b3f727f
1 Parent(s): 4cb0a6c

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +90 -14
README.md CHANGED
@@ -1,18 +1,94 @@
1
- This directory specializes in the Google T5 xxl LLM.
2
 
3
- Specifically, it focuses on
4
- https://huggingface.co/mcmonkey/google_t5-v1_1-xxl_encoderonly/
5
- because that is the one most used by AI generative models at the moment
6
 
7
- * dictionary.T5.fullword
8
- This file is filtered from the full list of tokens in this model
9
- (generated by dumptokens.py),
10
- and has only the tokens that are full standalone words
11
- (and are also plain English ascii. Sorry, fancy languages)
12
 
13
- * dictionary.both
14
- Words that are common across CLIP-L and T5. Generated by
15
- sort dictionary.T5.fullword ../dictionary.fullword |uniq -c |awk '$1 =="2"{print $2}' >dict.both
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
16
 
17
- * showtokens.py
18
- Given a word, or string of words, shows how T5 tokenizes it
 
1
+ # tokenspace directory
2
 
3
+ This directory contains utilities for the purpose of browsing the
4
+ "token space" of CLIP ViT-L/14
 
5
 
6
+ Primary tools are:
 
 
 
 
7
 
8
+ * "calculate-distances.py": allows command-line browsing of words and their neighbours
9
+ * "graph-embeddings.py": plots graph of full values of two embeddings
10
+
11
+
12
+ ## (clipmodel,cliptextmodel)-calculate-distances.py
13
+
14
+ Loads the generated embeddings, reads in a word, calculates "distance" to every
15
+ embedding, and then shows the closest "neighbours".
16
+
17
+ To run this requires the files "embeddings.safetensors" and "dictionary",
18
+ in matching format
19
+
20
+ You will need to rename or copy appropriate files for this as mentioned below.
21
+
22
+ Note that SD models use cliptextmodel, NOT clipmodel
23
+
24
+ ## graph-textmodels.py
25
+
26
+ Shows the difference between the same word, embedded by CLIPTextModel
27
+ vs CLIPModel
28
+
29
+ ## graph-embeddings.py
30
+
31
+ Run the script. It will ask you for two text strings.
32
+ Once you enter both, it will plot the graph and display it for you
33
+
34
+ Note that this tool does not require any of the other files; just that you
35
+ have the requisite python modules installed. (pip install -r requirements.txt)
36
+
37
+ ### embeddings.safetensors
38
+
39
+ You can either copy one of the provided files, or generate your own.
40
+ See generate-embeddings.py for that.
41
+
42
+ Note that you muist always use the "dictionary" file that matchnes your embeddings file
43
+
44
+ ### embeddings.allids.safetensors
45
+
46
+ DO NOT USE THIS ONE for programs that expect a matching dictionary.
47
+ This one is purely numeric based.
48
+ Its intention is more for research datamining, but it does have a matching
49
+ graph front end, graph-byid.py
50
+
51
+
52
+ ### dictionary
53
+
54
+ Make sure to always use the dictionary file that matches your embeddings file.
55
+
56
+ The "dictionary.fullword" file is pulled from fullword.json, which is distilled from "full words"
57
+ present in the ViT-L/14 CLIP model's provided token dictionary, called "vocab.json".
58
+ Thus there are only around 30,000 words in it
59
+
60
+ If you want to use the provided "embeddings.safetensors.huge" file, you will want to use the matching
61
+ "dictionary.huge" file, which has over 300,000 words
62
+
63
+ This huge file comes from the linux "wamerican-huge" package, which delivers it under
64
+ /usr/share/dict/american-english-huge
65
+
66
+ There also exists a "american-insane" package
67
+
68
+
69
+ ## generate-embeddings.py
70
+
71
+ Generates the "embeddings.safetensor" file, based on the "dictionary" file present.
72
+ Takes a few minutes to run, depending on size of the dictionary
73
+
74
+ The shape of the embeddings tensor, is
75
+ [number-of-words][768]
76
+
77
+ Note that yes, it is possible to directly pull a tensor from the CLIP model,
78
+ using keyname of text_model.embeddings.token_embedding.weight
79
+
80
+ This will NOT GIVE YOU THE RIGHT DISTANCES!
81
+ Hence why we are calculating and then storing the embedding weights actually
82
+ generated by the CLIP process
83
+
84
+
85
+ ## fullword.json
86
+
87
+ This file contains a collection of "one word, one CLIP token id" pairings.
88
+ The file was taken from vocab.json, which is part of multiple SD models in huggingface.co
89
+
90
+ The file was optimized for what people are actually going to type as words.
91
+ First all the non-(/w) entries were stripped out.
92
+ Then all the garbage punctuation and foreign characters were stripped out.
93
+ Finally, the actual (/w) was stripped out, for ease of use.
94