ppbrown commited on
Commit
bac0dfe
1 Parent(s): c992dd1

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +18 -7
README.md CHANGED
@@ -1,12 +1,9 @@
1
- ---
2
- license: cc-by-4.0
3
- ---
4
  # tokenspace directory
5
 
6
  This directory contains utilities for the purpose of browsing the
7
  "token space" of CLIP ViT-L/14
8
 
9
- Long term goal is to be able to literally browse from word to
10
  "nearby" word
11
 
12
 
@@ -15,15 +12,29 @@ Long term goal is to be able to literally browse from word to
15
 
16
  Generates the "embeddings.safetensor" file!
17
 
18
- Basically goes through the fullword.json file, and
19
  generates a standalone embedding object for each word.
20
  Shape of the embeddings tensor, is
21
  [number-of-words][768]
22
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
23
  ## fullword.json
24
 
25
  This file contains a collection of "one word, one CLIP token id" pairings.
26
  The file was taken from vocab.sdturbo.json
27
- First all the non-</w> entries were stripped out.
28
  Then all the garbage punctuation and foreign characters were stripped out.
29
- Finally, the actual </w> was stripped out, for ease of use.
 
 
 
 
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
+ Long term goal is to be able to literally browse from word to
7
  "nearby" word
8
 
9
 
 
12
 
13
  Generates the "embeddings.safetensor" file!
14
 
15
+ Basically goes through the fullword.json file, and
16
  generates a standalone embedding object for each word.
17
  Shape of the embeddings tensor, is
18
  [number-of-words][768]
19
 
20
+ Note that it is possible to directly pull a tensor from the CLIP model,
21
+ key of text_model.embeddings.token_embedding.weight
22
+
23
+ This will NOT GIVE YOU THE RIGHT DISTANCES!
24
+ Hence why we are calculating and then storing the embedding weights actually
25
+ used by the CLIP process
26
+
27
+
28
+ ## generate-distances.py
29
+
30
+ Loads the prior generated embeddings, and then tries to calculate a full matrix
31
+ of distances between all tokens
32
+
33
+
34
  ## fullword.json
35
 
36
  This file contains a collection of "one word, one CLIP token id" pairings.
37
  The file was taken from vocab.sdturbo.json
38
+ First all the non-(/w) entries were stripped out.
39
  Then all the garbage punctuation and foreign characters were stripped out.
40
+ Finally, the actual (/w) was stripped out, for ease of use.