Upload T5/dumptokens.py with huggingface_hub
Browse files- T5/dumptokens.py +53 -0
T5/dumptokens.py
ADDED
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/env python
|
2 |
+
|
3 |
+
"""
|
4 |
+
According to model,
|
5 |
+
vocab_size: 32128
|
6 |
+
|
7 |
+
But it actually caps out at 32099
|
8 |
+
"""
|
9 |
+
|
10 |
+
|
11 |
+
from transformers import T5Tokenizer,T5EncoderModel
|
12 |
+
import torch
|
13 |
+
import charade
|
14 |
+
|
15 |
+
T="mcmonkey/google_t5-v1_1-xxl_encoderonly"
|
16 |
+
|
17 |
+
tokenizer = T5Tokenizer.from_pretrained(T)
|
18 |
+
|
19 |
+
startword= tokenizer.convert_ids_to_tokens(3)
|
20 |
+
#print (startword)
|
21 |
+
|
22 |
+
# id should be a numeral
|
23 |
+
def print_token_from_id(id):
|
24 |
+
decoded_tokens = tokenizer.convert_ids_to_tokens(id)
|
25 |
+
print(decoded_tokens+" : " + str(id))
|
26 |
+
|
27 |
+
# print if it has the marker indicating it is a standalone word,
|
28 |
+
# not just a building block
|
29 |
+
def print_if_word(id):
|
30 |
+
decoded_tokens = tokenizer.convert_ids_to_tokens(id)
|
31 |
+
if decoded_tokens.startswith(startword):
|
32 |
+
print(decoded_tokens[1:] +" : " + str(id))
|
33 |
+
|
34 |
+
# standalone word, AND doesnt have any foreign non-ascii7 chars
|
35 |
+
def print_if_asciiword(id):
|
36 |
+
decoded_tokens = tokenizer.convert_ids_to_tokens(id)
|
37 |
+
if decoded_tokens.startswith(startword):
|
38 |
+
aword=decoded_tokens[1:]
|
39 |
+
if len(aword) <1:
|
40 |
+
return
|
41 |
+
estr=str(aword.encode())
|
42 |
+
if '\\x' in estr:
|
43 |
+
return
|
44 |
+
|
45 |
+
print(aword +" : " , id)
|
46 |
+
|
47 |
+
for id in range(4,32099):
|
48 |
+
#print_token_from_id(id)
|
49 |
+
#print_if_word(id)
|
50 |
+
print_if_asciiword(id)
|
51 |
+
|
52 |
+
|
53 |
+
|