Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,89 @@
|
|
1 |
-
---
|
2 |
-
license: apache-2.0
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
datasets:
|
4 |
+
- cc100
|
5 |
+
language:
|
6 |
+
- tr
|
7 |
+
library_name: peft
|
8 |
+
pipeline_tag: token-classification
|
9 |
+
---
|
10 |
+
|
11 |
+
```python
|
12 |
+
from peft import PeftModel, prepare_model_for_kbit_training
|
13 |
+
from transformers import T5ForTokenClassification, BitsAndBytesConfig
|
14 |
+
import torch
|
15 |
+
|
16 |
+
model_id = "google/byt5-small"
|
17 |
+
|
18 |
+
bnb_config = BitsAndBytesConfig(
|
19 |
+
load_in_4bit=True,
|
20 |
+
bnb_4bit_quant_type="nf4",
|
21 |
+
bnb_4bit_use_double_quant=True,
|
22 |
+
bnb_4bit_compute_dtype=torch.bfloat16,
|
23 |
+
)
|
24 |
+
|
25 |
+
model = T5ForTokenClassification.from_pretrained(model_id,
|
26 |
+
num_labels=2,
|
27 |
+
torch_dtype=torch.bfloat16,
|
28 |
+
quantization_config=bnb_config,
|
29 |
+
device_map="auto",)
|
30 |
+
model = prepare_model_for_kbit_training(model)
|
31 |
+
model = PeftModel.from_pretrained(model, 'bite-the-byte/byt5-small-deASCIIfy-TR')
|
32 |
+
|
33 |
+
def test_mask(data):
|
34 |
+
"""
|
35 |
+
Masks the padded tokens in the input.
|
36 |
+
Args:
|
37 |
+
data (list): List of strings.
|
38 |
+
Returns:
|
39 |
+
dataset (list): List of dictionaries.
|
40 |
+
"""
|
41 |
+
|
42 |
+
dataset = list()
|
43 |
+
for sample in data:
|
44 |
+
new_sample = dict()
|
45 |
+
|
46 |
+
input_tokens = [i + 3 for i in sample.encode('utf-8')]
|
47 |
+
input_tokens.append(0) # eos token
|
48 |
+
new_sample['input_ids'] = torch.tensor([input_tokens], dtype=torch.int64)
|
49 |
+
|
50 |
+
# Create attention mask
|
51 |
+
attention_mask = [1] * len(input_tokens) # Attend to all tokens
|
52 |
+
new_sample['attention_mask'] = torch.tensor([attention_mask], dtype=torch.int64)
|
53 |
+
|
54 |
+
dataset.append(new_sample)
|
55 |
+
|
56 |
+
return dataset
|
57 |
+
|
58 |
+
def rewrite(model, data):
|
59 |
+
"""
|
60 |
+
Rewrites the input text with the model.
|
61 |
+
Args:
|
62 |
+
model (torch.nn.Module): Model.
|
63 |
+
data (dict): Dictionary containing 'input_ids' and 'attention_mask'.
|
64 |
+
Returns:
|
65 |
+
output (str): Rewritten text.
|
66 |
+
"""
|
67 |
+
|
68 |
+
with torch.no_grad():
|
69 |
+
data = {k: v.to(model.device) for k, v in data.items()}
|
70 |
+
pred = torch.argmax(model(**data).logits, dim=2)
|
71 |
+
|
72 |
+
output = list() # save the indices of the characters as list of integers
|
73 |
+
|
74 |
+
# Conversion table for Turkish characters {100: [300, 350], ...}
|
75 |
+
en2tr = {en: tr for tr, en in zip(list(map(list, map(str.encode, list('ÜİĞŞÇÖüığşçö')))), list(map(ord, list('UIGSCOuigsco'))))}
|
76 |
+
|
77 |
+
for inp, lab in zip((data['input_ids'] - 3)[0].tolist(), pred[0].tolist()):
|
78 |
+
if lab and inp in en2tr:
|
79 |
+
# if the model predicts a diacritic, replace it with the corresponding Turkish character
|
80 |
+
output.extend(en2tr[inp])
|
81 |
+
elif inp >= 0: output.append(inp)
|
82 |
+
return bytes(output).decode()
|
83 |
+
|
84 |
+
def try_it(text, model):#=model):
|
85 |
+
sample = test_mask([text])
|
86 |
+
return rewrite(model, sample[0])
|
87 |
+
|
88 |
+
try_it('Cekoslovakyalilastiramadiklarimizdan misiniz?', model)
|
89 |
+
```
|