File size: 2,874 Bytes
d120873
 
 
 
37ded96
 
d120873
 
 
d22cb09
 
 
 
d120873
d22cb09
 
95b5309
 
d22cb09
b8db721
95b5309
 
 
49ae858
d22cb09
 
 
 
 
 
 
 
 
49ae858
95b5309
 
 
d22cb09
 
 
 
 
 
 
 
 
 
 
 
37ded96
 
2e5046d
088720e
e7d7478
088720e
 
2e5046d
 
 
 
e7d7478
088720e
 
 
e7d7478
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
from nltk.tokenize import sent_tokenize
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch
import src.exception.Exception.Exception as ExceptionCustom
# Use a pipeline as a high-level helper
from transformers import pipeline

METHOD = "TRANSLATE"

def paraphraseTranslateMethod(requestValue: str, model: str):
    exception = ExceptionCustom.checkForException(requestValue, METHOD)
    if exception:
        return "", exception

    tokenized_sent_list = sent_tokenize(requestValue)
    result_value = []
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    
    for SENTENCE in tokenized_sent_list:
        if model == 'roen':
            tokenizerROMENG = AutoTokenizer.from_pretrained("BlackKakapo/opus-mt-ro-en")
            modelROMENG = AutoModelForSeq2SeqLM.from_pretrained("BlackKakapo/opus-mt-ro-en")
            modelROMENG.to(device)
            input_ids = tokenizerROMENG(SENTENCE, return_tensors='pt').to(device)
            output = modelROMENG.generate(
                input_ids=input_ids.input_ids,
                do_sample=True,
                max_length=512,
                top_k=90,
                top_p=0.97,
                early_stopping=False
            )
            result = tokenizerROMENG.batch_decode(output, skip_special_tokens=True)[0]
        else:
            tokenizerENGROM = AutoTokenizer.from_pretrained("BlackKakapo/opus-mt-en-ro")
            modelENGROM = AutoModelForSeq2SeqLM.from_pretrained("BlackKakapo/opus-mt-en-ro")
            modelENGROM.to(device)
            input_ids = tokenizerENGROM(SENTENCE, return_tensors='pt').to(device)
            output = modelENGROM.generate(
                input_ids=input_ids.input_ids,
                do_sample=True,
                max_length=512,
                top_k=90,
                top_p=0.97,
                early_stopping=False
            )
            result = tokenizerENGROM.batch_decode(output, skip_special_tokens=True)[0]
        result_value.append(result)

    return " ".join(result_value).strip(), model

def gemma(requestValue: str, model: str = 'Gargaz/gemma-2b-romanian-better'):
    prompt = f"Translate this to Romanian using a formal tone. Only return the translation: {requestValue}"
    messages = [{"role": "user", "content": f"Translate this text to Romanian using a formal tone. Only return the translated text: {requestValue}"}]
    if '/' not in model:
        model = 'Gargaz/gemma-2b-romanian-better'
    pipe = pipeline(
        "text-generation",
        model=model,
        device=-1,
        max_new_tokens=256,          # Keep short to reduce verbosity
        do_sample=False            # Use greedy decoding for determinism
    ) 
    output = pipe(messages, num_return_sequences=1, return_full_text=False)
    # return output[0]["generated_text"].strip(), model
    return output, model