File size: 7,353 Bytes
4c6879b |
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 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 |
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import sys
import os
# Get the absolute path of IndicTransToolkit
indictrans_path = "/content/Voice-to-Text-Translation-System-Leveraging-Whisper-and-IndicTrans2/IndicTrans2/huggingface_interface/IndicTransToolkit/IndicTransToolkit"
sys.path.append(indictrans_path)
from processor import IndicProcessor
# Check if GPU is available
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
def translate_text(transcription, target_lang, src_lang):
mapping = {
"Assamese": "asm_Beng", "Bengali": "ben_Beng", "Bodo": "brx_Deva", "Dogri": "doi_Deva",
"Gujarati": "guj_Gujr", "Hindi": "hin_Deva", "Kannada": "kan_Knda",
"Kashmiri(Perso-Arabic script)": "kas_Arab", "Kashmiri(Devanagari script)": "kas_Deva",
"Konkani": "kok_Deva", "Maithili": "mai_Deva", "Malayalam": "mal_Mlym",
"Manipuri(Bengali script)": "mni_Beng", "Manipuri(Meitei script)": "mni_Mtei",
"Marathi": "mar_Deva", "Nepali": "nep_Deva", "Odia": "ory_Orya",
"Punjabi": "pan_Guru", "Sanskrit": "san_Deva", "Santali(Ol Chiki script)": "sat_Olck",
"Sindhi(Perso-Arabic script)": "snd_Arab", "Sindhi(Devanagari script)": "snd_Deva",
"Tamil": "tam_Taml", "Telugu": "tel_Telu", "Urdu": "urd_Arab","English":"eng_Latn",
}
if target_lang in mapping:
tgt_lang = mapping[target_lang]
if src_lang == tgt_lang:
return "Detected Language and Target Language cannot be same"
if src_lang == "eng_Latn":
model_name = "prajdabre/rotary-indictrans2-en-indic-1B"
else:
model1_name ="prajdabre/rotary-indictrans2-indic-en-1B"
model2_name = "prajdabre/rotary-indictrans2-en-indic-1B"
translations = indic_indic(model1_name,model2_name, src_lang, target_lang,transcription)
return translations
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# Load model in 8-bit quantization
model = AutoModelForSeq2SeqLM.from_pretrained(
model_name,
trust_remote_code=True,
torch_dtype=torch.float16,
#load_in_8bit=True,
attn_implementation="flash_attention_2"
).to(DEVICE)
ip = IndicProcessor(inference=True)
input_sentences = [transcription]
batch = ip.preprocess_batch(input_sentences, src_lang=src_lang, tgt_lang=tgt_lang)
# Tokenize the sentences and generate input encodings
inputs = tokenizer(
batch,
truncation=True,
padding="longest",
return_tensors="pt",
max_length=2048,
)
# Move inputs to the correct device (only inputs, NOT model)
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
# Generate translations using the model
with torch.inference_mode():
generated_tokens = model.generate(
**inputs,
num_beams=5,
length_penalty=1.5,
repetition_penalty=2.0,
num_return_sequences=1,
max_new_tokens=2048,
early_stopping=True
)
# Move generated tokens to CPU before decoding
generated_tokens = generated_tokens.cpu().tolist()
# Decode the generated tokens into text
with tokenizer.as_target_tokenizer():
generated_tokens = tokenizer.batch_decode(
generated_tokens,
skip_special_tokens=True,
clean_up_tokenization_spaces=True
)
# Postprocess the translations
translations = ip.postprocess_batch(generated_tokens, lang=tgt_lang)
print(type(translations))
translations =str(translations).strip("'")
return translations
def indic_indic(model1_name,model2_name,src_lang,tgt_lang,transcription,intermediate_lng ="eng_Latn",):
tokenizer = AutoTokenizer.from_pretrained(model1_name, trust_remote_code=True)
# Load model in 8-bit quantization
model = AutoModelForSeq2SeqLM.from_pretrained(
model1_name,
trust_remote_code=True,
torch_dtype=torch.float16,
#load_in_8bit=True,
attn_implementation="flash_attention_2"
).to(DEVICE)
ip = IndicProcessor(inference=True)
input_sentences = [transcription]
batch = ip.preprocess_batch(input_sentences, src_lang=src_lang, tgt_lang=intermediate_lng)
# Tokenize the sentences and generate input encodings
inputs = tokenizer(
batch,
truncation=True,
padding="longest",
return_tensors="pt",
max_length=2048,
)
# Move inputs to the correct device (only inputs, NOT model)
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
# Generate translations using the model
with torch.inference_mode():
generated_tokens = model.generate(
**inputs,
num_beams=10,
length_penalty=1.5,
repetition_penalty=2.0,
num_return_sequences=1,
max_new_tokens=2048,
early_stopping=True
)
# Move generated tokens to CPU before decoding
generated_tokens = generated_tokens.cpu().tolist()
# Decode the generated tokens into text
with tokenizer.as_target_tokenizer():
generated_tokens = tokenizer.batch_decode(
generated_tokens,
skip_special_tokens=True,
clean_up_tokenization_spaces=True
)
# Postprocess the translations
translations1 = ip.postprocess_batch(generated_tokens, lang=tgt_lang)
translations1 =str(translations).strip("'")
tokenizer = AutoTokenizer.from_pretrained(model2_name, trust_remote_code=True)
# Load model in 8-bit quantization
model = AutoModelForSeq2SeqLM.from_pretrained(
model2_name,
trust_remote_code=True,
torch_dtype=torch.float16,
#load_in_8bit=True,
attn_implementation="flash_attention_2"
).to(DEVICE)
ip = IndicProcessor(inference=True)
input_sentences = [translations1]
batch = ip.preprocess_batch(input_sentences, src_lang=intermediate_lng, tgt_lang=tgt_lang)
# Tokenize the sentences and generate input encodings
inputs = tokenizer(
batch,
truncation=True,
padding="longest",
return_tensors="pt",
max_length=2048,
)
# Move inputs to the correct device (only inputs, NOT model)
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
# Generate translations using the model
with torch.inference_mode():
generated_tokens = model.generate(
**inputs,
num_beams=10,
length_penalty=1.5,
repetition_penalty=2.0,
num_return_sequences=1,
max_new_tokens=2048,
early_stopping=True
)
# Move generated tokens to CPU before decoding
generated_tokens = generated_tokens.cpu().tolist()
# Decode the generated tokens into text
with tokenizer.as_target_tokenizer():
generated_tokens = tokenizer.batch_decode(
generated_tokens,
skip_special_tokens=True,
clean_up_tokenization_spaces=True
)
# Postprocess the translations
translations = ip.postprocess_batch(generated_tokens, lang=tgt_lang)
return translations
|