|
|
import torch |
|
|
import spaces |
|
|
from transformers import pipeline, AutoModelForSeq2SeqLM, AutoTokenizer, GenerationConfig |
|
|
from peft import PeftModel, PeftConfig |
|
|
import os |
|
|
import unicodedata |
|
|
from huggingface_hub import login |
|
|
|
|
|
max_length = 512 |
|
|
auth_token = os.getenv('HF_SPACE_TOKEN') |
|
|
login(token=auth_token) |
|
|
|
|
|
|
|
|
@spaces.GPU |
|
|
def goai_traduction(text, src_lang, tgt_lang): |
|
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu") |
|
|
|
|
|
if src_lang == "fra_Latn" and tgt_lang == "mos_Latn": |
|
|
model_id = "ArissBandoss/nllb-200-distilled-600M-finetuned-fr-to-mos-V4" |
|
|
elif src_lang == "mos_Latn" and tgt_lang == "fra_Latn": |
|
|
model_id = "ArissBandoss/3b-new-400" |
|
|
else: |
|
|
model_id = "ArissBandoss/nllb-200-distilled-600M-finetuned-fr-to-mos-V4" |
|
|
|
|
|
print(f"Chargement du modèle: {model_id}") |
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id, token=auth_token) |
|
|
model = AutoModelForSeq2SeqLM.from_pretrained(model_id, token=auth_token).to(device) |
|
|
|
|
|
print(f"Texte brut ({len(text)} caractères / {len(text.split())} mots):") |
|
|
print(text) |
|
|
|
|
|
print(f"Configuration du modèle:") |
|
|
print(f"- tokenizer.model_max_length: {tokenizer.model_max_length}") |
|
|
print(f"- Position embeddings shape: {model.model.encoder.embed_positions.weights.shape}") |
|
|
print(f"- decoder.embed_positions shape: {model.model.decoder.embed_positions.weights.shape}") |
|
|
|
|
|
|
|
|
tokenizer.src_lang = src_lang |
|
|
|
|
|
|
|
|
inputs = tokenizer(text, return_tensors="pt", truncation=False).to(device) |
|
|
input_ids = inputs["input_ids"][0] |
|
|
|
|
|
print("Tokens d'entrée:") |
|
|
print(f"- Nombre de tokens: {input_ids.shape[0]}") |
|
|
print(f"- Premiers tokens: {input_ids[:10].tolist()}") |
|
|
print(f"- Derniers tokens: {input_ids[-10:].tolist()}") |
|
|
|
|
|
|
|
|
tgt_lang_id = tokenizer.convert_tokens_to_ids(tgt_lang) |
|
|
print(f"Token ID de la langue cible ({tgt_lang}): {tgt_lang_id}") |
|
|
|
|
|
for length_penalty in [1.0, 1.5, 2.0]: |
|
|
for num_beams in [5, 10]: |
|
|
print(f"\nTest avec length_penalty={length_penalty}, num_beams={num_beams}") |
|
|
|
|
|
|
|
|
outputs = model.generate( |
|
|
**inputs, |
|
|
forced_bos_token_id=tgt_lang_id, |
|
|
max_new_tokens=2048, |
|
|
early_stopping=False, |
|
|
num_beams=num_beams, |
|
|
no_repeat_ngram_size=0, |
|
|
bad_words_ids=bad_words_ids, |
|
|
length_penalty=length_penalty |
|
|
) |
|
|
|
|
|
|
|
|
translation = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0] |
|
|
|
|
|
print(f"Traduction ({len(translation)} caractères / {len(translation.split())} mots):") |
|
|
print(translation) |
|
|
output_ids = outputs[0] |
|
|
print(f"- Nombre de tokens générés: {output_ids.shape[0]}") |
|
|
print(f"- Premiers tokens générés: {output_ids[:10].tolist()}") |
|
|
print(f"- Derniers tokens générés: {output_ids[-10:].tolist()}") |
|
|
|
|
|
return translation |
|
|
|
|
|
def real_time_traduction(input_text, src_lang, tgt_lang): |
|
|
return goai_traduction(input_text, src_lang, tgt_lang) |
