Update goai_helpers/goai_traduction.py
Browse files- goai_helpers/goai_traduction.py +18 -14
goai_helpers/goai_traduction.py
CHANGED
|
@@ -18,32 +18,36 @@ def goai_traduction(text, src_lang, tgt_lang):
|
|
| 18 |
if src_lang == "fra_Latn" and tgt_lang == "mos_Latn":
|
| 19 |
model_id = "ArissBandoss/nllb-200-distilled-600M-finetuned-fr-to-mos-V4"
|
| 20 |
elif src_lang == "mos_Latn" and tgt_lang == "fra_Latn":
|
| 21 |
-
model_id = "ArissBandoss/mos2fr-
|
| 22 |
else:
|
| 23 |
model_id = "ArissBandoss/nllb-200-distilled-600M-finetuned-fr-to-mos-V4"
|
| 24 |
-
|
| 25 |
tokenizer = AutoTokenizer.from_pretrained(model_id, token=auth_token)
|
| 26 |
model = AutoModelForSeq2SeqLM.from_pretrained(model_id, token=auth_token).to(device)
|
| 27 |
-
|
|
|
|
| 28 |
tokenizer.src_lang = src_lang
|
|
|
|
|
|
|
| 29 |
inputs = tokenizer(text, return_tensors="pt").to(device)
|
| 30 |
-
|
| 31 |
-
#
|
| 32 |
tgt_lang_id = tokenizer.convert_tokens_to_ids(tgt_lang)
|
| 33 |
-
|
| 34 |
-
# Génération
|
| 35 |
outputs = model.generate(
|
| 36 |
**inputs,
|
| 37 |
forced_bos_token_id=tgt_lang_id,
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
|
|
|
| 42 |
)
|
| 43 |
-
|
|
|
|
| 44 |
translation = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
|
| 45 |
-
|
| 46 |
-
print(translation)
|
| 47 |
return translation
|
| 48 |
|
| 49 |
def real_time_traduction(input_text, src_lang, tgt_lang):
|
|
|
|
| 18 |
if src_lang == "fra_Latn" and tgt_lang == "mos_Latn":
|
| 19 |
model_id = "ArissBandoss/nllb-200-distilled-600M-finetuned-fr-to-mos-V4"
|
| 20 |
elif src_lang == "mos_Latn" and tgt_lang == "fra_Latn":
|
| 21 |
+
model_id = "ArissBandoss/mos2fr-5B-800-fixed" # Modèle réparé
|
| 22 |
else:
|
| 23 |
model_id = "ArissBandoss/nllb-200-distilled-600M-finetuned-fr-to-mos-V4"
|
| 24 |
+
|
| 25 |
tokenizer = AutoTokenizer.from_pretrained(model_id, token=auth_token)
|
| 26 |
model = AutoModelForSeq2SeqLM.from_pretrained(model_id, token=auth_token).to(device)
|
| 27 |
+
|
| 28 |
+
# Configuration du tokenizer
|
| 29 |
tokenizer.src_lang = src_lang
|
| 30 |
+
|
| 31 |
+
# Tokenisation
|
| 32 |
inputs = tokenizer(text, return_tensors="pt").to(device)
|
| 33 |
+
|
| 34 |
+
# ID du token de langue cible
|
| 35 |
tgt_lang_id = tokenizer.convert_tokens_to_ids(tgt_lang)
|
| 36 |
+
|
| 37 |
+
# Génération avec les paramètres optimaux
|
| 38 |
outputs = model.generate(
|
| 39 |
**inputs,
|
| 40 |
forced_bos_token_id=tgt_lang_id,
|
| 41 |
+
max_new_tokens=1024,
|
| 42 |
+
num_beams=5,
|
| 43 |
+
early_stopping=False,
|
| 44 |
+
no_repeat_ngram_size=0,
|
| 45 |
+
length_penalty=1.0
|
| 46 |
)
|
| 47 |
+
|
| 48 |
+
# Décodage
|
| 49 |
translation = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
|
| 50 |
+
|
|
|
|
| 51 |
return translation
|
| 52 |
|
| 53 |
def real_time_traduction(input_text, src_lang, tgt_lang):
|