Add the M2M100 model from Facebook

app.py

@@ -7,25 +7,25 @@ Author: Didier Guillevic
 Date: 2024-09-07
 """
 import spaces
-import torch
 import gradio as gr
 import langdetect
+from typing import List
 
 import logging
 logger = logging.getLogger(__name__)
 logging.basicConfig(level=logging.INFO)
 
+from deep_translator import GoogleTranslator
+from model_spacy import nlp_xx
+
 import model_translation as translation
-from model_translation import tokenizer_multilingual, model_multilingual
-from model_translation import tokenizer_m2m100, model_m2m100
 
-from deep_translator import GoogleTranslator
 
-from model_spacy import nlp_xx
+def detect_language(text):
+    lang = langdetect.detect(text)
+    return lang
+
 
-#
-# Translate given input text
-#
 def build_text_chunks(text, src_lang, sents_per_chunk):
     """
     Given a text:
@@ -71,18 +71,21 @@ def build_text_chunks(text, src_lang, sents_per_chunk):
 
     return chunks
 
-def translate_with_model(
-        text, tokenizer, model, src_lang, sents_per_chunk,
-        input_max_length=512, output_max_length=512):
-    
-    # Build text chunks (using sents_per_chunk and translation.max_words_per_chunk)
-    chunks = build_text_chunks(text, src_lang, sents_per_chunk)
-    logger.info(f"LANG: {src_lang}, TEXT: {text[:20]}, NB_CHUNKS: {len(chunks)}")
 
-    # Translate chunks
+def translate_with_Helsinki(
+        chunks, src_lang, tgt_lang, input_max_length, output_max_length) -> str:
+    """Translate the chunks with the Helsinki model
+    """
+    if src_lang not in translation.src_langs:
+        return (
+            f"ISSUE: currently no model for language '{src_lang}'. "
+             "If wrong language, please specify language."
+        )
+    logger.info(f"LANG: {src_lang}, TEXT: {chunks[0][:50]}...")
+    tokenizer, model = translation.get_tokenizer_model_for_src_lang(src_lang)
+
     translated_chunks = []
     for chunk in chunks:
-
         # NOTE: The 'fa' (Persian) model has multiple target languages to choose from.
         # We need to specifiy the desired languages among: fra ita por ron spa
         #   https://huggingface.co/Helsinki-NLP/opus-mt-tc-big-fa-itc
@@ -91,92 +94,61 @@ def translate_with_model(
             chunk = ">>fra<< " + chunk
 
         inputs = tokenizer(
-            chunk, return_tensors="pt",
-            max_length=input_max_length,
+            chunk, return_tensors="pt", max_length=input_max_length,
             truncation=True, padding="longest").to(model.device)
-        
-        outputs = model.generate(
-            **inputs,
-            max_length=output_max_length)
-        
+        outputs = model.generate(**inputs, max_length=output_max_length)
         translated_chunk = tokenizer.batch_decode(
             outputs, skip_special_tokens=True)[0]
-        
         #logger.info(f"Text: {chunk}")
         #logger.info(f"Translation: {translated_chunk}")
-        
         translated_chunks.append(translated_chunk)
 
     return '\n'.join(translated_chunks)
 
 
-def detect_language(text):
-    lang = langdetect.detect(text)
-    return lang
-
-
-def translate_with_bilingual_model(
-        text, src_lang, tgt_lang, sents_per_chunk
-    ):
-    """
-    Translate with Helsinki bilingual models
-    """
-    if src_lang not in translation.src_langs:
-        return (
-            f"ISSUE: currently no model for language '{src_lang}'. "
-             "If wrong language, please specify language."
-        )
-    logger.info(f"LANG: {src_lang}, TEXT: {text[:50]}...")
-    tokenizer, model = translation.get_tokenizer_model_for_src_lang(src_lang)
-    translated_text_bilingual_model = translate_with_model(
-        text, tokenizer, model, src_lang, sents_per_chunk)
-    return translated_text_bilingual_model
-
-
 @spaces.GPU
-def translate_with_m2m100_model(
-        text: str,
+def translate_with_m2m100(
+        chunks: List[str],
         src_lang: str,
-        tgt_lang: str,
-        sents_per_chunk: int=5):
-    """
-    Translate with the m2m100 model
+        tgt_lang: str) -> str:
+    """Translate with the m2m100 model
     """
-    tokenizer_m2m100.src_lang = src_lang
-    input_ids = tokenizer_m2m100(
-        text, return_tensors="pt").input_ids.to(model_m2m100.device)
-    outputs = model_m2m100.generate(
-        input_ids=input_ids,
-        forced_bos_token_id=tokenizer_m2m100.get_lang_id(tgt_lang))
-    translated_text = tokenizer_m2m100.batch_decode(
-        outputs[0], skip_special_tokens=True)
-    return translated_text
+    m2m100 = translation.ModelM2M100()
+    m2m100.tokenizer.src_lang = src_lang
+
+    translated_chunks = []
+    for chunk in chunks:
+        input_ids = m2m100.tokenizer(
+            chunk, return_tensors="pt").input_ids.to(m2m100.model.device)
+        outputs = m2m100.model.generate(
+            input_ids=input_ids,
+            forced_bos_token_id=m2m100.tokenizer.get_lang_id(tgt_lang))
+        translated_chunk = m2m100.tokenizer.batch_decode(
+            outputs, skip_special_tokens=True)[0]
+        translated_chunks.append(translated_chunk)
+    
+    return '\n'.join(translated_chunks)
 
 
 @spaces.GPU
-def translate_with_multilingual_model(
-        text: str,
+def translate_with_MADLAD(
+        chunks: List[str],
         tgt_lang: str,
-        sents_per_chunk: int=5,
         input_max_length: int=512,
-        output_max_length: int=512):
+        output_max_length: int=512) -> str:
+    """Translate with Google MADLAD model
     """
-    Translate the given text into English (default "easy" language)
-    """
-    chunks = build_text_chunks(text, None, sents_per_chunk)
-    translated_chunks = []
+    madlad = translation.ModelMADLAD()
 
     for chunk in chunks:
         input_text = f"<2{tgt_lang}> {text}"
-        logger.info(f" Translating: {input_text[:30]}")
-        input_ids = tokenizer_multilingual(
-            input_text, return_tensors="pt",
-            max_length=input_max_length,
-            truncation=True, padding="longest").input_ids.to(
-                model_multilingual.device)
-        outputs = model_multilingual.generate(
+        #logger.info(f" Translating: {input_text[:30]}")
+        input_ids = madlad.tokenizer(
+            input_text, return_tensors="pt", max_length=input_max_length,
+            truncation=True, padding="longest").input_ids.to(madlad.model.device)
+        outputs = madlad.model.generate(
             input_ids=input_ids, max_length=output_max_length)
-        translated_chunk = tokenizer_multilingual.decode(
+        translated_chunk = madlad.tokenizer.decode(
             outputs[0], skip_special_tokens=True)
         translated_chunks.append(translated_chunk)
     
@@ -195,37 +167,19 @@ def translate_text(
     src_lang = src_lang if (src_lang and src_lang != "auto") else detect_language(text)
     tgt_lang = 'en' # Default "easy" language
 
-    #
-    # Bilingual (Helsinki model)
-    #
-    translated_text_bilingual_model = translate_with_bilingual_model(
-        text, src_lang, tgt_lang, sents_per_chunk
-    )
-
-    #
-    # m2m100 model
-    #
-    translated_text_m2m100_model = translate_with_m2m100_model(
-        text, src_lang, tgt_lang, sents_per_chunk
-    )
+    chunks = build_text_chunks(text, src_lang, sents_per_chunk)
 
-    #
-    # Multilingual model (Google MADLAD)
-    #
-    
-    translated_text_multilingual_model = translate_with_multilingual_model(
-        text, tgt_lang, sents_per_chunk, input_max_length, output_max_length)
-    
-    #
-    # Google Translate
-    #
+    translated_text_Helsinki = translate_with_Helsinki(
+        chunks, src_lang, tgt_lang, input_max_length, output_max_length)
+    translated_text_m2m100 = translate_with_m2m100(chunks, src_lang, tgt_lang)
+    translated_text_MADLAD = translate_with_MADLAD(chunks, tgt_lang)
     translated_text_google_translate = GoogleTranslator(
         source='auto', target='en').translate(text=text)
 
     return (
-        translated_text_bilingual_model,
-
-        translated_text_multilingual_model,
+        translated_text_Helsinki,
+        translated_text_m2m100,
+        translated_text_MADLAD,
         translated_text_google_translate
     )
 
@@ -244,19 +198,19 @@ with gr.Blocks() as demo:
         label="Text to translate",
         render=False
     )
-    output_text_bilingual_model = gr.Textbox(
+    output_text_Helsinki = gr.Textbox(
         lines=6,
         label="Bilingual translation model (Helsinki NLP)",
         render=False
     )
-    output_text_m2m100_model = gr.Textbox(
+    output_text_m2m100 = gr.Textbox(
         lines=6,
         label="Facebook m2m100 translation model (**small**)",
         render=False
     )
-    output_text_multilingual_model = gr.Textbox(
+    output_text_MADLAD = gr.Textbox(
         lines=6,
-        label="Multilingual translation model (**small** Google MADLAD)",
+        label="Google MADLAD translation model (**small**)",
         render=False
     )
     output_text_google_translate = gr.Textbox(
@@ -272,7 +226,7 @@ with gr.Blocks() as demo:
         render=False
     )
     src_lang = gr.Radio(
-        choices=["auto", "ar", "en", "fa", "fr", "he", "ja", "zh"], value="auto",
+        choices=["auto", "ar", "en", "fa", "fr", "he", "zh"], value="auto",
         label="Source language",
         render=False
     )
@@ -282,22 +236,17 @@ with gr.Blocks() as demo:
         ["ریچارد مور، رئیس سازمان مخفی اطلاعاتی بریتانیا (ام‌آی‌۶) در دیدار ویلیام برنز، رئیس سازمان اطلاعات مرکزی آمریکا (سیا) گفت همچنان احتمال اقدام ایران علیه اسرائیل در واکنش به ترور اسماعیل هنیه، رهبر حماس وجود دارد. آقای برنز نیز در این دیدار فاش کرد که در سال اول جنگ اوکراین، «خطر واقعی» وجود داشت که روسیه به استفاده از «تسلیحات هسته‌ای تاکتیکی» متوسل شود. این دو مقام امنیتی هشدار دادند که «نظم جهانی» از زمان جنگ سرد تا کنون تا این حد «در معرض تهدید» نبوده است.", "fa"],
         ["Clément Delangue est, avec Julien Chaumond et Thomas Wolf, l’un des trois Français cofondateurs de Hugging Face, une start-up d’intelligence artificielle (IA) de premier plan. Valorisée à 4,2 milliards d’euros après avoir levé près de 450 millions d’euros depuis sa création en 2016, cette société de droit américain est connue comme la plate-forme de référence où développeurs et entreprises publient des outils et des modèles pour faire de l’IA en open source, c’est-à-dire accessible gratuitement et modifiable.", "fr"],
         ["يُعد تفشي مرض جدري القردة قضية صحية عالمية خطيرة، ومن المهم محاولة منع انتشاره للحفاظ على سلامة الناس وتجنب العدوى. د. صموئيل بولاند، مدير الحوادث الخاصة بمرض الجدري في المكتب الإقليمي لمنظمة الصحة العالمية في أفريقيا، يتحدث من كينشاسا في جمهورية الكونغو الديمقراطية، ولديه بعض النصائح البسيطة التي يمكن للناس اتباعها لتقليل خطر انتشار المرض.", "ar"],
-        ["【ワシントン＝冨山優介】米ボーイングの新型宇宙船「スターライナー」は７日午前０時（日本時間７日午後１時）過ぎ、米ニューメキシコ州のホワイトサンズ宇宙港に着地し、地球に帰還した。スターライナーは米宇宙飛行士２人を乗せて６月に打ち上げられ、国際宇宙ステーション（ＩＳＳ）に接続したが、機体のトラブルが解決できず、無人でＩＳＳから離脱した。", "ja"],
         ["張先生稱，奇瑞已經凖備在西班牙生產汽車，並決心採取「本地化」的方式進入歐洲市場。此外，他也否認該公司的出口受益於不公平補貼。奇瑞成立於1997年，是中國最大的汽車公司之一。它已經是中國最大的汽車出口商，並且制定了進一步擴張的野心勃勃的計劃。", "zh"],
         ["ברוכה הבאה, קיטי: בית הקפה החדש בלוס אנג'לס החתולה האהובה והחברים שלה מקבלים בית קפה משלהם בשדרות יוניברסל סיטי, שם תוכלו למצוא מגוון של פינוקים מתוקים – החל ממשקאות ועד עוגות", "he"],
     ]
-
-    outputs = gr.Row(
-        
-    )
     
     gr.Interface(
         fn=translate_text,
         inputs=[input_text, src_lang,],
         outputs=[
-            output_text_bilingual_model,
-            output_text_multilingual_model,
-            output_text_m2m100_model,
+            output_text_Helsinki,
+            output_text_m2m100,
+            output_text_MADLAD,
             output_text_google_translate,
         ],
         additional_inputs=[sentences_per_chunk,],
@@ -309,7 +258,7 @@ with gr.Blocks() as demo:
 
     with gr.Accordion("Documentation", open=False):
         gr.Markdown("""
-            - Models: serving bilingual models from Helsinki NLP and multilingual model from Google MADLAD.
+            - Models: serving Helsinki NLP, Facebook M2M100 and Google MADLAD models.
             - Basic: processing of long paragraph / document to be enhanced.
             - Most examples are copy/pasted from BBC news international web sites.
         """)

model_translation.py

@@ -10,7 +10,70 @@ Date: 2024-03-16
 
 import torch
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+from transformers import M2M100Tokenizer, M2M100ForConditionalGeneration
+
+
+class Singleton(type):
+    _instances = {}
+    def __call__(cls, *args, **kwargs):
+        if cls not in cls._instances:
+            cls._instances[cls] = super(Singleton, cls).__call__(*args, **kwargs)
+        return cls._instances[cls]
 
+class ModelM2M100(metaclass=Singleton):
+    """Loads an instance of the M2M100 model (418M).
+    """
+    def __init__(self):
+        self._model_name = "facebook/m2m100_418M"
+        self._tokenizer = M2M100Tokenizer.from_pretrained(self._model_name)
+        self._model = M2M100ForConditionalGeneration.from_pretrained(
+            self._model_name,
+            device_map="auto",
+            torch_dtype=torch.float16,
+            low_cpu_mem_usage=True
+        )
+    
+    @property
+    def model_name(self):
+        return self._model_name
+
+    @property
+    def tokenizer(self):
+        return self._tokenizer
+
+    @property
+    def model(self):
+        return self._model
+
+class ModelMADLAD(metaclass=Singleton):
+    """Loads an instance of the Google MADLAD model (3B).
+    """
+    def __init__(self, model_name):
+        self._model_name = "google/madlad400-3b-mt"
+        self._tokenizer = AutoTokenizer.from_pretrained(
+            self.model_name, use_fast=True
+        )
+        self._model = AutoModelForSeq2SeqLM.from_pretrained(
+            self._model_name,
+            device_map="auto",
+            torch_dtype=torch.float16,
+            low_cpu_mem_usage=True
+        )
+    
+    @property
+    def model_name(self):
+        return self._model_name
+    
+    @property
+    def tokenizer(self):
+        return self._tokenizer
+
+    @property
+    def model(self):
+        return self._model
+
+
+# Bi-lingual individual models
 src_langs = set(["ar", "en", "fa", "fr", "he", "ja", "zh"])
 model_names = {
     "ar": "Helsinki-NLP/opus-mt-ar-en",
@@ -18,7 +81,6 @@ model_names = {
     "fa": "Helsinki-NLP/opus-mt-tc-big-fa-itc",
     "fr": "Helsinki-NLP/opus-mt-fr-en",
     "he": "Helsinki-NLP/opus-mt-tc-big-he-en",
-    "ja": "Helsinki-NLP/opus-mt-jap-en",
     "zh": "Helsinki-NLP/opus-mt-zh-en",
 }
 
@@ -57,30 +119,3 @@ def get_tokenizer_model_for_src_lang(src_lang: str) -> (AutoTokenizer, AutoModel
 # - Let's set to some number of words somewhat lower than that threshold
 # - e.g. 200 words
 max_words_per_chunk = 200
-
-#
-# Multilingual language pairs
-#
-from transformers import M2M100Tokenizer, M2M100ForConditionalGeneration
-
-model_name_m2m100 = "facebook/m2m100_418M"
-tokenizer_m2m100 = M2M100Tokenizer.from_pretrained(model_name_m2m100)
-model_m2m100 = M2M100ForConditionalGeneration.from_pretrained(
-    model_name_m2m100,
-    device_map="auto",
-    torch_dtype=torch.float16,
-    low_cpu_mem_usage=True
-)
-
-#
-# Multilingual translation model
-#
-model_MADLAD_name = "google/madlad400-3b-mt"
-#model_MADLAD_name = "google/madlad400-7b-mt-bt"
-tokenizer_multilingual = AutoTokenizer.from_pretrained(model_MADLAD_name, use_fast=True)
-model_multilingual = AutoModelForSeq2SeqLM.from_pretrained(
-    model_MADLAD_name,
-    device_map="auto",
-    torch_dtype=torch.float16,
-    low_cpu_mem_usage=True
-)