back to t5

app.py

@@ -1,169 +1,99 @@
 import gradio as gr
-from transformers import pipeline
+from transformers import T5ForConditionalGeneration, T5Tokenizer
 
-# Initialiser les pipelines de traduction pour chaque paire de langues
-# Utiliser des modèles Helsinki-NLP qui sont spécifiquement entraînés pour la traduction
-translation_models = {
-    "English-French": pipeline("translation", model="Helsinki-NLP/opus-mt-en-fr"),
-    "French-English": pipeline("translation", model="Helsinki-NLP/opus-mt-fr-en"),
-    "English-Spanish": pipeline("translation", model="Helsinki-NLP/opus-mt-en-es"),
-    "Spanish-English": pipeline("translation", model="Helsinki-NLP/opus-mt-es-en"),
-    "English-Japanese": pipeline("translation", model="Helsinki-NLP/opus-mt-en-jap"),
-    "Japanese-English": pipeline("translation", model="Helsinki-NLP/opus-mt-jap-en"),
-    "English-Chinese": pipeline("translation", model="Helsinki-NLP/opus-mt-en-zh"),
-    "Chinese-English": pipeline("translation", model="Helsinki-NLP/opus-mt-zh-en"),
-    "French-Spanish": pipeline("translation", model="Helsinki-NLP/opus-mt-fr-es"),
-    "Spanish-French": pipeline("translation", model="Helsinki-NLP/opus-mt-es-fr"),
-    "French-Japanese": pipeline("translation", model="Helsinki-NLP/opus-mt-fr-jap"),
-    "Japanese-French": pipeline("translation", model="Helsinki-NLP/opus-mt-jap-fr"),
-    "French-Chinese": pipeline("translation", model="Helsinki-NLP/opus-mt-fr-zh"),
-    "Chinese-French": pipeline("translation", model="Helsinki-NLP/opus-mt-zh-fr"),
-    "Spanish-Japanese": pipeline("translation", model="Helsinki-NLP/opus-mt-es-jap"),
-    "Japanese-Spanish": pipeline("translation", model="Helsinki-NLP/opus-mt-jap-es"),
-    "Spanish-Chinese": pipeline("translation", model="Helsinki-NLP/opus-mt-es-zh"),
-    "Chinese-Spanish": pipeline("translation", model="Helsinki-NLP/opus-mt-zh-es"),
-    "Japanese-Chinese": pipeline("translation", model="Helsinki-NLP/opus-mt-jap-zh"),
-    "Chinese-Japanese": pipeline("translation", model="Helsinki-NLP/opus-mt-zh-jap")
-}
+# Initialize T5 model and tokenizer
+tokenizer = T5Tokenizer.from_pretrained("t5-base")
+model = T5ForConditionalGeneration.from_pretrained("t5-base")
 
-# Mappages des noms des langues
-language_names = {
-    "English": "English",
-    "French": "French",
-    "Japanese": "Japanese",
-    "Spanish": "Spanish",
-    "Chinese": "Chinese"  # Renommé de "Mandarin" à "Chinese" pour correspondre aux noms des modèles
+# Supported language pairs for T5-base
+languages = ["English", "French", "Japanese", "Spanish", "Chinese"]
+language_codes = {
+    "English": "en",
+    "French": "fr",
+    "Japanese": "ja",
+    "Spanish": "es",
+    "Chinese": "zh"  # Using Chinese instead of Mandarin for T5 compatibility
 }
 
+# Translation function using T5
 def translate(text, source_lang, target_lang):
-    """Traduire le texte en utilisant le modèle approprié"""
-    if not text:
-        return ""
-    
     if source_lang == target_lang:
         return text
     
-    # Construire la clé du modèle
-    model_key = f"{source_lang}-{target_lang}"
+    source_code = language_codes[source_lang]
+    target_code = language_codes[target_lang]
+    
+    # Format the input as expected by T5
+    task_prefix = f"translate {source_code} to {target_code}: "
+    input_text = task_prefix + text
+    
+    # Tokenize and generate translation
+    inputs = tokenizer(input_text, return_tensors="pt", padding=True)
+    
+    output_sequences = model.generate(
+        input_ids=inputs["input_ids"],
+        attention_mask=inputs["attention_mask"],
+        max_length=512,
+        do_sample=False
+    )
     
-    # Vérifier si ce modèle existe
-    if model_key in translation_models:
-        try:
-            result = translation_models[model_key](text)
-            return result[0]["translation_text"]
-        except Exception as e:
-            return f"Erreur de traduction: {str(e)}"
-    else:
-        return f"La traduction de {source_lang} vers {target_lang} n'est pas prise en charge."
+    # Decode the output
+    translation = tokenizer.decode(output_sequences[0], skip_special_tokens=True)
+    return translation
 
-# Exemples de textes dans différentes langues avec des étiquettes descriptives
+# Example texts for each language
 examples = {
-    "English": {
-        "Basic greeting": "Hello, how are you today?",
-        "Travel plans": "I would like to visit Japan someday.",
-        "Food preference": "I like potatoes and vegetables.",
-        "Weather": "It's a beautiful sunny day today.",
-        "Question": "Where is the nearest train station?"
-    },
-    "French": {
-        "Salutation": "Bonjour, comment allez-vous aujourd'hui?",
-        "Voyage": "J'aimerais visiter le Japon un jour.",
-        "Nourriture": "J'aime les pommes de terre et les légumes.",
-        "Météo": "Il fait beau et ensoleillé aujourd'hui.",
-        "Question": "Où est la gare la plus proche?"
-    },
-    "Japanese": {
-        "挨拶": "こんにちは、今日はお元気ですか？",
-        "旅行": "いつか日本を訪れたいです。",
-        "食べ物": "ジャガイモと野菜が好きです。",
-        "天気": "今日は晴れていて美しい一日です。",
-        "質問": "最寄りの駅はどこですか？"
-    },
-    "Spanish": {
-        "Saludo": "Hola, ¿cómo estás hoy?",
-        "Viaje": "Me gustaría visitar Japón algún día.",
-        "Comida": "Me gustan las patatas y las verduras.",
-        "Clima": "Hoy es un hermoso día soleado.",
-        "Pregunta": "¿Dónde está la estación de tren más cercana?"
-    },
-    "Chinese": {
-        "问候": "你好，今天好吗？",
-        "旅行": "我希望有一天能去日本。",
-        "食物": "我喜欢土豆和蔬菜。",
-        "天气": "今天是个阳光明媚的美好日子。",
-        "问题": "最近的火车站在哪里？"
-    }
+    "English": ["I went to the supermarket yesterday.", "The weather is beautiful today."],
+    "French": ["Je suis allé au supermarché hier.", "Le temps est magnifique aujourd'hui."],
+    "Japanese": ["昨日スーパーマーケットに行きました。", "今日の天気は素晴らしいです。"],
+    "Spanish": ["Fui al supermercado ayer.", "El clima está hermoso hoy."],
+    "Chinese": ["我昨天去了超市。", "今天天气很好。"]
 }
 
-# Créer l'interface Gradio
+# Create Gradio interface
 with gr.Blocks() as demo:
-    gr.Markdown("# Traduction Multilingue")
+    gr.Markdown("# Multilingual Translation App (T5-base)")
     
     with gr.Row():
         source_lang = gr.Dropdown(
-            choices=list(language_names.keys()),
-            label="Langue source",
+            languages, 
+            label="Source Language", 
             value="English"
         )
         target_lang = gr.Dropdown(
-            choices=list(language_names.keys()),
-            label="Langue cible",
+            languages, 
+            label="Target Language", 
             value="French"
         )
     
     with gr.Row():
         with gr.Column():
-            source_text = gr.Textbox(label="Texte source", lines=5)
-            translate_btn = gr.Button(value="Traduire")
-        
+            input_text = gr.Textbox(label="Source Text", placeholder="Enter text to translate...")
+            translate_btn = gr.Button(value="Translate")
         with gr.Column():
-            target_text = gr.Textbox(label="Résultat de la traduction", lines=5)
+            output_text = gr.Textbox(label="Translated Text")
     
-    # Ajouter une section d'exemples
-    with gr.Accordion("Exemples", open=True):
-        example_dropdown = gr.Dropdown(
-            choices=list(examples["English"].keys()),
-            label="Sélectionner un exemple",
-            interactive=True,
-            value=list(examples["English"].keys())[0] if examples["English"] else None
-        )
-        load_example_btn = gr.Button(value="Charger l'exemple")
-    
-    # Mettre à jour le menu déroulant d'exemples lorsque la langue source change
-    def update_examples_dropdown(lang):
-        return gr.Dropdown.update(
-            choices=list(examples.get(lang, {}).keys()),
-            value=list(examples.get(lang, {}).keys())[0] if examples.get(lang, {}) else None
-        )
+    # Dynamic examples based on source language
+    example_component = gr.Examples(examples=examples["English"], inputs=input_text)
     
-    source_lang.change(
-        update_examples_dropdown,
-        inputs=[source_lang],
-        outputs=[example_dropdown]
-    )
-    
-    # Charger l'exemple sélectionné dans le texte source
-    def load_selected_example(example_label, lang):
-        return examples.get(lang, {}).get(example_label, "")
+    # Update examples when source language changes
+    def update_examples(lang):
+        return gr.Examples(examples=examples.get(lang, []), inputs=input_text)
     
-    load_example_btn.click(
-        load_selected_example,
-        inputs=[example_dropdown, source_lang],
-        outputs=[source_text]
-    )
+    source_lang.change(update_examples, inputs=source_lang, outputs=example_component)
     
-    # Configurer la fonction de traduction
+    # Translation function connections
     translate_btn.click(
-        translate,
-        inputs=[source_text, source_lang, target_lang],
-        outputs=target_text
+        translate, 
+        inputs=[input_text, source_lang, target_lang], 
+        outputs=output_text
     )
     
-    # Traduire également lorsque la touche Entrée est appuyée dans la zone de texte source
-    source_text.submit(
-        translate,
-        inputs=[source_text, source_lang, target_lang],
-        outputs=target_text
+    # Also translate when Enter key is pressed
+    input_text.submit(
+        translate, 
+        inputs=[input_text, source_lang, target_lang], 
+        outputs=output_text
     )
 
 if __name__ == "__main__":

requirements.txt

@@ -1,5 +1,4 @@
 gradio>=3.50.2
 transformers>=4.35.0
 torch>=2.0.0
-sentencepiece>=0.1.99
-sacremoses>=0.0.53
+sentencepiece>=0.1.99