app.py

import gradio as gr
from transformers import MT5Tokenizer, MT5ForConditionalGeneration
import torch

css_code = f"""
.container {{ 
    max-width: 1500px; 
    margin: auto; 
    padding: 20px; 
    background-color: rgba(255, 255, 255, 0.8); /* 半透明背景使内容更清晰 */
    border-radius: 8px;
}}
.title {{ text-align: center; margin-bottom: 30px; }}
.title h1 {{ color: #2c3e50; }}
.title h2 {{ color: #34495e; }}
.gr-button {{ border-radius: 8px; }}
.footer {{ max-width: 1500px; text-align: center; margin-top: 30px; color: #666; }}
"""

loaded_models = {}

def get_model_and_tokenizer(model_choice, device):
    model_repo = "MHBS-IHB/fish-mt5"
    if model_choice in loaded_models:
        return loaded_models[model_choice]
    
    tokenizer = MT5Tokenizer.from_pretrained(model_repo, subfolder=model_choice)
    model = MT5ForConditionalGeneration.from_pretrained(model_repo, subfolder=model_choice)
    model.to(device)
    loaded_models[model_choice] = (model, tokenizer)
    return model, tokenizer

def predict(model, tokenizer, input_text, device, translation_type):
    model.eval()
    if translation_type == "Chinese to Latin":
        input_text = f"translate Chinese to Latin: {input_text}"
    elif translation_type == "Latin to Chinese":
        input_text = f"translate Latin to Chinese: {input_text}"
    
    inputs = tokenizer(
        input_text,
        return_tensors='pt',
        max_length=100,
        truncation=True
    ).to(device)
    
    with torch.no_grad():
        outputs = model.generate(
            input_ids=inputs['input_ids'],
            attention_mask=inputs['attention_mask'],
            max_length=50,
            num_beams=5,
            early_stopping=True
        )
    
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

def gradio_predict(input_text, translation_type, model_choice):
    model, tokenizer = get_model_and_tokenizer(model_choice, DEVICE)
    return predict(model=model, tokenizer=tokenizer, 
                   input_text=input_text, device=DEVICE, 
                   translation_type=translation_type)

def clear_text():
    return "", ""

with gr.Blocks(
    theme=gr.themes.Soft(),
    css=css_code

) as iface:
    # 顶部标题
    gr.Markdown(
        """
        <div class="title">
            <h1>🐟 世界鱼类拉汉互译 🐠</h1>
            <h2>Dual Latin-Chinese Translation of Global Fish Species</h2>
        </div>
        """,
        elem_classes="container"
    )

    with gr.Row(elem_classes="container"):
        with gr.Column(scale=1):
            input_text = gr.Textbox(
                label="Enter Fish Species Names (Chinese or Latin)", 
                placeholder="例如：中华鲟 / Acipenser sinensis"
            )
            # 翻译方向改为“Bidirectional Translation”并提供其他备选项
            translation_type = gr.Radio(
                choices=["Chinese to Latin", "Latin to Chinese"],
                label="Bidirectional translation",
                value="Chinese to Latin"
            )
        with gr.Column(scale=1):
            output_text = gr.Textbox(
                label="Translation result"
            )
            model_choice = gr.Dropdown(
            choices=["fish_mt5_small", "fish_mt5_base", "fish_mt5_large"],
            label="Select Model",
            value="fish_mt5_large"
            )  
    
    with gr.Row(elem_classes="container"):
        translate_btn = gr.Button("Translate 🔄", variant="primary")
        clear_btn = gr.Button("Clear 🗑️", variant="secondary")
    
    
    gr.Markdown(
        """
        <div class="footer">
            <p>🌊 Powered by fine-tuned MT5 Model | The model might take a while to load for the first time, so please be patient—but once it's loaded, translations are lightning fast! 🚀</p>
        </div>
        """,
        elem_classes="container"
    )
    
    
    translate_btn.click(
        fn=gradio_predict,
        inputs=[input_text, translation_type, model_choice],
        outputs=output_text
    )
    clear_btn.click(
        fn=clear_text,
        inputs=[],
        outputs=[input_text, output_text]
    )
    
    gr.Examples(
        examples=[
            ["中华鲟", "Chinese to Latin", "fish_mt5_small"],
            ["Acipenser sinensis", "Latin to Chinese", "fish_mt5_small"]
        ],
        inputs=[input_text, translation_type, model_choice],
        outputs=output_text,
        label="📋 Translation Examples"
    )

if __name__ == "__main__":
    iface.launch()