app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from struct_caption import StructCaptioner
from fusion_caption import FusionCaptioner

# Khởi tạo mô hình
struct_captioner = StructCaptioner("Skywork/SkyCaptioner-V1")
fusion_captioner = FusionCaptioner("Qwen/Qwen3-8B")

# Tải mô hình dịch tiếng Việt
translation_model = AutoModelForSeq2SeqLM.from_pretrained("VietAI/envit5-translation")
translation_tokenizer = AutoTokenizer.from_pretrained("VietAI/envit5-translation")

def translate_to_vietnamese(text):
    inputs = translation_tokenizer(f"en: {text}", return_tensors="pt", padding=True)
    outputs = translation_model.generate(**inputs, max_length=512)
    return translation_tokenizer.decode(outputs[0], skip_special_tokens=True)

# Giao diện người dùng Gradio
with gr.Blocks() as demo:
    gr.Markdown("<h1 style='text-align: center;'>SkyCaptioner-V1</h1>")
    
    with gr.Row():
        with gr.Column(scale=0.5):
            video_input = gr.Video(label="Upload Video", interactive=True, format="mp4")
            btn_struct = gr.Button("Generate Struct Caption")
        with gr.Column():
            struct_caption_output = gr.Code(label="Struct Caption", language="json", lines=25, interactive=False)

    with gr.Row():
        with gr.Column(scale=0.5):
            task_input = gr.Radio(label="Task Type", choices=["t2v", "i2v"], value="t2v", interactive=True)
            btn_fusion = gr.Button("Generate Fusion Caption")
        with gr.Column():
            fusion_caption_output = gr.Textbox(label="Fusion Caption", value="", interactive=False)

    @gr.Interface(fn=generate_struct_caption, inputs=video_input, outputs=struct_caption_output)
    def generate_struct_caption(video):
        struct_caption = struct_captioner(video)
        return struct_caption

    @gr.Interface(fn=generate_fusion_caption, inputs=[struct_caption_output, task_input], outputs=fusion_caption_output)
    def generate_fusion_caption(struct_caption_str, task):
        fusion_caption = fusion_captioner(struct_caption_str, task)
        if task == "t2v":
            return fusion_caption
        else:
            return translate_to_vietnamese(fusion_caption)

    gr.Examples(
        examples=[["./examples/1.mp4"], ["./examples/2.mp4"], ["./examples/3.mp4"], ["./examples/4.mp4"]],
        inputs=video_input,
        label="Example Videos"
    )

    demo.launch()