app.py

import os
import requests
import gradio as gr

from utils import get_token

url_caption = os.environ["CAPTION_NODE"]
url_vqa = os.environ["VQA_NODE"]


def image_caption(file_path):
    token = get_token()

    files = {"file": open(file_path, "rb")}
    headers = {"X-Auth-Token": token}
    resp = requests.post(url_caption,
                         files=files,
                         headers=headers,
                         verify=False)
    resp = resp.json()
    desc = resp["inference_result"]["instances"]["image"][0]
    return desc


def vqa(file_path, question):
    token = get_token()

    files = {"file": open(file_path, "rb")}
    question = {"question": question}
    headers = {"X-Auth-Token": token}
    resp = requests.post(url_vqa,
                         files=files,
                         data=question,
                         headers=headers,
                         verify=False)
    resp = resp.json()
    ans = resp["inference_result"]["instances"]
    return ans


def read_content(file_path):
    with open(file_path, 'r', encoding='utf-8') as f:
        content = f.read()
    return content


examples_caption = [
    os.path.join(os.path.dirname(__file__), "examples/caption/00.jpg"),
    os.path.join(os.path.dirname(__file__), "examples/caption/01.jpg"),
    os.path.join(os.path.dirname(__file__), "examples/caption/02.jpg"),
    os.path.join(os.path.dirname(__file__), "examples/caption/03.jpg"),
    os.path.join(os.path.dirname(__file__), "examples/caption/04.jpg"),
    os.path.join(os.path.dirname(__file__), "examples/caption/05.jpg")
]
examples_vqa = [
    os.path.join(os.path.dirname(__file__), "examples/vqa/00.jpg"),
    os.path.join(os.path.dirname(__file__), "examples/vqa/01.jpg"),
    os.path.join(os.path.dirname(__file__), "examples/vqa/02.jpg"),
    os.path.join(os.path.dirname(__file__), "examples/vqa/03.jpg"),
    os.path.join(os.path.dirname(__file__), "examples/vqa/04.jpg"),
    os.path.join(os.path.dirname(__file__), "examples/vqa/05.jpg")
]

css = """
.gradio-container {background-image: url('file=./background.jpg'); background-size:cover; background-repeat: no-repeat;}

#infer {
    background: linear-gradient(to bottom right, #FFD8B4, #FFB066);
    border: 1px solid #ffd8b4;
    border-radius: 8px;
    color: #ee7400
}
"""

with gr.Blocks(css=css) as demo:
    gr.HTML(read_content("./header.html"))
    gr.Markdown("# MindSpore Zidongtaichu ")
    gr.Markdown(
        "\nOPT (Omni-Perception Pre-Trainer) is the abbreviation of the full-scene perception pre-training model.  "
        " It is an important achievement of the Chinese Academy of Sciences Automation and Huawei on the road to exploring general artificial intelligence."
        " The modal 100 billion large model, the Chinese name is Zidong.Taichu."
        " supports efficient collaboration among different modalities of text, vision, and voice,"
        " and can support industrial applications such as film and television creation, industrial quality inspection, and intelligent driving."
    )

    with gr.Tab("以图生文 (Image Caption)"):
        with gr.Row():
            caption_input = gr.Image(
                type="filepath",
                value=examples_caption[0],
            )
            caption_output = gr.TextArea(label="description",
                                         interactive=False)
        caption_button = gr.Button("Submit", elem_id="infer")
        gr.Examples(
            examples=examples_caption,
            inputs=caption_input,
        )

    caption_button.click(image_caption,
                         inputs=[caption_input],
                         outputs=[caption_output])

    with gr.Tab("视觉问答 (VQA)"):
        with gr.Row():
            with gr.Column():
                q_pic_input = gr.Image(type="filepath",
                                       label="step1: select a picture")
                gr.Examples(
                    examples=examples_vqa,
                    inputs=q_pic_input,
                )
            with gr.Column():
                vqa_question = gr.TextArea(
                    label="step2: question",
                    lines=5,
                    placeholder="please enter a question related to the picture"
                )
                vqa_answer = gr.TextArea(label="answer",
                                         lines=5,
                                         interactive=False)
                vqa_button = gr.Button("Submit", elem_id="infer")

    vqa_button.click(vqa,
                     inputs=[q_pic_input, vqa_question],
                     outputs=[vqa_answer])

    with gr.Accordion("Open for More!"):
        gr.Markdown(
            "- If you want to know more about the foundation models of MindSpore, please visit "
            "[The Foundation Models Platform for Mindspore](https://xihe.mindspore.cn/)"
        )
        gr.Markdown(
            "- If you want to know more about OPT models, please visit "
            "[OPT](https://gitee.com/mindspore/zidongtaichu)")
        gr.Markdown(
            "- Try [zidongtaichu model on the Foundation Models Platform for Mindspore]"
            "(https://xihe.mindspore.cn/modelzoo/taichug)")

demo.queue(concurrency_count=5)
demo.launch(enable_queue=True)