app.py

from gtts import gTTS
from io import BytesIO
import base64
from PIL import Image
import cv2
import numpy as np
import gradio as gr
from ultralyticsplus import YOLO
from base64 import b64encode
from speech_recognition import AudioFile, Recognizer
import numpy as np
from utils import tts, read_image_file, pil_to_base64, base64_to_pil, get_hist
from scipy.spatial import distance as dist

model = YOLO('ultralyticsplus/yolov8s')
CLASS = model.model.names
defaul_bot_voice = "おはいようございます"
area_thres = 0.3

def infer(image, last_seen):
    results = model.predict(image, show=False)[0]
    masks, boxes = results.masks, results.boxes
    area_image = image.width * image.height
    voice_bot = None
    most_close = 0
    out_img = None
    diff_value = 0.5
    if boxes is not None:
        for xyxy, conf, cls in zip(boxes.xyxy, boxes.conf, boxes.cls):
            if int(cls) != 0:
                continue
            box = xyxy.tolist()
            area_rate = (box[2] - box[0]) * (box[3] - box[1]) / area_image
            if area_rate >= most_close:
                out_img = image.crop(tuple(box)).resize((128, 128))
                most_close = area_rate
    if last_seen != "":
        last_seen = base64_to_pil(last_seen)
        if out_img is not None:
            diff_value = dist.euclidean(get_hist(out_img), get_hist(last_seen))
    print(most_close, diff_value)
    if most_close >= area_thres and diff_value >= 0.5:
        voice_bot = tts(defaul_bot_voice, language="ja")
    return out_img, voice_bot
    
iface = gr.Interface(
    fn=infer,
    title="aisatsu api",
    inputs=[gr.Image(label="image", type="pil", shape=(960, 640)), gr.Textbox(label="last seen", value="")],
    outputs=[gr.Image(label="output image"), gr.Textbox(label="output voice")],
    article = "Author: <a href=\"https://huggingface.co/vumichien\">Vu Minh Chien</a>.",
).launch(enable_queue=True, debug=True)