main.py

from ultralytics import YOLO
from base64 import b64encode
from speech_recognition import AudioFile, Recognizer
import numpy as np
from scipy.spatial import distance as dist

from sahi.utils.cv import read_image_as_pil
from fastapi import FastAPI, File, UploadFile, Form
from utils import tts, read_image_file, pil_to_base64, base64_to_pil, get_hist
from typing import Optional
from huggingface_hub import hf_hub_download

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

app = FastAPI()

@app.get("/")
def read_root():
    return {"Message": "Application startup complete"}


@app.post("/aisatsu_api/")
async def predict_api(
        file: UploadFile = File(...),
        last_seen: Optional[str] = Form(None)
):
    image = read_image_file(await file.read())
    results = model.predict(image, show=False)[0]
    image = read_image_as_pil(image)
    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((64, 64))
                most_close = area_rate
    if last_seen is not None:
        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 {
        "voice": voice_bot,
        "image": pil_to_base64(out_img) if out_img is not None else None
    }