app.py

from fastai.vision.all import *
from io import BytesIO
import requests
import streamlit as st

import numpy as np
import torch
import time
import cv2
from numpy import random

import sys
sys.path.append('./yolov7')  # 確保 yolov7 目錄在 Python 路徑中

from models.experimental import attempt_load
from utils.general import check_img_size, non_max_suppression, scale_coords
from utils.plots import plot_one_box

def letterbox(img, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32):
    shape = img.shape[:2]  # 當前形狀 [高度, 寬度]
    if isinstance(new_shape, int):
        new_shape = (new_shape, new_shape)

    # 縮放比例（新/舊）
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    if not scaleup:  # 僅縮小，不放大（以獲得更好的測試 mAP）
        r = min(r, 1.0)

    # 計算填充
    ratio = r, r  # 寬度、高度比例
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # 寬高填充
    if auto:  # 最小矩形
        dw, dh = np.mod(dw, stride), np.mod(dh, stride)  # 寬高填充
    elif scaleFill:  # 拉伸
        dw, dh = 0.0, 0.0
        new_unpad = (new_shape[1], new_shape[0])
        ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # 寬度、高度比例

    dw /= 2  # 將填充分成兩邊
    dh /= 2

    if shape[::-1] != new_unpad:  # 調整大小
        img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # 添加邊框
    return img, ratio, (dw, dh)

def detect_modify(img0, model, conf, imgsz=640, conf_thres=0.25, iou_thres=0.45):
    st.image(img0, caption="您的圖片", use_column_width=True)

    stride = int(model.stride.max())  # 模型步幅
    imgsz = check_img_size(imgsz, s=stride)  # 檢查圖片大小

    # 填充調整大小
    img0 = cv2.cvtColor(np.asarray(img0), cv2.COLOR_RGB2BGR)
    img = letterbox(img0, imgsz, stride=stride)[0]
    # 轉換
    img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR轉RGB, 到3x416x416
    img = np.ascontiguousarray(img)

    # 獲取名稱和顏色
    names = model.module.names if hasattr(model, 'module') else model.names
    colors = [[random.randint(0, 255) for _ in range(3)] for _ in names]

    # 進行推理
    old_img_w = old_img_h = imgsz
    old_img_b = 1

    t0 = time.time()
    img = torch.from_numpy(img).to(device)
    img = img / 255.0
    if img.ndimension() == 3:
        img = img.unsqueeze(0)

    # 推理
    with torch.no_grad():  # 計算梯度會導致GPU內存洩漏
        pred = model(img)[0]

    # 應用NMS（非極大值抑制）
    pred = non_max_suppression(pred, conf, iou_thres)

    # 處理檢測
    gn = torch.tensor(img0.shape)[[1, 0, 1, 0]]  # 正規化增益 whwh

    det = pred[0]
    if len(det):
        # 將框從img_size縮放到im0大小
        det[:, :4] = scale_coords(img.shape[2:], det[:, :4], img0.shape).round()

        # 打印結果
        s = ''
        for c in det[:, -1].unique():
            n = (det[:, -1] == c).sum()  # 每類檢測數量
            s += f"{n} {names[int(c)]}{'s' * (n > 1)}, "  # 添加到字符串

        # 寫入結果
        for *xyxy, conf, cls in reversed(det):
            label = f'{names[int(cls)]} {conf:.2f}'
            plot_one_box(xyxy, img0, label=label, color=colors[int(cls)], line_thickness=3)

    f"""
    ### 預測結果:
    """
    img0 = cv2.cvtColor(np.asarray(img0), cv2.COLOR_BGR2RGB)
    st.image(img0, caption="預測結果", use_column_width=True)

# 設置參數
weight_path = './best.pt'
imgsz = 640
conf = 0.6
conf_thres = 0.25
iou_thres = 0.45
device = torch.device("cpu")
path = "./"

# 加載模型
model = attempt_load(weight_path, map_location=torch.device('cpu'))  # 加載FP32模型

def process_video(video_path, model, conf, imgsz=640, conf_thres=0.25, iou_thres=0.45):
    cap = cv2.VideoCapture(video_path)
    stframe = st.empty()

    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break

        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        detect_modify(frame, model, conf=conf, imgsz=imgsz, conf_thres=conf_thres, iou_thres=iou_thres)
        stframe.image(frame, channels="RGB")

    cap.release()

"""
# 籃球畫面物件捕捉
使用者可以將您想測試的圖片或影片上傳，模型將會把球員、籃球與籃框這三個物件捕捉出來，並將其附上對應的框與信任指數，
而使用者也可以藉由拖曳畫面上的指標，來選擇自己想要探測的信任指數。
"""
option = st.radio("", ["上傳圖片", "圖片 URL", "上傳影片"])
conf = st.slider("選擇置信度閾值:", min_value=0.0, max_value=1.0, value=0.6)

if option == "上傳圖片":
    uploaded_file = st.file_uploader("請上傳圖片。")

    if uploaded_file is not None:
        img = PILImage.create(uploaded_file)
        detect_modify(img, model, conf=conf, imgsz=imgsz, conf_thres=conf_thres, iou_thres=iou_thres)
elif option == "圖片 URL":
    url = st.text_input("請輸入網址。")
    if url != "":
        try:
            response = requests.get(url)
            pil_img = PILImage.create(BytesIO(response.content))
            detect_modify(pil_img, model, conf=conf, imgsz=imgsz, conf_thres=conf_thres, iou_thres=iou_thres)
        except:
            st.text("讀取圖像有問題", url)
elif option == "上傳影片":
    uploaded_video = st.file_uploader("請上傳影片。", type=["mp4", "avi", "mov", "mkv"])

    if uploaded_video is not None:
        video_path = uploaded_video.name
        with open(video_path, mode='wb') as f:
            f.write(uploaded_video.read())
        process_video(video_path, model, conf=conf, imgsz=imgsz, conf_thres=conf_thres, iou_thres=iou_thres)