YOLOv5代码详解（test.py部分）

目录

 

2. test.py

2.1 设置超参数

2.2 设置任务（验证，测试，学习）

2.3 测试函数

2.3.1 初始化模型

2.3.2 判断设备类型并仅使用一张GPU进行测试

2.3.3 获取配置文件路径和文件参数

2.3.4 数据获取

2.3.5 计算map数据

2.3.6 打印结果（图片，速度），保存结果至json，并返回结果

---

2. test.py

该部分主要用于运行train.py时，计算每个epoch的mAP。
PS，与train.py相似的部分就不再阐述。

2.1 设置超参数

权重，数据，batch size，图像尺寸，使用哪张显卡，数据增强，计算mAP。if __name__ == '__main__':

    parser = argparse.ArgumentParser(prog='test.py')
    parser.add_argument('--weights', type=str, default='weights/best.pt', help='model.pt path')
    parser.add_argument('--data', type=str, default='data/coco.yaml', help='*.data path')
    parser.add_argument('--batch-size', type=int, default=16, help='size of each image batch')
    parser.add_argument('--img-size', type=int, default=608, help='inference size (pixels)')
    parser.add_argument('--conf-thres', type=float, default=0.001, help='object confidence threshold')
    parser.add_argument('--iou-thres', type=float, default=0.65, help='IOU threshold for NMS')
    parser.add_argument('--save-json', action='store_true', help='save a cocoapi-compatible JSON results file')
    parser.add_argument('--task', default='', help="'val', 'test', 'study'")
    parser.add_argument('--device', default='1', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset')
    parser.add_argument('--augment', action='store_true', help='augmented inference')
    parser.add_argument('--verbose', action='store_true', help='report mAP by class')
    opt = parser.parse_args()
    opt.img_size = check_img_size(opt.img_size)
    opt.save_json = opt.save_json or opt.data.endswith('coco.yaml')
    opt.data = check_file(opt.data)  # check file
    print(opt)

2.2 设置任务（验证，测试，学习）

    # task = 'val', 'test', 'study'
    if opt.task in ['val','test']:  # (default) run normally
        test(opt.data,
             opt.weights,
             opt.batch_size,
             opt.img_size,
             opt.conf_thres,
             opt.iou_thres,
             opt.save_json,
             opt.single_cls,
             opt.augment,
             opt.verbose)

    elif opt.task == 'study':  # run over a range of settings and save/plot
        for weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']:
            f = 'study_%s_%s.txt' % (Path(opt.data).stem, Path(weights).stem)  # filename to save to
            x = list(range(288, 896, 64))  # x axis
            y = []  # y axis
            for i in x:  # img-size
                print('\nRunning %s point %s...' % (f, i))
                r, _, t = test(opt.data, weights, opt.batch_size, i, opt.conf_thres, opt.iou_thres, opt.save_json)
                y.append(r + t)  # results and times
            np.savetxt(f, y, fmt='%10.4g')  # save
        os.system('zip -r study.zip study_*.txt')
        # plot_study_txt(f, x)  # plot

“study”应该是可以修改预训练权重的操作。

2.3 测试函数

2.3.1 初始化模型

判断模型是否存在，若不存在则训练为假，移除之前的测试结果，下载模型。

    # Initialize/load model and set device
    if model is None:
        training = False
        device = torch_utils.select_device(opt.device, batch_size=batch_size)
        half = device.type != 'cpu'  # half precision only supported on CUDA

        # Remove previous
        for f in glob.glob('test_batch*.jpg'):
            os.remove(f)

        # Load model
        google_utils.attempt_download(weights)
        model = torch.load(weights, map_location=device)['model'].float()  # load to FP32
        torch_utils.model_info(model)
        model.fuse()
        model.to(device)
        if half:
            model.half()  # to FP16

        # Multi-GPU disabled, incompatible with .half()
        # if device.type != 'cpu' and torch.cuda.device_count() > 1:
        #     model = nn.DataParallel(model)

    else:  # called by train.py
        training = True
        device = next(model.parameters()).device  # get model device
        # half disabled https://github.com/ultralytics/yolov5/issues/99
        half = False  # device.type != 'cpu' and torch.cuda.device_count() == 1
        if half:
            model.half()  # to FP16

2.3.2 判断设备类型并仅使用一张GPU进行测试

判断设备类型并仅仅单GPU支持一半的精度。

    # Half
    half = device.type != 'cpu' and torch.cuda.device_count() == 1  # half precision only supported on single-GPU
    if half:
        model.half()  # to FP16

2.3.3 获取配置文件路径和文件参数

获取配置文件yaml中的nc参数。    # Configure
    model.eval()
    with open(data) as f:
        data = yaml.load(f, Loader=yaml.FullLoader)  # model dict
    nc = 1 if single_cls else int(data['nc'])  # number of classes
    iouv = torch.linspace(0.5, 0.95, 10).to(device)  # iou vector for [email protected]:0.95
    niou = iouv.numel()

2.3.4 数据获取

测试数据的获取
使用NMS
初始化图片大小
获取测试或验证图片路径
用梯度下降运行模型、计算损失和运行NMS
计算每张测试图片的数据并保存到json文件中
画出某几张batch图片（包括GT和预测）

    # Dataloader
    if dataloader is None:  # not training
        merge = opt.merge  # use Merge NMS
        img = torch.zeros((1, 3, imgsz, imgsz), device=device)  # init img
        _ = model(img.half() if half else img) if device.type != 'cpu' else None  # run once
        path = data['test'] if opt.task == 'test' else data['val']  # path to val/test images
        dataloader = create_dataloader(path, imgsz, batch_size, int(max(model.stride)), opt,
                                       hyp=None, augment=False, cache=False, pad=0.5, rect=True)[0]

    seen = 0
    names = model.names if hasattr(model, 'names') else model.module.names
    coco91class = coco80_to_coco91_class()
    s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', '[email protected]', '[email protected]:.95')
    p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0.
    loss = torch.zeros(3, device=device)
    jdict, stats, ap, ap_class = [], [], [], []
    for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)):
        img = img.to(device)
        img = img.half() if half else img.float()  # uint8 to fp16/32
        img /= 255.0  # 0 - 255 to 0.0 - 1.0
        targets = targets.to(device)
        nb, _, height, width = img.shape  # batch size, channels, height, width
        whwh = torch.Tensor([width, height, width, height]).to(device)

        # Disable gradients
        with torch.no_grad():
            # Run model
            t = torch_utils.time_synchronized()
            inf_out, train_out = model(img, augment=augment)  # inference and training outputs
            t0 += torch_utils.time_synchronized() - t

            # Compute loss
            if training:  # if model has loss hyperparameters
                loss += compute_loss([x.float() for x in train_out], targets, model)[1][:3]  # GIoU, obj, cls

            # Run NMS
            t = torch_utils.time_synchronized()
            output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, merge=merge)
            t1 += torch_utils.time_synchronized() - t

        # Statistics per image
        for si, pred in enumerate(output):
            labels = targets[targets[:, 0] == si, 1:]
            nl = len(labels)
            tcls = labels[:, 0].tolist() if nl else []  # target class
            seen += 1

            if pred is None:
                if nl:
                    stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls))
                continue

            # Append to text file
            # with open('test.txt', 'a') as file:
            #    [file.write('%11.5g' * 7 % tuple(x) + '\n') for x in pred]

            # Clip boxes to image bounds
            clip_coords(pred, (height, width))

            # Append to pycocotools JSON dictionary
            if save_json:
                # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ...
                image_id = int(Path(paths[si]).stem.split('_')[-1])
                box = pred[:, :4].clone()  # xyxy
                scale_coords(img[si].shape[1:], box, shapes[si][0], shapes[si][1])  # to original shape
                box = xyxy2xywh(box)  # xywh
                box[:, :2] -= box[:, 2:] / 2  # xy center to top-left corner
                for p, b in zip(pred.tolist(), box.tolist()):
                    jdict.append({'image_id': image_id,
                                  'category_id': coco91class[int(p[5])],
                                  'bbox': [round(x, 3) for x in b],
                                  'score': round(p[4], 5)})

            # Assign all predictions as incorrect
            correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device)
            if nl:
                detected = []  # target indices
                tcls_tensor = labels[:, 0]

                # target boxes
                tbox = xywh2xyxy(labels[:, 1:5]) * whwh

                # Per target class
                for cls in torch.unique(tcls_tensor):
                    ti = (cls == tcls_tensor).nonzero().view(-1)  # prediction indices
                    pi = (cls == pred[:, 5]).nonzero().view(-1)  # target indices

                    # Search for detections
                    if pi.shape[0]:
                        # Prediction to target ious
                        ious, i = box_iou(pred[pi, :4], tbox[ti]).max(1)  # best ious, indices

                        # Append detections
                        for j in (ious > iouv[0]).nonzero():
                            d = ti[i[j]]  # detected target
                            if d not in detected:
                                detected.append(d)
                                correct[pi[j]] = ious[j] > iouv  # iou_thres is 1xn
                                if len(detected) == nl:  # all targets already located in image
                                    break

            # Append statistics (correct, conf, pcls, tcls)
            stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls))        # Plot images
        if batch_i < 1:
            f = 'test_batch%g_gt.jpg' % batch_i  # filename
            plot_images(img, targets, paths, f, names)  # ground truth
            f = 'test_batch%g_pred.jpg' % batch_i
            plot_images(img, output_to_target(output, width, height), paths, f, names)  # predictions

2.3.5 计算map数据

    # Compute statistics
    stats = [np.concatenate(x, 0) for x in zip(*stats)]  # to numpy
    if len(stats):
        p, r, ap, f1, ap_class = ap_per_class(*stats)
        p, r, ap50, ap = p[:, 0], r[:, 0], ap[:, 0], ap.mean(1)  # [P, R, [email protected], [email protected]:0.95]
        mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean()
        nt = np.bincount(stats[3].astype(np.int64), minlength=nc)  # number of targets per class
    else:
        nt = torch.zeros(1)

2.3.6 打印结果（图片，速度），保存结果至json，并返回结果

    # Print results
    pf = '%20s' + '%12.3g' * 6  # print format
    print(pf % ('all', seen, nt.sum(), mp, mr, map50, map))

    # Print results per class
    if verbose and nc > 1 and len(stats):
        for i, c in enumerate(ap_class):
            print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i]))

    # Print speeds
    t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size)  # tuple
    if not training:
        print('Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t)

    # Save JSON
    if save_json and map50 and len(jdict):
        imgIds = [int(Path(x).stem.split('_')[-1]) for x in dataloader.dataset.img_files]
        f = 'detections_val2017_%s_results.json' % \
            (weights.split(os.sep)[-1].replace('.pt', '') if weights else '')  # filename
        print('\nCOCO mAP with pycocotools... saving %s...' % f)
        with open(f, 'w') as file:
            json.dump(jdict, file)

        try:
            from pycocotools.coco import COCO
            from pycocotools.cocoeval import COCOeval

            # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
            cocoGt = COCO(glob.glob('../coco/annotations/instances_val*.json')[0])  # initialize COCO ground truth api
            cocoDt = cocoGt.loadRes(f)  # initialize COCO pred api

            cocoEval = COCOeval(cocoGt, cocoDt, 'bbox')
            cocoEval.params.imgIds = imgIds  # image IDs to evaluate
            cocoEval.evaluate()
            cocoEval.accumulate()
            cocoEval.summarize()
            map, map50 = cocoEval.stats[:2]  # update results ([email protected]:0.95, [email protected])
        except:
            print('WARNING: pycocotools must be installed with numpy==1.17 to run correctly. '
                  'See https://github.com/cocodataset/cocoapi/issues/356')

    # Return results
    model.float()  # for training
    maps = np.zeros(nc) + map
    for i, c in enumerate(ap_class):
        maps[c] = ap[i]
    return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t

————————————————
版权声明：本文为CSDN博主「Liaojiajia2019」的原创文章，遵循CC 4.0 BY-SA版权协议，转载请附上原文出处链接及本声明。
原文链接：https://blog.csdn.net/mary_0830/article/details/107081030