你好，我在训练自己标注的数据时会出现标签越界的报错，然后加入一些调试代码后出现以下错误：

`Traceback (most recent call last):

File "train.py", line 159, in train(net, train_loader, loss_dict, optimizer, scheduler, logger, epoch, metric_dict, cfg.use_aux) File "train.py", line 71, in train loss = calc_loss(loss_dict, results, logger, global_step) File "train.py", line 41, in calc_loss assert torch.all(results['seg_label'] < 5), results['seg_label'][results['seg_label'] >= 5] # 5 is num_lanes + 1

AssertionError: tensor([255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255], device='cuda:0') `

请问这个是怎么造成的呢，我标注的车道线并没有超过四条。

Hi there, I am currently willing to compile the C++ evaluation tools for culane following the instructions in INSTALL.md. There is a subtle difference of vs version between mine(vs2019 community) and the one used in the given command, "cmake .. -G "Visual Studio 15 2017 Win64" " which is vs2017. I have tried some possible commands according to my setup like, cmake .. -G "Visual Studio 16 2019 Win64" , cmake .. -G "Visual Studio 15 2017 Win64" , cmake .. -G "Visual Studio 16 2019" -A x64 which I found here , cmake .. -G "Visual Studio 16 2019 Win64" -A x64 , but neither works...

Here are the cmake error messages. Is there any chance I can fix this up without install a 2017vs by modifying the cmakelist or any other configuration? Thanks in advance!

I also try to directly open the folder with vs2019, but get a following error.

CMake Error at E:\workstation\lanedet\UltraFast\evaluation\culane\CMakeLists.txt:18 (string): string sub-command REPLACE requires at least four arguments. culane_evaluator E:\workstation\lanedet\UltraFast\evaluation\culane\CMakeLists.txt 18

I changed the input size to 256*256 and change anchor to [108, 116, 125, 133, 142, 151, 160, 168, 177, 186, 195, 203, 212, 220, 229, 237, 246, 255], but when I test my model, the lane that predict by my model is really close to center like this: I change my hyper parameter according to this issue:https://github.com/cfzd/Ultra-Fast-Lane-Detection/issues/124

The orange line is the result which trained with following cfg:

Config (path: configs/culane.py): {'dataset': 'CULane', 'data_root': '/data/home/xxxx/dataset/CULane', 'epoch': 50, 'batch_size': 128, 'optimizer': 'SGD', 'learning_rate': 0.1, 'weight_decay': 0.0001, 'momentum': 0.9, 'scheduler': 'multi', 'steps': [25, 38], 'gamma': 0.1, 'warmup': 'linear', 'warmup_iters': 695, 'use_aux': True, 'griding_num': 200, 'backbone': '18', 'sim_loss_w': 0.0, 'shp_loss_w': 0.0, 'note': 'First train on CULane', 'log_path': './train_log', 'finetune': None, 'resume': None, 'test_model': None, 'test_work_dir': './tmp'}

The blue line is the result which trained with following cfg:

Config (path: configs/culane.py): {'dataset': 'CULane', 'data_root': '/data/home/xxxx/dataset/CULane', 'epoch': 50, 'batch_size': 128, 'optimizer': 'SGD', 'learning_rate': 0.1, 'weight_decay': 0.0001, 'momentum': 0.9, 'scheduler': 'multi', 'steps': [25, 38], 'gamma': 0.1, 'warmup': 'linear', 'warmup_iters': 695, 'use_aux': True, 'griding_num': 200, 'backbone': '18', 'sim_loss_w': 1, 'shp_loss_w': 1, 'note': 'Train on CULane with other loss', 'log_path': './train_log', 'finetune': None, 'resume': None, 'test_model': None, 'test_work_dir': './tmp'}

The difference between cfgs is that the orange one do not applied sim_loss and shp_loss ( the weight of them are set to 0.0 ) and blue one applied with weight 1.0 and 1.0 ( which used in the paper ). Such kind of difference made loss ill. According my observation, the loss gone into mad in second epoch and here are some log from stdout:

.......
 100%|█████████▉| 694/695 [42:47<00:04,  4.02s/it, data_time=1.159, iou=0.194, loss=nan                                                    100%|█████████▉| 694/695 [42:48<00:04,  4.02s/it, data_time=0.055, iou=0.194, loss=nan                                                    100%|██████████| 695/695 [42:48<00:00,  2.96s/it, data_time=0.055, iou=0.194, loss=nan                                                    100%|██████████| 695/695 [42:48<00:00,  3.70s/it, data_time=0.055, iou=0.194, loss=nan                                                    , net_time=0.213, top1=0.001, top2=0.002, top3=0.003]
  0%|          | 0/695 [00:14<?, ?it/s, data_time=13.690, iou=0.194, loss=nan, net_tim                                                      0%|          | 1/695 [00:14<2:51:37, 14.84s/it, data_time=13.690, iou=0.194, loss=na                                                      0%|          | 1/695 [00:16<2:51:37, 14.84s/it, data_time=0.043, iou=0.194, loss=nan                                                      0%|          | 2/695 [00:16<2:04:03, 10.74s/it, data_time=0.043, iou=0.194, loss=nan                                                      0%|          | 2/695 [00:17<2:04:03, 10.74s/it, data_time=0.043, iou=0.195, loss=nan                                                      0%|          | 3/695 [00:17<1:30:50,  7.88s/it, data_time=0.043, iou=0.195, loss=nan                                                      0%|          | 3/695 [00:18<1:30:50,  7.88s/it, data_time=0.059, iou=0.195, loss=nan  
......

Is that gradient explosion? But it seems quite good in first epoch. Furthermore, I use single tesla V100 to train and batchsize is set to 128 according former issue

大佬，我在测试model时发现车辆在变道时，检测出的车道会扭曲变形严重，变道结束后，效果就恢复良好了，分析了下，原因感觉是变道过程中车辆骑在车道线上行驶，导致模型对这条骑着的车道线进行分类时，无法区分是左车道还是右车道，所以产生一个中间临界状态，不知这个解释是否正确，以及这种变道过程中检测效果变差的case该如何解决呢， 拜谢大佬！

你好，我用您提供的demo.py测试时候出现： Expected 4-dimensional input for 4-dimensional weight [64, 3, 7, 7], but got 3-dimensional input of size [3, 288, 800] instead 请问这可能是什么问题呢，该如何解决呢？比较小白，请指教，谢谢了。代码如下：

` import torch, os, cv2 from model.model import parsingNet from utils.common import merge_config from utils.dist_utils import dist_print import torch import scipy.special, tqdm import numpy as np import torchvision.transforms as transforms from data.dataset import LaneTestDataset

if name == "main": torch.backends.cudnn.benchmark = True args, cfg = merge_config() dist_print('start testing...') assert cfg.backbone in ['18','34','50','101','152','50next','101next','50wide','101wide']

if cfg.dataset == 'CULane':
    cls_num_per_lane = 18
elif cfg.dataset == 'Tusimple':
    cls_num_per_lane = 56
else:
    raise NotImplementedError

net = parsingNet(pretrained = False, backbone=cfg.backbone,cls_dim = (cfg.griding_num+1,cls_num_per_lane,4),
                use_aux=False).cuda() # we dont need auxiliary segmentation in testing

state_dict = torch.load(cfg.test_model, map_location='cpu')['model']
compatible_state_dict = {}
for k, v in state_dict.items():
    if 'module.' in k:
        compatible_state_dict[k[7:]] = v
    else:
        compatible_state_dict[k] = v

net.load_state_dict(compatible_state_dict, strict=False)
net.eval()

img_transforms = transforms.Compose([
    transforms.Resize((288, 800)),
    transforms.ToTensor(),
    transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])


cap = cv2.VideoCapture("0000.mp4")
frame = cap.read()
fourcc = cv2.VideoWriter_fourcc(*'MJPG')
vout = cv2.VideoWriter(str(111)+'.avi', fourcc , 30.0, (1280,720))
from PIL import Image
print('加载cuda是否成功：',torch.cuda.is_available())
while 1:
    rval,frame = cap.read()
    if rval == False:
        break
    cv2.imwrite("ssss.jpg",frame)
    img_ = Image.open("ssss.jpg")

    imgs = img_transforms(img_)
    imgs = imgs.cuda()
    with torch.no_grad():
        out = net(imgs)

    col_sample = np.linspace(0, 800 - 1, cfg.griding_num)
    col_sample_w = col_sample[1] - col_sample[0]


    out_j = out[0].data.cpu().numpy()
    out_j = out_j[:, ::-1, :]
    prob = scipy.special.softmax(out_j[:-1, :, :], axis=0)
    idx = np.arange(cfg.griding_num) + 1
    idx = idx.reshape(-1, 1, 1)
    loc = np.sum(prob * idx, axis=0)
    out_j = np.argmax(out_j, axis=0)
    loc[out_j == cfg.griding_num] = 0
    out_j = loc

    # import pdb; pdb.set_trace()
    # vis = cv2.imread(os.path.join(cfg.data_root,names[0]))
    for i in range(out_j.shape[1]):
        if np.sum(out_j[:, i] != 0) > 2:
            for k in range(out_j.shape[0]):
                if out_j[k, i] > 0:
                    ppp = (int(out_j[k, i] * col_sample_w * 1280 / 800) - 1, int(720 - k * 20) - 1)
                    cv2.circle(img_,ppp,5,(0,255,0),-1)
    vout.write(img_)
    
vout.release()

`

Hi,

I've been recently experimenting with using different feature extractor, in this case MobilenetV3 and efficientNet-B2. Results with MobilenetV3 was acceptable but in the case of efficientNet I came across a weird phenomenon.

First, I changed the following things:

The Dense head has now been replaced by 4 branches each containing 3 sets of FC at 256, 256 and 128. When using the segmentation branch I adopted a NAS-net like structure essentially pooling from the last layer and a previous block where the stride is at 16. Also in efficientNet instead of the convolution with filter 8, I used convolution with filter 32.

In both case I am not using structural losses.

in the case of Mobilenet the experiment worked. I was able to get a functional model. However with efficienNet although the loss was a lot smaller the predicitons seems to be just random. Any thoughts on this?

Thanks again for the paper and repository.

1、因图片数据需要，把culane数据集裁剪出中间部分，图像大小为1280*590，那个row_anchor保持相同18个； 2、使用默认超参数culane.py，那个shm和shp的损失权重设置为0，进行训练，那个relation_loss和relation_dis损失一直上升，训练25epoch后，检测效果不好 3、有些人提到shm和shp的损失权重分别设置为0.8，0.5，进行训练，那个relation_dis损失下降，relation_loss一直上升，同样检测效果不好

考虑culane的场景比较丰富，加上自己场景组成混合数据集，进行训练，不知道能否达到目的？

I trained a model on CULane, using ResNet-34 as the backbone, and the model's F1 on the test set was only 70.5. The only modification I did was changing the backbone line in the config file to backbone = '34'. Are there any more changes that are necessary to achieve the reported 72.3% F1 in the paper?

I wanted to train the model with an input shape of 256 x 512 rather than 288 x 800. While changing the shape I get lots of errors with the unsupported shape. How do we try to make this transition easy and code to work for every different input shape? or u recommend converting all FC layers to conv

Hello @cfzd,

thank you for such amazing work.

I want to know if there are any methods to extract the array of detected lane pixels? As I want to try to calculate the lane curvature using the detected lanes.

Any approaches that I can use?

Thanks.

在使用自己的数据跑ufld算法时，我发现用test.py测试出的结果并不能很好的代表模型的能力。比如我使用训练过的数据进行测试，测试的准确率仅仅在0.3左右，但是用demo.py跑出的可视化结果可以看出，推理结果是与标注基本一致的（毕竟是训练过的数据），所以我仔细对比了一下两个代码的推理部分，发现在demo.py中，代码用到了row_anchor作为评价标准， ppp = (int(out_j[k, i] * col_sample_w * img_w / 800) - 1, int(img_h * (row_anchor[cls_num_per_lane-1-k]/288)) - 1 ) 而在test.py引用的eval_wrapper.py中，row_anchor部分被写成如下： fp.write( '%d %d ' % (int(out_j[k, i] * col_sample_w * 1640 / 800) - 1, int(590 - k * 20) - 1)) 其中int(590 - k * 20) - 1并未考虑resize成288的操作，区域的划分也比较粗暴，所以我改成了demo.py中的写法，再次测试训练过的数据的结果接近1。所以我想测试部分可能并没有完全展现出算法的性能。我仅在自己的数据上进行了测试，那么我猜想在culane数据集上是否已可以通过这样的改动提升算法的准确率呢，大家有兴趣可以尝试一下。

Hi @cfzd , thanks for your great work, it's really impressive!

i have trained ufast on my dataset, and i find running test.py ,i can get the accuaacy 、fp、fn， but how can i get top1 top2 top3? hope your reply!