Traceback (most recent call last): File "/home/fhl/src/rgbt-det/multispectral-object-detection/train.py", line 1028, in train_rgb_ir(hyp, opt, device, tb_writer) File "/home/fhl/src/rgbt-det/multispectral-object-detection/train.py", line 519, in train_rgb_ir model = Model(opt.cfg or ckpt['model'].yaml, ch=6, nc=nc, anchors=hyp.get('anchors')).to(device) # create File "/home/fhl/src/rgbt-det/multispectral-object-detection/models/yolo_test.py", line 188, in init self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch]) # model, savelist File "/home/fhl/src/rgbt-det/multispectral-object-detection/models/yolo_test.py", line 529, in parse_model elif m is CMAFF: NameError: name 'CMAFF' is not defined

在common.py文件中发现CMAFF未定义,请问是什么原因导致的?

Thank you for your nice work.

I train the model with the option as below :

parser.add_argument('--cfg', type=str, default='./models/transformer/yolov5l_fusion_transformerx3_FLIR_aligned.yaml', help='model.yaml path')

Then, I can't reproduce the results on FLIR dataset in the paper.

The results of training process

The result of best checkpoint

Could you provide the checkpoint on FLIR dataset?

Also, i got difference baseline result when i train the add-fusion model.

# train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/] train_rgb: /home/fqy/Data/FLIR_ADAS_1_3/align/yolo/rgb/align_train.txt # 128 images val_rgb: /home/fqy/Data/FLIR_ADAS_1_3/align/yolo/rgb/align_validation.txt # 128 images train_ir: /home/fqy/Data/FLIR_ADAS_1_3/align/yolo/ir/align_train.txt # 128 images val_ir: /home/fqy/Data/FLIR_ADAS_1_3/align/yolo/ir/align_validation.txt # 128 images 对于这种方式的目录结构和.txt的内容不是很清楚，所以想像你要一份yolov5 format的llvip数据，非常感谢！

How do you do? May I ask why CLS is always 0 when I train Epoch gpu_mem box obj cls total labels img_size 3/199 7.64G 0.04769 0.02945 0 0.07714 7 640: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 6013/6013 [24:45<00:00, 4.05it/s] Class Images Labels P R [email protected] [email protected] [email protected]:.95: 100%|████████████████████████████████████████████████████████████████████████████████████████████████| 866/866 [01:19<00:00, 10.83it/s] all 3463 7931 0.821 0.803 0.802 0.289 0.371

I can't reproduce the result of paper with provided checkpoint. There is a small difference. Could you give me the label files of FLIR datsets..? Also, Do you have the mAP result of single modality on each of the topics..? and Could you provide it ..?

Thank you so much, looking forward to your reply!

paper

result with provided checkpoint

python .\detect_twostream.py --weights=./pre_models/yolov5l_transformerx3_llvip_s1024_bs32_e200 --source1=C:/HMZ/work/project/datasets/LLVIP/infrared/test --source2=C:/HMZ/work/project/datasets/LLVIP/visible/test --img-size=1024 --device=0

According to the image in your paper, the YAML file should be two streams of input RGB image and IR image respectively. Why the input of the second stream [-4, 1, Focus, [64, 3]]，-4 means from layer 6 instead of IR images? Why not for example [-10,1, Focus, [64, 3]] （of course this will not work）

作者你好， 如这里所述，LLVIP数据集自发表以后被更新了一次，更新后的LLVIP数据集的行人检测精度也提高了，即： 从图中可以看出，更新后的LLVIP仅在infrared模态上的检测结果（0.67）就优于您论文中的结果（0.636）： 考虑到LLVIP被更新过一次，于是我将您的代码在新的LLVIP上重新训练了一次，batch_size=8x4gpus, image_size=1024. cfg is here, hyp is here. 我训练的结果为0.656，远低于LLVIP中仅使用infrared的结果。 为了实验的公平性，我已确认您的代码和LLVIP开源的代码都是YOLOv5 (version5.0)。请看LLVIP作者的回复，以及您之前的回复.

以上实验结果令我非常困惑，请问您是否能提供在新的LLVIP数据集上的CFT的mAP呢？这对后来者与您的方法做公平对比非常重要

期待您的回复

Namespace(augment=False, batch_size=64, conf_thres=0.001, data='/home/delll/mark1/multispectral-object-detection/data/multispectral/FLIR_aligned.yaml', device='', exist_ok=False, img_size=640, iou_thres=0.5, name='exp', project='runs/test', save_conf=True, save_hybrid=False, save_json=False, save_txt=True, single_cls=False, task='val', verbose=False, weights='/home/delll/mark1/multispectral-object-detection/best.pt') /home/delll/mark1/multispectral-object-detection/data/multispectral/FLIR_aligned.yaml YOLOv5 🚀 387dc3e torch 1.13.0+cu117 CPU

Fusing layers... Model Summary: 1015 layers, 206209480 parameters, 0 gradients val Traceback (most recent call last): File "/home/delll/mark1/multispectral-object-detection/utils/datasets.py", line 875, in init raise Exception(f'{prefix}{path_rgb} does not exist') Exception: val: /home/fqy/Data/FLIR_ADAS_1_3/align/yolo/rgb/align_validation.txt does not exist

During handling of the above exception, another exception occurred:

Traceback (most recent call last): File "test.py", line 339, in opt=opt File "test.py", line 94, in test prefix=colorstr(f'{task}: '))[0] File "/home/delll/mark1/multispectral-object-detection/utils/datasets.py", line 236, in create_dataloader_rgb_ir prefix=prefix) File "/home/delll/mark1/multispectral-object-detection/utils/datasets.py", line 899, in init raise Exception(f'{prefix}Error loading data from {path_rgb,path_ir}: {e}\nSee {help_url}') Exception: val: Error loading data from ('/home/fqy/Data/FLIR_ADAS_1_3/align/yolo/rgb/align_validation.txt', '/home/fqy/Data/FLIR_ADAS_1_3/align/yolo/ir/align_validation.txt'): val: /home/fqy/Data/FLIR_ADAS_1_3/align/yolo/rgb/align_validation.txt does not exist See https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data

#Can't get attribute 'SPPF' on <module 'models.common' from 'G:\pythonfile\multispectral-object-detection-main\multispectral-object-detection-main\models\common.py'> How to solve this problem?