Hi, nice work. I am very interested in your paper, which gives me a lot of hints. I am a little confused about the formula 2 in the paper. May I ask why the negative vector is written like this? May I ask can you share your derivation process to help me understand? Thank you very much.

what's the gt_inst_mask dimension here: https://github.com/duwt/OrienMask/blob/61a0aa8f6b34c699962f83bb6c8ce557337ea701/eval/orienmask_yolo_loss.py#L259 ?

is it same wh with original image size or just the mask patch same wh with box?

Thanks your works! I has confused about recover mask, and got orien[9, 2, H, W] convert to Tensor [1, 18, H, W] in tensorrt , I has recover boxs in origin image, ,the x<bbox.widthorien_threshold && y< bbox.height orien_threshold in code , I just need to iterate through this 18-channel picture (Tensor [1, 18, H, W] )to filter out the pixel values?

Hello, while running the training module, I have encountered a strange error which says that the nms_cpu file cannot be imported. See below for error description:

Please note that the training module was initiated using python 3.7. When using Python 3.9, the error is slightly different which says that there is a loop in imports ... Please see below:

Could you tell me why this is happening? Your help is immensely appertained.

很感谢您的工作，我在测试和评估都没遇到问题。 在默认设置，训练的时候出现pred_wh not finite。我查看issue，Increasing the learning rate warm-up steps from 500 to 1000 makes the training procedure more stable. 但是还是会报错。之后我又把warm-up的step改为500，将config中设置(n_gpu=1, batch_size=8, accumulate=2还是会出现一样的错误。

pytorch1.8+cuda11.1，py=3.8 单卡

Hello, When I try to run the code I got this error.

Traceback (most recent call last): File "/home/swap/miniconda3/envs/orienmask/lib/python3.7/site-packages/IPython/core/interactiveshell.py", line 3441, in run_code exec(code_obj, self.user_global_ns, self.user_ns) File "", line 1, in runfile('/home/swap/Downloads/OrienMask-master/train.py', wdir='/home/swap/Downloads/OrienMask-master') File "/home/swap/.local/share/JetBrains/Toolbox/apps/PyCharm-C/ch-0/202.7660.27/plugins/python-ce/helpers/pydev/_pydev_bundle/pydev_umd.py", line 197, in runfile pydev_imports.execfile(filename, global_vars, local_vars) # execute the script File "/home/swap/.local/share/JetBrains/Toolbox/apps/PyCharm-C/ch-0/202.7660.27/plugins/python-ce/helpers/pydev/_pydev_imps/_pydev_execfile.py", line 18, in execfile exec(compile(contents+"\n", file, 'exec'), glob, loc) File "/home/swap/Downloads/OrienMask-master/train.py", line 33, in raise AssertionError("Configuration file need to be specified.") AssertionError: Configuration file need to be specified.

I am not sure if I understand it right, but I am very confused about this line code:

in the postprocess abtain masks. Where you multiply grid_sizes, which defined as 17x17, 34x34, 68x68 and same as nH, nW since you have this code:

But the boxes wh have this operation before:

why divide nH and nW and then multiply it back later. Is that necessary?

(orienmask) swap@pop-os:~/Downloads/OrienMask-master$ CUDA_VISIBLE_DEVICES=0 python train.py -c orienmask_yolo_coco_544_anchor4_fpn_plus [DarkNet53] Load pretrained model checkpoints/pretrained/pretrained_darknet53.pth Set checkpoint directory: checkpoints/OrienMaskAnchor4FPNPlus_0901_125918 2021-09-01 12:59:18,239

2021-09-01 12:59:18,240 [EPOCH 1] 2021-09-01 12:59:18,240 Train on epoch 1 2%|▌ | 682/29572 [03:55<2:44:38, 2.92it/s, lr=6.40e-04, loss=284.6246]pred_wh not finite 2%|▌ | 682/29572 [03:55<2:46:02, 2.90it/s, lr=6.40e-04, loss=284.6246]

In train phase, why does "pred_wh not finite" appear and how to fix it ?? thanks. 2021-08-13 11:54:59,609 [EPOCH 1] 2021-08-13 11:54:59,609 Train on epoch 1 0%|▍ | 1/313 [00:05<28:19, 5.45s/it, loss=-1.0000, lr=-1.00e0]2021-08-13 11:55:05,191 Reducer buckets have been rebuilt in this iteration. 92%|██████████████████████████████████████████████████████████████████████████████████████████████████████████▏ | 289/313 [02:44<00:13, 1.76it/s, lr=5.50e-04, loss=1257.3200]pred_wh not finite pred_wh not finite 92%|██████████████████████████████████████████████████████████████████████████████████████████████████████████▏ | 289/313 [02:44<00:13, 1.76it/s, lr=5.50e-04, loss=1257.3200] INFO:torch.distributed.elastic.agent.server.api:[default] worker group successfully finished. Waiting 300 seconds for other agents to finish.

Thanks your works! I has confused about recover mask, and got orien(92HW) convert to Tensor (118HW) in tensorrt , I has recover boxs in origin image, ,the x<bbox.widthorien_threshold && y< bbox.height orien_threshold in code , I just need to iterate through this 18-channel picture (Tensor (118H*W) )to filter out the pixel values?

作者你好 我最近在调试你的程序，发现一个很奇怪的问题，最开始我在A100服务器，cuda11.3，pytorch1.10的环境下训练，一切正常。但是换了其他的环境，训练过程中就会出现loss很大，最后报pred_wh not finite的错误。由于是梯度爆炸，我尝试把学习率改小就训练正常了。请问作者，这个问题如何解决，可以告诉我，你的训练环境吗？谢谢，期待你的回复

I use CUDA_VISIBLE_DEVICES=0,1 bash ./dist_train.sh -c orienmask_yolo_coco_544_anchor4_fpn_plus to train. There also a error "pred_wh not finite" I just modify the datasets with cityscapes ,and modify image-size [1024,512],num_classes=8 and so on. and i don not know why?

I am using exactly same backbone of YOLOX-s, but the AP is very low, but the mask and box seems normal. AP about 21.

Do u know why? Have a test with some light backbone models? Hope you can did more experiment on different backbone on the performance changes.