zhang-tao-whu / e2ec Goto Github PK

Q1： Is it possible to effectively fit the regression points to the real points in discrete spatial dimensions with a simple l1 loss?

Thank you very much for your work, do you know why this method needs to train more than 100 epochs, the convergence speed is very slow, which makes training a lot of inconvenience, some instance segmentation methods only need more than 20 epochs, is it because of CenterNet?

Thanks for your excellent work!
I wonder when the multi GPU training codes will be released, which is very important for reproduction and consequent research.
I really appreciate your generosity and kindness.

When there are more than 100 instances in each image, which parameter should I modify?

Hi, thank you for your work! I was wondering if it is possible to keep the org resolution during visualization, to create example Images like in your paper.

Thank you!

Hello! Thank you for your amazing work.

I used your network (with the coco config) to train on a dataset of custom imagery with coco style object instance annotations. Now I wish to generate predictions for the test split of this dataset for which I have no annotation file. I notice you are using an empty annotation file to perform evaluations on the coco test dev split. Could you kindly guide me on how I can perform single image predictions (object polygons and seg masks for each image) on my custom dataset using the model I trained? Looking forward to your reply.

Thank you!

Have you done some comparative experiments about the model with the same backbone？For example, replacing your backbone to resnet50?

sir, 2 days ago, I saw the project related to DVIS, but now, I cannot found that repository.. :( Do you have any plan to reopen that project?

Hi, thank you very much for your work. Is the pre-trained model (eg. model_sbd.pth) you provided the training result parameters of the 149th epoch? Or the epoch with the best performance among all the training epochs? Because no relevant information was found in the model.pth for you provided.

I want to train myself datasets,I did it with the readme,but it occured error!Could you help me solve it problem?

SBD数据集的格式跟其他数据集不太像
所以套用readme中的格式会报错
可以提供下SBD数据集的visualize方式吗
万分感谢！

您好，我是新手，请问您的代码是基于Linux的吗

How to train ous dataset?
when i train in my dataset it get

pos_loss = torch.log(pred) * torch.pow(1 - pred, 2) * pos_inds
RuntimeError: The size of tensor a (80) must match the size of tensor b (2) at non-singleton dimension 1

Is the classes number problem?
I can not find the classes config

我使用了mmcv中的dcn，并按照您的说明和问题23修改了代码，除此以外没有修改。运行时有一些警告但没有报错，但准确率一直是0，可视化图像也是错误的，coco和kitti两个数据集都是一样的情况，想请教一下是不是哪里操作不当？
测试时使用的代码：
python test.py kitti --checkpoint model/model_kitti.pth --stage coarse
python visualize.py coco data/coco --checkpoint model/model_coco.pth --with_nms True --output_dir data/result/coco
得到的可视化图像是一团混乱的线条，能看到线条下方有缩小的原图
运行时的警告：
load model: model/model_kitti.pth
loading annotations into memory...
Done (t=3.27s)
creating index...
index created!
命令语法不正确。
loading annotations into memory...
Done (t=3.30s)
creating index...
index created!
0%|

UserWarning: Named tensors and all their associated APIs are an experimental feature and subject to change. Please do not use them for anything important until they are released as stable. (Triggered internally at ..\c10/core/TensorImpl.h:1156.)
return torch.max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)

UserWarning: Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details.
"Default grid_sample and affine_grid behavior has changed "

dear author:
我看到您论文里写用dla-34作为backbone，dla-34是什么呢，有没有参考资料？

hi，你好，我将cfg.model.use_dcn 设置为False，但是训练程序运行起来还是会import dcn导致出错，这块该如何处理呢？

Hello, when we train on the KITTI dataset, the input size is (384,896) or (512,896)?

哪里出了问题，我自己的数据集只有一个细胞类别，‘ct-hm’设为2。

I set use_dcn = False,dowload the pretrained model and the dataset to test .The code run successfully, but the result is 0.

Due to my GPU device limit, I change the batchsize from 32 to 8 in cityscapes training, but it got an unexpected error that in the computation of loss, the tensor missmatch the size. I am comfused what's wrong with this error if I lower the batchsize. Or this code only can train cityscapes in batchsize of 32?

感谢作者大大的作品·！
请问有没有办法在visualize的时候，即显示轮廓，又能把物体的bbox记录、显示出来。像下面这样：

I have several questions about MDA and DML.

1. In DML, the key point is needed to calculate the loss. How to obtain the key point in the training? Does it comes from annotation information?
2. In MDA, how to use the fixed vertices? Is it used to split the contour to several segments, and use there segments to calculate loss? It seems an improvement for the Segment-wise Matching Scheme in DANSE . Am I correct?

The model actually produces five polygons: init, coarse, final1, final2, and final3.
Coarse is refined using the proposed global deformation module, while final* are refined using the circular convolution module.
What if we use all global deformation modules or circular convolution modules for all the polygon refinement?

您好，十分想学习您的工作成果，但是很多数据集无法访问了，可以麻烦您更新一下吗？感谢！

Hi, thank you for your work! Could you update the download link of cityscapes_anno.tar.gz file? Now click on it to show that the link has expired.

Thank you!

In the file dataset/data_loader.py, you have set num_workers=batch_size in the function make_ddp_train_loader. Is there any specific reason for this? You mentioned in issue #13 that you've done this for convenience. Can you please elaborate on this?

This looks like it should instead be num_workers=train.cfg.num_workers. Please let me know if this is correct.

Thank you!

Hi, I tried the multi-gpu training code but the program always got stuck after a few iterations.

Environment:

• pytorch 1.7.1
• cuda 10.2
• gcc version 7.5.0
• Ubuntu 18.04.3 LTS

Reproduce the bug:
CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node 4 train_net_ddp.py --config_file coco --gpus 4

Output:

  File "/home/a/anaconda3/envs/e2ec/lib/python3.7/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/home/aa/anaconda3/envs/e2ec/lib/python3.7/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/home/a/anaconda3/envs/e2ec/lib/python3.7/site-packages/torch/distributed/launch.py", line 260, in <module>
    main()
  File "/home/a/anaconda3/envs/e2ec/lib/python3.7/site-packages/torch/distributed/launch.py", line 253, in main
    process.wait()
  File "/home/a/anaconda3/envs/e2ec/lib/python3.7/subprocess.py", line 1019, in wait
    return self._wait(timeout=timeout)
  File "/home/a/anaconda3/envs/e2ec/lib/python3.7/subprocess.py", line 1653, in _wait
    (pid, sts) = self._try_wait(0)
  File "/home/a/anaconda3/envs/e2ec/lib/python3.7/subprocess.py", line 1611, in _try_wait
    (pid, sts) = os.waitpid(self.pid, wait_flags)
KeyboardInterrupt```

Thanks for the excellent paper and code.
But my experiments (DCN disabled) on KINS dataset always deadlock at epoch 20 with GPU and CPU busy.
I use command python -m torch.distributed.launch --nproc_per_node 2 train_net_ddp.py --config_file kitti --bs 4 --gpus 2.
Is there any advice to check? Thanks advance.

BTW, what are the minimum bs and epoch in KINS dataset? Bs 64 and epoch 150 seem too huge.

Hello, thank you for your awesome work.
I am currently training a model on the Cityscapes dataset. Apart from setting the batch size to 8 in the config, everything else matches your config file. However, the final result I obtained is only around 0.27 AP, which is significantly lower than the 0.34 provided with your pretrained model. I would like to ask if there are any differences between the config used for your best model and the one you provided?

dear author:
1、return self.setloss(ini_pred_poly, pred_polys_, gt_polys, keyPointsMask) 为什么不直接用pred_polys_去计算最近距离，而是采用ini_pred_poly去计算最近距离呢？
2、ground truth的轮廓坐标如何获得？

Hi, thanks for your excellent work!

I tried to train this work on COCO dataset, but it converges very slowly and needs to train more than 100 epochs.
I notice that you mention it is because of CenterNet(detector) in previous issue, and you rebuild this network in FCOS version.

I wonder when will the FCOS version be released?
Or can you explain how do I change the detector? Which files and what functions should I change?
Thanks! :)

Hi, thanks for your wonderful work. But I found a bug in your README file. In the H3 title, Testing on COCO, the pretrained model URL should be gpcv.whu.edu.cn/member/ZhangTao/model.zip. However, Github converts it into a relative link in your repo, resulting in a 404 page. You can add https:// ahead of your pretrained model URL to avoid it.

Thanks very much for your work, I have some questions about get_gcn_feature function. In this function, we have input 'cnn_feature','img_poly' . Our main focus is to extract the features of img_polys‘ points from the feature map, is this correct? If so, we need to use torch.nn.functional.grid_sample to finish this task. So why in your code, you didn't normalize img_polys between [-1,1] to target the location in cnn_feature instead
img_poly[..., 0] = img_poly[..., 0] / (w / 2.) - 1
img_poly[..., 1] = img_poly[..., 1] / (h / 2.) - 1

I dont understand how this works, I hope you can help me out.

In torch.nn.functional.grid_sample(input, grid, mode='bilinear', padding_mode='zeros', align_corners=None)
grid specifies the sampling pixel locations normalized by the input spatial dimensions. Therefore, it should have most values in the range of [-1, 1]

Hello, Could you tell me how to modify the size when I want to test the coco datasets, because the coco dataset is of any size.I have try once,but the AP was 0, so I think weather the size is wrong.

使用Readme里面的visualize.py推理代码，预训练权重是作者提供的COCO权重，可视化结果为什么是这样的？换了其他图片和权重还是乱七八糟的线条

when i run test.py show the error:
python=3.7

and i think the same one like CharlesShang/DCNv2#24.
can u help me to solve this problem?
Thanks！

author您好，请问一下有没有在Windows系统下的安装方法呢？

Hello, if the number of instances in one batch is zero, the smooth L1 loss(Linit, Lcoarse, Liter) will be 'nan'. How can I address this problem?

Thanks for your work!
But I am confused about the Multi-direction alignment (MDA) part, can you explain in more detail?
Which part of the code does the MDA correspond to? If I want to try to change the value of M, where should I adjust it?