xiangyu-cas / aicity2020-voc-reid Goto Github PK

File "/home/hlq/llj/vehicle-ReID-baseline-master/vehiclereid/losses/circle_loss.py", line 23, in call
sim_mat = F.linear(F.normalize(bn_feat), F.normalize(self.weight))
File "/home/hlq/anaconda3/lib/python3.6/site-packages/torch/nn/functional.py", line 1408, in linear
output = input.matmul(weight.t())
RuntimeError: Expected object of backend CUDA but got backend CPU for argument #2 'mat2'

Greetings,

I can't seem to reproduce the results that you've achieved on VeRI using your pre-trained models.

I ran the script (scripts/veri/test.sh) with the pretrained models from here.

However, my mAP is only 37.7% compared to the >80% shown in your log files. Here's my log.

Could you please offer some detailed instructions on how to test on VeRi using your pre-trained weights?

Thank you!

I can't find aicity20/0326-search/augmix/best.pth

Hi, could you please clarify whether the results on VeRi-776 in your paper are obtained using image-to-track or image-to-image test fashion?

感谢作者开源如此优秀的代码，但是有个问题一直困扰着我
fm[0:4, :] = 0
fm[12:16, :] = 0
fm[:, 0:4] = 0
fm[:, 12:16] = 0
在weakly_supervised_crop_aug.py文件中这几行代码的作用是什么呢，这个需要根据特征图大小来修改吗？

您好，
我在VeRi-766上复现了基于ResNet50_ibn_a基础网络的训练结果，mAP达到了81.7%。

之后运行test.sh，对模型进行推理时，mAP只有65.6%。请问可能是什么原因造成？
谢谢！

Hello,

If we just want to reproduce the result on AICity 2020 Challenge of your code, I have downloaded the pretrained model 0409-ensemble but i get the error that I do not have the file feats.npy. ¿Should this file be in the pretrained folder 0409-ensemble?

Thanks

Hello, author
Where is the code for Data Augmentation: Weakly Supervised Detection?

Hi
Thank you for your great job.
I am trying to run test with the trained model you provided but I can't find feats.npy file.

Following is the error I got:.

Traceback (most recent call last):
File "./tools/aicity20/compute_distmat_from_feats.py", line 14, in
feat = np.load(src_dir + '/' + 'feats.npy')
File "/home/husan/.virtualenvs/re-id/lib/python3.8/site-packages/numpy/lib/npyio.py", line 416, in load
fid = stack.enter_context(open(os_fspath(file), "rb"))
FileNotFoundError: [Errno 2] No such file or directory: './output/aicity20/0409-ensemble/ReOriID//feats.npy'

大佬，我又来了，能留个邮箱或wx吗？
reid-strong-baseline 里面triplet loss和ID loss由BN层隔开,分别用了不同的feature
在你代码里用的是同一个feature，请问这个地方你是怎么考虑的？

您好，打扰一下，请教您一个问题，如果我只想拿到车辆最终的表示向量，可不可以不去计算向量间的dist?就是说，给我一张车辆的图片，我就直接获取这个车辆的表示向量？

请问数据集有网盘链接吗？
我看aicity官网还要输入秘钥才给下载

我看有人尝试在罗浩的baseline加过多卡和syncBN，但是掉点比较明显，不知是什么原因；看到前辈也加了 from apex.parallel import DistributedDataParallel as DDP 但是没有后续了

您好，我在尝试运行weakly_supervised_crop_aug.py 文件时始终存在以下错误：
/usr/local/lib/python3.6/dist-packages/torchvision/transforms/functional.py:114: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:143.)
img = torch.from_numpy(pic.transpose((2, 0, 1))).contiguous()
因此我始终无法生成cropped数据集，然后在test阶段一直报错无法找到AICity20_Cropped数据集。请问test可以只使用下载的AICity ReID 或者Simulation数据集吗？我在aicity20.py中更改了query和gallery路径，全部改成了下载的AICity ReID所在路径， 但程序仍然试图访问cropped数据集，麻烦作者大大告诉一下解决方案，谢谢。

您好，请问下您在训练Orientation ReID和Camera ReID分支的时候数据集aicity20-ReOri/ 和aicity20-ReCam/ 是如何设置的呢？谢谢！

GeForce GTX 1080
When I set FP16=True or false, the GPU memory doesn't change

@Xiangyu-CAS
I want to ask you something if you can help me.
What does compute_distmat_from_feats.py do?
When I run compute_distmat_from_feats.py I have got this one (is something wrong with this output):

python ./tools/aicity20/compute_distmat_from_feats.py --src_dir ./output/aicity20/0409-ensemble/ReCamID/
./tools/aicity20/compute_distmat_from_feats.py:19: UserWarning: This overload of addmm_ is deprecated:
addmm_(Number beta, Number alpha, Tensor mat1, Tensor mat2)
Consider using one of the following signatures instead:
addmm_(Tensor mat1, Tensor mat2, *, Number beta, Number alpha) (Triggered internally at /pytorch/torch/csrc/utils/python_arg_parser.cpp:882.)
distmat.addmm_(1, -2, feat, feat.t())

it works thank you!!

https://github.com/Xiangyu-CAS/AICity2020-VOC-ReID/blob/master/scripts/aicity20/test.sh
在生成vehicle distance matrix的时候，submit.py被调用3次，分别是“50”“101”“next”。
请问为什么需要这样做啊？谢谢！

I can't find the code of the generation of augment dataset,so I don't have ''AIC20_ReID_Cropped" directory.Can you help me?

Hi，很棒的工程！有个问题请教下，制作数据集AIC20_ReID_Cropped时，需要用到0326-search/augmix/best.pth，这个模型应该去哪里下载呢，一直没有找到哎，求帮忙

Hi,
I found that euclidean distance was used in your code to calculate distmat, while it was cosine similarity reported in your paper. As these two distance are opposite in expressing similarity, why is the calculation formula for the fusion distance matrix in the code still the same as paper's:
original_dist = original_dist - 0.1 * ori_dist - 0.1 * cam_dist
Because the ori_distance and cam_distance should be penalty, should "-" be changed to "+" ? But when I tried this so, the mAP on VeRi-776 decreased, which is worse than the original distance.
Looking forward to your reply, thanks!

I was wondering which version of cv2 and python3.7 are you using? I tried python3.7.7 and it gives some unexpected errors when installing other required packages like yacs. Thank you!

Most reid baseline set about dozens or even 100+ epoches. And do you use Adam with lr=0.00035? Have you tried SGD with lr=0.01? Thanks a lot!

and is there ResNeSt101_ibn_a ? THanks!

hi xiangyu：
to train reCam model，how to generate below dir ？

你好，我在训练veri数据集时，用的log日志里的参数，只复现到了80.7，比日志差0.9个点

• 请问81.6在训练时用到了弱监督增强了吗？
• 训练0.9个点的偏差正常吗？

Hello~项目已经成功跑起来了，识别效果很棒！还有个小问题想请教下您，
在运行的时候，数据集中有AIC20_REID,AIC20_REID_Cropped，和AIC20_ReID_Simulation，AIC20_REID中又有image_train,image_query,image_test,
我想问下，如果只提供AIC20_REID，然后其内部仅有image_test和image_query，没有image_train，可以识别出image_test中与image_query对应的汽车图片吗？因为我这样运行后，会报image_train中数据找不到的错误，不知道是否可以通过改动代码来实现这种需求，还是说算法本身并不支持这样做，必须要有image_train文件夹