siyu-c / acar-net Goto Github PK

hello!
I want to konw the mean of 'time','midframe' in ava_train_v2.2_with_fair_0.9.pkl,so can you explain for me ?

Thank you again for sharing this code. I like it more than Slowfast.

Can you please share '.pkl' scripts for a given '.csv' from fair and gt '.csv' given https://github.com/Siyu-C/ACAR-Net/tree/master/annotations.

Many thanks
G.

Hello
Thanks for your great work!!!
I have a question about actor localization: wheather your method support actor localization in spatial or not ？

Hello
@Siyu-C @junting
Thanks for your great work!
I have a question when I run this code:
When I set batch size=8 and gpu =2 ,the mAP is very low,but batch size=1,gpu=1,the mAP is very high.So Can you tell me why?I do not konw how to explain this problem.

Hi~ thanks for your great job!
I noticed that you reported results with SlowFast R-101+NL as the backbone which was pretrained on K600. Is it the pretrained model from the SlowFast repo? If it is the same in SlowFast repo, I wonder how to adapt it into this repo.
Thanks a lot~

Hello, thank you very much for your work. How can I train my own data set with your method? I look forward to receiving your reply.

Hello
I want to konw when will all code and models be available?I want to learn about train.pkl.
Thanks!!!!

您好，请问能否提供下基于 ucf101-24数据集的 ACAR-Net的代码呢？

Hello, thank you very much for your work. I noticed that in the backbone network SlowFast, there is no downsampling in the last ResNet stage, which is inconsistent with the official SlowFast source code. May I ask what was the consideration behind this decision? Additionally, in this scenario, when using the official pre-trained models of SlowFast, will it cause undesirable effects?

Thank you very much for your work！
Can you provide test cases ？to run inference with model(s) on wild video(s).
Thank you very much！

I want to evaluate testdataset by used my trained model,by it seemed run very slowly.Can you give me an example about evaluating ?Thanks!!!
@Siyu-C @junting

Is there any inference scripts on video files with out annotations?

Great job @junting @Siyu-C. I wanna try a deeper backbone. Could you tell when will you release pretrained slowfast_resnet101 model? Thanks

input a video, output spatio-temporal behavior detection video.
Thank you very much

Hello，thanks for your good job!But i want to konw the ACAR wheather support test mode?except train and validation.Thanks!!!
@junting @Siyu-C

hi, thanks for your awewsome work. Can you tell me about model test speed? Count on input video resolution, video duration , fps and so on.

There is a quesiton about the loss.py
line7: pose_logits = nn.Softmax(dim=1)(logits[:, :13])
line8: interact_logits = nn.Sigmoid()(logits[:, 13:]))
The number of person pose is 14, Should the "13" in line7 and line8 to be changed to "14"?

Does the code include the implementation of ACFB?

Thanks for your contributions, I wanna make custom dataset(AVA) and using your net to fine tuning, but I don't know how to create AVA dataset. Could you share your points with me? Thanks a lot

According to the paper, you use CE loss for pose label instead of bce loss. But in Ava dataset there exist some boxes without pose annotation (according to my calculation, about 1500 boxes) and also some boxes with more than one pose annotation. How do you deal with that?

Plan to release training code and models?

Could you also release your annotation files for v2.1? Thanks a lot.

Nice work and thank you for sharing your code!
Could you please tell me about your computing resources on this task? E.g., how many and what type of GPU did you use and how long is the training time? Thank you!

[2022-05-05 23:18:34,655][        main.py][line: 280][    INFO] Epoch [1]       Iter [53880/184378]     Time 0.238 (0.412)   Data 0.000 (0.192)      Loss 0.0788 (0.0725)                                                 
[2022-05-05 23:18:39,059][        main.py][line: 280][    INFO] Epoch [1]       Iter [53900/184378]     Time 0.296 (0.220)   Data 0.000 (0.000)      Loss 0.0805 (0.0908)                                                 
/opt/conda/conda-bld/pytorch_1646756402876/work/aten/src/ATen/native/cuda/Loss.cu:115: operator(): block: [0,
0,0], thread: [96,0,0] Assertion `input_val >= zero && input_val <= one` failed.                             /opt/conda/conda-bld/pytorch_1646756402876/work/aten/src/ATen/native/cuda/Loss.cu:115: operator(): block: [0,
0,0], thread: [97,0,0] Assertion `input_val >= zero && input_val <= one` failed.  

您好，我按照您给的配置SLOWFAST_R101_ACAR_HR2O_DEPTH1.yaml训练，nproc_per_node=1其他为默认，数据集也是按您提供的工具分割出的图片，显卡为3080ti，报错代码如上。我debug了过程，发现

        ret = model(data)
        num_rois = ret['num_rois']
        outputs = ret['outputs']
        targets = ret['targets']

这个outputs出来的数据全是[nan,nan,nan,...]

使用SLOWFAST_R50_ACAR_HR2O.yaml这个配置好像可以正常运行，我不知道问题出在哪里，期待得到您的回复，谢谢！

Whats would be the values base_lr and warmp_lr for 8Gpu training with one batch sample per GPU, effective batch size of 8.

base_lr is minimum LR and warmp_lr is LR after completing warm-up iterations, in this after the first epoch LR would be warmup_lr, is that right?
if base_lr=0.008 and warmp_lr is 0.064
Training start
epoch 0 LR = 0.008
epoch 1 LR = 0.064
epoch 2 LR = 0.064

is that correct?

Thanks for your great job. In paper, you mentioned you set batch size as 32. However, you set batch size as 1 in config file, and set 8 gpus as default. The total batch size are 8 which is different from paper. Could you share the settings of experiments in your paper? Thank you

@Siyu-C @junting ,hello, i get the mAP 0.00,when i run the code on dataset Something-else.Can you tell me why the mAP is 0.00（but the loss is not 0.0000）.Thanks!!!

Hi, this work is pretty cool !
I come across several problems when I use the repo.

1. What's the difference between ava_train_v2.2.pkl & ava_train_v2.2_with_fair_0.9.pkl ? What's the meaning of "with_fair_0.9" or "with_fair_0.85"? I print some data of these two pkl file mentioned above, and found that the one with "with_fair_0.9" got more bbox but the corresponding person_idx value is -1.

2. How can I generate the pkl file from original ava csv format?

3. How can I test arbitrary video input? It seems it can only support the ava val dataset which already got the bounding box value.

Thanks!

@junting @Siyu-C
hello，thanks for your great works!
By the way,where is the SLOWFAST_R101_K600.pth.tar from PySlowfast? I can not find it.
I have found this:

but i can not load them!

When I run this code ,it occurs : RuntimeError: NCCL error in: /pytorch/torch/lib/c10d/ProcessGroupNCCL.cpp:32, unhandled cuda error, NCCL version 2.4.8

I take some solutions,but they seem to not work.

my script is :CUDA_VISIBLE_DEVICES='0,1' python main.py --config /home/wsm/ACAR-Net/configs/AVA/SLOWFAST_R50_baseline.yaml --nproc_per_node 8 --backend nccl --master_port 31114

How to solve it?
@junting

@junting @Siyu-C
hello
I want to use other backbones like s3d, and r2plus1d, can you give me some advice?

Hi,

Are you updating ACFB features during training? From the paper I can see that the feature bank (ACFB) is initially filled with features from a trained ACAR-Net and that these features are used as key-values in the attention module. However, I am not sure if these memory features are further updated.

Thanks!

Which is the map that this code achieves? I was expecting to get 27.83 as the paper says. However, I am only able to get a 27.1. I am using batch 4 with 8 GPUS (32 total) and 0.064 as warmup_lr. Am I missing something?

您好，请问能否提供下基于 ucf101-24数据集的 ACAR-Net的代码呢？

hello
@junting @Siyu-C
I am sorry to bother you.
I have a problem with calc_mAP.py.
In line 39, the parameter: capacity=0(capacity: Maximum number of labeled boxes allowed for each example. Default is 0 where there is no limit.) whether has a special role or it will not affect the final mAP?
Thanks!!!!

请问 clip和clip之间的bbox是怎么连接起来的？我看代码是一开始就roialign了，就是bbox是所有clip的feat亚，好像不是按照每个clip几个bbox的模式？

Hi authors,
Will we release it?

Thanks for your ACAR-net paper.
I am looking forward to deeper learning this work.

Can you tell me how to generate pkl files of other data set labels? I like your work very much and look forward to receiving your reply.