Two-Stream Adaptive Graph Convolutional Networks for Skeleton-Based Action Recognition in CVPR19

LShi

Last update: Dec 26, 2022

Related tags

Deep Learning 2s-AGCN

Overview

2s-AGCN

Two-Stream Adaptive Graph Convolutional Networks for Skeleton-Based Action Recognition in CVPR19

Note

~~PyTorch version should be 0.3! For PyTorch0.4 or higher, the codes need to be modified.~~
Now we have updated the code to >=Pytorch0.4.
A new model named AAGCN is added, which can achieve better performance.

Data Preparation

Download the raw data from NTU-RGB+D and Skeleton-Kinetics. Then put them under the data directory:

 -data\  
   -kinetics_raw\  
     -kinetics_train\
       ...
     -kinetics_val\
       ...
     -kinetics_train_label.json
     -keintics_val_label.json
   -nturgbd_raw\  
     -nturgb+d_skeletons\
       ...
     -samples_with_missing_skeletons.txt

Preprocess the data with

python data_gen/ntu_gendata.py

python data_gen/kinetics-gendata.py.
Generate the bone data with:

python data_gen/gen_bone_data.py

Training & Testing

Change the config file depending on what you want.

`python main.py --config ./config/nturgbd-cross-view/train_joint.yaml`

`python main.py --config ./config/nturgbd-cross-view/train_bone.yaml`

To ensemble the results of joints and bones, run test firstly to generate the scores of the softmax layer.

`python main.py --config ./config/nturgbd-cross-view/test_joint.yaml`

`python main.py --config ./config/nturgbd-cross-view/test_bone.yaml`

Then combine the generated scores with:

`python ensemble.py` --datasets ntu/xview

Citation

Please cite the following paper if you use this repository in your reseach.

@inproceedings{2sagcn2019cvpr,  
      title     = {Two-Stream Adaptive Graph Convolutional Networks for Skeleton-Based Action Recognition},  
      author    = {Lei Shi and Yifan Zhang and Jian Cheng and Hanqing Lu},  
      booktitle = {CVPR},  
      year      = {2019},  
}

@article{shi_skeleton-based_2019,
    title = {Skeleton-{Based} {Action} {Recognition} with {Multi}-{Stream} {Adaptive} {Graph} {Convolutional} {Networks}},
    journal = {arXiv:1912.06971 [cs]},
    author = {Shi, Lei and Zhang, Yifan and Cheng, Jian and LU, Hanqing},
    month = dec,
    year = {2019},
}

Contact

For any questions, feel free to contact: [email protected]

Comments

Memory overloading issue

First of all, thanks a lot for making your code public. I am trying to do the experiment on NTU RGB D 120 dataset and I have split the data into training and testing in CS as given in the NTU-RGB D 120 paper. I have 63026 training samples and 54702 testing samples. I am trying to train the model on a GPU cluster but after running for one epoch, my model exceeds the memory limit: I try to clear the cache explicitly using gc.collect but the model still continues to grow in size. It will be great if you can help regarding this.

opened by Anirudh257 46
I got some wrong when I was training the net

首先我是得到了下面这个error，

注释掉该参数后，got another error

I got this error ,but I don't know how to solve. Could you give me some advice?

Traceback (most recent call last): File "/home/sues/Desktop/2s-AGCN-master/main.py", line 550, in processor.start() File "/home/sues/Desktop/2s-AGCN-master/main.py", line 491, in start self.train(epoch, save_model=save_model) File "/home/sues/Desktop/2s-AGCN-master/main.py", line 372, in train loss.backward() File "/home/sues/anaconda3/envs/2sAGCN/lib/python3.5/site-pac[kages/torch/autograd/variable.py", line 167, in backward torch.autograd.backward(self, gradient, retain_graph, create_graph, retain_variables) File "/home/sues/anaconda3/envs/2sAGCN/lib/python3.5/site-packages/torch/autograd/init.py", line 99, in backward variables, grad_variables, retain_graph) RuntimeError: cuda runtime error (59) : device-side assert triggered at /pytorch/torch/lib/THC/generic/THCTensorMath.cu:26 /pytorch/torch/lib/THCUNN/ClassNLLCriterion.cu:101: void cunn_ClassNLLCriterion_updateOutput_kernel(Dtype *, Dtype *, Dtype *, long *, Dtype *, int, int, int, int, long) [with Dtype = float, Acctype = float]: block: [0,0,0], thread: [0,0,0] Assertion `t >= 0 && t < n_classes

opened by Dongjiuqing 10
内存分配不够，Unable to allocate 29.0 GiB for an array with shape (7790126400,) and data type float32

当运行python data_gen/gen_bone_data.py这据代码时，会在 File "data_gen/gen_bone_data.py", line 62, in data = np.load('./data/{}/{}_data.npy'.format(dataset, set)) 处遇到 MemoryError: Unable to allocate 29.0 GiB for an array with shape (7790126400,) and data type float32 这样的错误，请问该如何解决呢？

opened by XieLinMofromsomewhere 7
augmentation in feeder

Hi, I want to know the data augmentation in the feeder has not improved? Does the length of the input have a big influence? Also, have you trained the model on the 120 dataset? How's the accuracy?

opened by VSunN 7
problem with gen_bone_data.py

你好，请问一下我跑gen_bone_data.py时报错，好像是矩阵的维度有问题，该怎么解决，谢谢 lin@lin:~/2s-AGCN-master/data_gen$ python gen_bone_data.py ntu/xsub train 4%|█▋ | 1/25 [06:49<2:43:40, 409.20s/it]Traceback (most recent call last): File "gen_bone_data.py", line 50, in fp_sp[:, :, :, v1, :] = data[:, :, :, v1, :] - data[:, :, :, v2, :] IndexError: index 20 is out of bounds for axis 3 with size 18 4%|█▋ | 1/25 [06:49<2:43:45, 409.41s/it]

opened by JaxferZ 5
Accuracy of aagcn

I ran your implemented code using J-AAGCN and NTU-RGBD CV dataset. But Accuracy is 94.64, not 95.1 in your paper. What is the difference? The batch size was 32, not 64 because of the resource limit. Are there any other things to be aware of? I use your implemented code.

opened by ilikeokoge 4
用released model做test的时候提示 Unexpected key(s) in state_dict:

python main.py --config ./config/nturgbd-cross-view/test_joint.yaml 这段代码能得到论文的结果。但是到了这段 python main.py --config ./config/nturgbd-cross-view/test_bone.yaml``，会提示RuntimeError: Error(s) in loading state_dict for Model:`

Unexpected key(s) in state_dict: "l1.gcn1.conv_res.0.weight", "l1.gcn1.conv_res.0.bias", "l1.gcn1.conv_res.1.weigh t", "l1.gcn1.conv_res.1.bias", "l1.gcn1.conv_res.1.running_mean", "l1.gcn1.conv_res.1.running_var", "l5.gcn1.conv_res.0.we ight", "l5.gcn1.conv_res.0.bias", "l5.gcn1.conv_res.1.weight", "l5.gcn1.conv_res.1.bias", "l5.gcn1.conv_res.1.running_mean ", "l5.gcn1.conv_res.1.running_var", "l8.gcn1.conv_res.0.weight", "l8.gcn1.conv_res.0.bias", "l8.gcn1.conv_res.1.weight", "l8.gcn1.conv_res.1.bias", "l8.gcn1.conv_res.1.running_mean", "l8.gcn1.conv_res.1.running_var".

看起来是这个pretrained模型与提供的代码不匹配，我怎么做才能得到结果呢！期待回复！

opened by tailin1009 3
dataload error

thank your source code, but when I run this code, The following error occurs: ValueError: num_samples should be a positive integer value, but got num_samples=0

I've run the program 'python data_gen/ntu_gendata.py 'before, and some documents were generated : train_data_joint.npy train_label.pkl val_data_joint.npy val_label.pkl

but their size are all 1K

How should I deal with, trouble you give directions.

thanks

opened by xuanshibin 3
RuntimeError: running_mean should contain 126 elements not 63 (example).

What is your elements for number of joints (18)? When I run your code, I got this error " RuntimeError: running_mean should contain 126 elements not 63". 63 means I change number of node. How to adjust these elements and how to get your elements 126 for your experiment?

opened by JasOlean 3
what is (N, C, T, V, M) in agcn.py?
thank you for sharing code and information :) I have some question about agcn.py code

what is (N, C, T, V, M) in agcn.py? i guess T is 300 frame, V is the similarity between nodes, M is number of men in one video, but i am not sure that it is right

are bone train code and joint train(agcn.py) code same? if it is not, is bone train code aagcn.py?
opened by lodado 2
No module named 'data_gen' and No such file or directory: '../data/kinetics_raw/kinetics_val'

When I run "python data_gen/ntu_gendata.py", gets the error : ModuleNotFoundError: No module named 'data_gen'.

When I run "python data_gen/kinetics_gendata.py", gets the error : FileNotFoundError: [Errno 2] No such file or directory: '../data/kinetics_raw/kinetics_val'.

My raw data has put in the ./data.

Needs your help！

opened by XiongXintyw 2
关于MS-AAGCN的运行问题

大佬您好！我十分有幸拜读了您的文章《Skeleton-Based Action Recognition with Multi-Stream Adaptive Graph Convolutional Networks》，受益匪浅！我已经跑通了2S-AGCN的代码，想和您请教一下MS-AAGCN的代码该如何运行呢？

opened by 15762260991 1
注意力模块中参数A的定义

在复现代码时找不到关于图卷积层中参数A的定义请问这个A指的是什么呢： class TCN_GCN_unit(nn.Module): def init(self, in_channels, out_channels, A, stride=1, residual=True, adaptive=True, attention=True):

opened by wangxx0101 1
关于自适应时，tanh和softmax函数的问题

大佬您好，有两个问题想请教一下。 ①tanh激活函数，它将返回一个范围在[- 1,1]的值，softmax激活函数返回一个[0, 1]的值，当我们建模关节之间的相关性时，如果使用tanh返回为负值的时候，是说明这两个关节负相关吗？ ②为什么tanh的效果会比softmax好一点，这个我不是太懂，您可以详细的讲解一下吗？

opened by blue-q 0
Where is the code for visualization in Figure 8 and 9?

Dear Authors,

I have already read your "Two-Stream Adaptive Graph Convolutional Networks for Skeleton-Based Action Recognition". In that paper, you showed some experimental results in Figure 8 and 9. I would like to know which part of the code for that. Or, how to use layers to show these result's visualization? If you answer my question, I will really appreciate you. Thank you.

opened by JasOlean 3

Releases(v0.0)

v0.0(Jul 31, 2019)

Pretrained models for joints and bones on cv and cs benchmarks of NTU-RGBD dataset.
Source code(tar.gz)
Source code(zip)
model.zip(49.39 MB)

Two-Stream Adaptive Graph Convolutional Networks for Skeleton-Based Action Recognition in CVPR19

Related tags

Overview

2s-AGCN

Note

Data Preparation

Training & Testing

Citation

Contact

Comments

Releases(v0.0)

v0.0(Jul 31, 2019)

Owner

LShi

[ICCV2021] Official code for "Channel-wise Topology Refinement Graph Convolution for Skeleton-Based Action Recognition"

Paddle-Skeleton-Based-Action-Recognition - DecoupleGCN-DropGraph, ASGCN, AGCN, STGCN

A pytorch reproduction of { Co-occurrence Feature Learning from Skeleton Data for Action Recognition and Detection with Hierarchical Aggregation }.

Official PyTorch implementation of "Contrastive Learning from Extremely Augmented Skeleton Sequences for Self-supervised Action Recognition" in AAAI2022.

The official TensorFlow implementation of the paper Action Transformer: A Self-Attention Model for Short-Time Pose-Based Human Action Recognition

Code for Two-stage Identifier: "Locate and Label: A Two-stage Identifier for Nested Named Entity Recognition"

Official implementation of ACTION-Net: Multipath Excitation for Action Recognition (CVPR'21).

This is the repository for the AAAI 21 paper [Contrastive and Generative Graph Convolutional Networks for Graph-based Semi-Supervised Learning].

Efficient Two-Step Networks for Temporal Action Segmentation (Neurocomputing 2021)

This repository implements and evaluates convolutional networks on the Möbius strip as toy model instantiations of Coordinate Independent Convolutional Networks.

PyTorch implementation for Convolutional Networks with Adaptive Inference Graphs

Allows including an action inside another action (by preprocessing the Yaml file). This is how composite actions should have worked.

Official Pytorch Implementation of 'Learning Action Completeness from Points for Weakly-supervised Temporal Action Localization' (ICCV-21 Oral)

Human Action Controller - A human action controller running on different platforms.

TDN: Temporal Difference Networks for Efficient Action Recognition

AEI: Actors-Environment Interaction with Adaptive Attention for Temporal Action Proposals Generation

PyTorch implementation of the R2Plus1D convolution based ResNet architecture described in the paper "A Closer Look at Spatiotemporal Convolutions for Action Recognition"

Web service for facial landmark detection, head pose estimation, facial action unit recognition, and eye-gaze estimation based on OpenFace 2.0