zhenhuat / stcformer Goto Github PK
View Code? Open in Web Editor NEW(CVPR2023)3D Human Pose Estimation with Spatio-Temporal Criss-cross Attention
(CVPR2023)3D Human Pose Estimation with Spatio-Temporal Criss-cross Attention
你好,十分感谢您的工作。请问可以看一下dataset文件夹的整体结构并且包含什么样的数据文件嘛。我运行时出现数据的问题:FileNotFoundError: [Errno 2] No such file or directory: '/tangzhenhua2/video_pose/MHFormer-main-old/dataset/data_3d_h36m.npz。期待您的回复,谢谢。
Hello, we have successfully replicated your work, and the results are excellent. However, we need to point out an error in your paper. In the table of your paper, you mistakenly reported the result of MACs as the result of FLOPs. In reality, FLOPs equals 2 times MACs. We reproduced the results for 27 frames using THOP, and the obtained value is 2178M, which represents MACs, not FLOPs. The correct FLOPs result should be 4356M.
您好,在复现您代码的过程中发现多GPU训练时总是报错,请问您这个实验设置直接设置多卡是不支持的吗?单卡的话可以正常运行,但是直接设置多卡就会报错:
TypeError: Caught TypeError in replica 1 on device 1.
不知道应该怎么解决呢?
作者您好,我是刚入门这个方向的新人,我想请问一下如果要在这个方向取得一定成果,1张4090显卡可以满足算力要求吗?
您好,请问run_3dhp_stc.py和run_stc.py中引用两个模型?为啥不都使用stcformer.py呢?看起来差别还蛮大的
你好,按照你提供的训练脚本:python run_stc.py -f 27 -b 128 --train 1 --layers 6 -s 9,训练出的模型有些问题,训练结果为:
2023/08/10 12:51:16 epoch: 2, lr: 0.0009600, loss: 0.0369, p1: 48.94, p2: 36.62
2023/08/10 13:22:25 epoch: 3, lr: 0.0009216, loss: 0.0316, p1: 46.50, p2: 35.82
2023/08/10 13:53:36 epoch: 4, lr: 0.0008847, loss: 0.0286, p1: 46.45, p2: 35.38
2023/08/10 14:24:49 epoch: 5, lr: 0.0008493, loss: 0.0266, p1: 47.20, p2: 35.70
2023/08/10 14:55:51 epoch: 6, lr: 0.0008154, loss: 0.0251, p1: 46.23, p2: 34.75
2023/08/10 15:26:56 epoch: 7, lr: 0.0007828, loss: 0.0240, p1: 45.34, p2: 34.90
2023/08/10 15:57:58 epoch: 8, lr: 0.0007514, loss: 0.0230, p1: 45.24, p2: 34.78
2023/08/10 16:42:51 epoch: 9, lr: 0.0007214, loss: 0.0223, p1: 45.56, p2: 34.96
2023/08/10 17:14:10 epoch: 10, lr: 0.0006925, loss: 0.0216, p1: 46.46, p2: 34.99
2023/08/10 17:45:28 epoch: 11, lr: 0.0006648, loss: 0.0210, p1: 45.84, p2: 35.18
2023/08/10 18:16:47 epoch: 12, lr: 0.0006382, loss: 0.0205, p1: 46.81, p2: 35.02
2023/08/10 18:47:58 epoch: 13, lr: 0.0006127, loss: 0.0201, p1: 46.30, p2: 34.85
python run_stc.py -f 81-b 128 --train 1 --layers 6 -s 9
2023/08/10 21:06:24 epoch: 1, lr: 0.0010000, loss: 0.0788, p1: 50.18, p2: 39.26
2023/08/10 22:22:16 epoch: 2, lr: 0.0009600, loss: 0.0393, p1: 46.84, p2: 36.61
2023/08/10 23:38:09 epoch: 3, lr: 0.0009216, loss: 0.0331, p1: 45.97, p2: 35.37
2023/08/11 00:54:04 epoch: 4, lr: 0.0008847, loss: 0.0298, p1: 45.23, p2: 35.14
2023/08/11 02:09:55 epoch: 5, lr: 0.0008493, loss: 0.0275, p1: 44.11, p2: 34.80
2023/08/11 03:25:47 epoch: 6, lr: 0.0008154, loss: 0.0259, p1: 44.48, p2: 34.46
2023/08/11 04:41:38 epoch: 7, lr: 0.0007828, loss: 0.0246, p1: 43.90, p2: 34.36
2023/08/11 05:57:30 epoch: 8, lr: 0.0007514, loss: 0.0236, p1: 45.46, p2: 34.39
2023/08/11 07:13:20 epoch: 9, lr: 0.0007214, loss: 0.0228, p1: 45.18, p2: 34.24
和你代码中给的预训练模型差异比较大,请指教?
Hi thank you for your research.
I want to test STCFormer model in the wild video.
How can test it? Is there any way to visualization skeleton in the wild video?
请教一下为什么为根据指示训练27帧的模型,没有在第6轮训出44.06的p1指标?
您好,请问一下,如果希望能复现论文中在Human3.6M上243帧测得的结果(STCFormer和STCFormer-L),对于243帧训练的参数完整设置是怎么样的呢?如果方便地话,希望您能提供训练的脚本,包括batchsize以及stride等参数设置,谢谢!
很棒的一个工作,可以分享一下3DHP的训练参数设置吗?感谢!
Could I ask if the general training is to run the set 60 epochs? Or is it ok to stop at the front.
你好作者,很棒的工作。请问一下在训练3DHP时候,logging.info('epoch: %d, lr: %.7f, loss: %.4f, MPJPE: %.2f, p1: %.2f' % (epoch, lr, loss, mpjpe, p1))中的p1是mpjpe还是pa-mpjpe?另外p1对应的是step里返回的 error_sum_test.avg?
大佬您好,请问图5的注意力可视化热图是将q和x相乘之后生成的吗?希望您能提供对应代码,谢谢!
请问作者您之前训练的81帧42.7和27帧的44.1是如何求加权平均的?
Dear Author,
Can you please share code and pretrained models for Human 3.6m and MPI-INF-3DHP?
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.