namedben / convolutional-pose-machines-pytorch Goto Github PK

这里在每个loss之后都乘上了一个常数heat_weight，训练结果符合预期，但是如果不乘上常数，结果很差，这是为什么？能否解释下。

Why is loss multiplied with heat_weight which seems to be a constant?

Hi @namedBen ,

Thanks for sharing this great repo. As far as I know, CPM only supports single pose estimation in the image. The follow-up paper of CPM, OpenPose, does support that but it is hard to implement and need lots of trick. Do you have any idea that we can extend current CPM to support two pose estimation? Thanks!

Hi, I read your cpm_model.py and found you assign dim = 1 when doing torch.cat.
But the tensor in pytroch is in c,h,w. So shouldn't dim be 0?

cpm_checkpoint_pth.tar 百度网盘打不开 失效了 麻烦可以重新上传下，谢谢，或者可以发送到我的邮箱[email protected]，不胜感激！！！

我跑训练的时候，内存一下就超16G了，这正常嘛？谢谢

HI, i am new to deep learning and am having a hard time following some of the dimensions. would love some help.
Thank you for the code, very clear and concise

Edit: I was able to follow dimensions by creating separate layers and pushing a dummy input of ones through it.
I have new doubts tho,

what is centre map, and why do we need it?

Could you please provide your pre-trained model?
Thanks

because it has a lot of man made logistic code mistake. I don't know why coder do this shit action.

Hi, thanks for your great work!
I was wondering why you chose [128, 128, 128] and [256, 256, 256] as mean and std for your dataset. What if we use those from ImageNet (ranging from 0 to 1)?

Test Iteration: 100 Time 123.292s / 100iters, (1.233) Data load 121.240s / 100iters, (2.424793)
Loss = 2535.75683594 (ave = 2515.50237061)

Loss1 = 472.91094971 (ave = 469.05812988)
Loss2 = 419.15380859 (ave = 412.87503113)
Loss3 = 409.17819214 (ave = 407.73377777)
Loss4 = 411.38052368 (ave = 409.59360992)
Loss5 = 412.31079102 (ave = 408.39463074)
Loss6 = 410.82272339 (ave = 407.84722137)
您好，我想问一下我的loss值一直下不去您有什么方法吗？数据集我用的是lsp数据集

你好，请问你的测试图像是直接resize到368，368,还是centercrop至368，368啊？还是做了一些其他的处理，因为直接用你的测试程序去测试lsp数据集里的图片，程序会报错，只有图片尺寸为368*368才不会报错，然后我采用resize或cetercrop处理图像然后再测试，除了你提供的两张测试图像，其他的效果都很差。

我想下载已经训练好的权重，但是百度网盘上已经被取消了，能否分享一个新的训练好的权重

hi , thanks for share your model. I downlaoded it but met a problem like this:
tar -xvf cpm_latest.pth.tar
tar: This does not look like a tar archive
tar: Skipping to next header
tar: Exiting with failure status due to previous errors

Is it be compressed properly?