leonhlj / rskp Goto Github PK

你好，感谢分享你的工作。由于实验室只有一台3090（不匹配cuda10）的服务器，所以我只能安装pytorch1.7.0，python3.6.6，cudatoolkit11.0.3。但是会在main_branch.py的30行mat_inv_x, _ = torch.solve(eye_x, eye_x - (w ** 2) * affinity_mat)遇到CUDA error: invalid configuration argument这一报错。不知道该如何解决？求指导一下。

Thank you for the excellent code. I am wondering if you could share the training config on ActivityNet dataset.

Classification map 96.154649
Detection map @ 0.100000 = 69.094885
Detection map @ 0.200000 = 63.238989
Detection map @ 0.300000 = 54.506203
Detection map @ 0.400000 = 45.565674
Detection map @ 0.500000 = 36.865112
Detection map @ 0.600000 = 24.011744
Detection map @ 0.700000 = 12.755108
average map = 43.719674

and I download your model offered, and only get the result about 44.7

Thans for your work.
I run the code and get the error:
RuntimeError: CUDA error: invalid configuration argument
It seems to be a gpu memory problem, how much memory you are using?

Thank you very much for your work.

def random_walk(x, y, w):
    x_norm = calculate_l1_norm(x)
    y_norm = calculate_l1_norm(y)
    eye_x = torch.eye(x.size(1)).float().to(x.device)
    latent_z = F.softmax(torch.einsum('nkd,ntd->nkt', [y_norm, x_norm]) * 5.0, 1)
    norm_latent_z = latent_z / (latent_z.sum(dim=-1, keepdim=True) + 1e-9)
    affinity_mat = torch.einsum('nkt,nkd->ntd', [latent_z, norm_latent_z])
    mat_inv_x, _ = torch.solve(eye_x, eye_x - (w ** 2) * affinity_mat)
    y2x_sum_x = w * torch.einsum('nkt,nkd->ntd', [latent_z, y]) + x
    refined_x = (1 - w) * torch.einsum('ntk,nkd->ntd', [mat_inv_x, y2x_sum_x])    
    return refined_x

I don't understand the logic of this part. Is there any paper about bipartite random walk (BiRW)? Or please elaborate on the logic of this part. Thank you very much.

Hello, can you provide a requirement.txt?

Thank you for your good work!
I'm just getting into this field, and I want to ask why some works use thumos14 dataset and some works use Thumos14reduced dataset. Are these two datasets exactly the same?

Hi thanks for your good work!

I have a question about the paper.

I can't understand how Z_t's dimension is t by n not l by n

Should t be considered the number of video snippets equal to l?

Thank you

Hi thanks for your good work!

I have a question about the source code.

The idxs value continues to return to zero only not the real video label index.

This is because the code refers to the 0th return value of np.where

I wonder if this is intended for now.

Thank You.

Thank you very much for providing the features. But this link can't be clicked now.

Can you save the features to Baidu online disk and share them, or send a copy to my email（ [email protected] ）。 Thank you very much and wish you a happy life.