Efficient-AI-Backbones

Efficient-AI-Backbones copied to clipboard

Thank you very much for your released code for VIG. I spent some time on reading your code carefully. But I am confused about some parts of code, could you please give some hints? @iamhankai

1. reduce_ratios = [4, 2, 1, 1] I am confused about the reduced_ratios. I read from the paper that the sizes of each stages is 1/2 of their precedure stages. So what the reduce ratios are used for?
2. in the code:

class MRConv2d(nn.Module):
    """
    Max-Relative Graph Convolution (Paper: https://arxiv.org/abs/1904.03751) for dense data type
    """
    def __init__(self, in_channels, out_channels, act='relu', norm=None, bias=True):
        super(MRConv2d, self).__init__()
        self.nn = BasicConv([in_channels*2, out_channels], act, norm, bias)

    def forward(self, x, edge_index, y=None):
        x_i = batched_index_select(x, edge_index[1])
        if y is not None:
            x_j = batched_index_select(y, edge_index[0])
        else:
            x_j = batched_index_select(x, edge_index[0])
        x_j, _ = torch.max(x_j - x_i, -1, keepdim=True)
        b, c, n, _ = x.shape
        x = torch.cat([x.unsqueeze(2), x_j.unsqueeze(2)], dim=2).reshape(b, 2 * c, n, _)
        return self.nn(x) 

What the y stands for? I am really confused about the y as I found y is calculated from x based on the reduce_ratios. But I can not relate y or reduce_ratios to formulars in your VIG or Deep GCN. Could you please give me some advices? Thank you!

Thanks for the attention. Here we reduce the number of nodes by reduce_ratio so as to reduce the computational cost of distance calculation.

Thanks for the attention. Here we reduce the number of nodes by reduce_ratio so as to reduce the computational cost of distance calculation. Thank you for your rapid reply. I also found that your dilation is always set to 1 in your code: min(idx // 4 + 1, max_dilation) . Does this because you have reduced the nodes by pooling based on reduce_ration? Thank you very much for your precious time!

Dilation is not always 1 since idx is the layer index in min(idx // 4 + 1, max_dilation).

hi @iamhankai I am confused with this line x_j, _ = torch.max(x_j - x_i, -1, keepdim=True) in MRConv code,

x_j=[b, c, n, k] is the selected node by computing the distance, after this line x_j=[b, c, n, 1]. It seems that the MRConv only consider one node with the largest distance?

@PanXiebit The torch.max is applied at element-wise level, not the vector-wise.

@iamhankai Thanks, I understand!