Point Transformer v1的参数量怎么计算

论文地址：https://arxiv.org/pdf/2012.09164.pdf
代码地址：https://github.com/Pointcept/Pointcept/blob/main/pointcept/models/point_transformer/point_transformer_seg.py

• 层数：5层encoder，5层decoder
• Channel的数目：[32, 64, 128, 256, 512]
• Block的数目：
  PointTransformerSeg26 -> [1, 1, 1, 1, 1]
  PointTransformerSeg38 -> [1, 2, 2, 2, 2]
  PointTransformerSeg50 -> [1, 2, 3, 5, 2]

第一层的定义：

 self.enc1 = self._make_enc(
            Bottleneck,
            32,
            1,
            8,
            stride=1,
            nsample=8,
        )  # N/1

设计主函数输出每一层encoder的参数量，不包含transition layer而且只有1个block：

if __name__ == '__main__':

    in_planes = [32, 64, 128, 256, 512]
    n_sample = [8, 16, 16, 16, 16]

    for i in range(5):

        model = Bottleneck(in_planes=in_planes[i], planes=in_planes[i], share_planes=8, nsample=n_sample[i])
        pytorch_total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)

        print(str(i) + ": " + str(pytorch_total_params/1024.0/1024.0) + "M")

结果：

0: 0.0055103302001953125M
1: 0.021043777465820312M
2: 0.08223152160644531M
3: 0.3250904083251953M
4: 1.2927417755126953M

可以看出point transformer的参数量比较小。