torch.index_select(input_tensor, 切片维度, 切片索引)
注意:切完之后,转onnx时会生成Gather节点;
torch自带切片操作:
start : end : step:
范围前闭后开,将其放在哪个维度上,就对那个维度起作用
torch.cat((a, b) , dim)
在已有的轴上拼接 矩阵,默认轴为0,给定轴的维度可以不同,其余轴的维度必须相同
三个操作的组合使用例子如下:
python
import torch
x = torch.randn(1, 18, 4, 4)
# print("x:",x)
print("x.shape:",x.shape)
indices_cls = torch.tensor([2, 5, 8, 11, 14, 17])
indices_point = torch.tensor([0,1, 3,4, 6,7, 9,10, 12,13, 15,16])
kpt_point = torch.index_select(x, 1, indices_point)
kpt_cls = torch.index_select(x, 1, indices_cls)
print("kpt_point.shape:",kpt_point.shape)
print("kpt_cls.shape:",kpt_cls.shape)
x_2 = torch.cat([kpt_point[:,0:2:1,],kpt_cls[:,0:1:1,],kpt_point[:,2:4:1,],kpt_cls[:,1:2:1,],kpt_point[:,4:6:1,],kpt_cls[:,2:3:1,],
kpt_point[:,6:8:1,],kpt_cls[:,3:4:1,],kpt_point[:,8:10:1,],kpt_cls[:,4:5:1,],kpt_point[:,10:12:1,],kpt_cls[:,5:6:1,]],1)
# print("x_2:",x_2)
print("x_2.shape:",x_2.shape)
打印组合前后tensor的输出形状和内容发现,前后一致:
python
x.shape: torch.Size([1, 18, 4, 4])
kpt_point.shape: torch.Size([1, 12, 4, 4])
kpt_cls.shape: torch.Size([1, 6, 4, 4])
x_2.shape: torch.Size([1, 18, 4, 4])