tensor view 使用与常见报错: shape ‘[4]‘ is invalid for input of size 6 +not compatible x.contiguous().view

原创已于 2022-04-29 10:00:22 修改 · 9.3k 阅读

1 ·

CC 4.0 BY-SA版权

文章标签：

#pytorch #深度学习 #机器学习

于 2021-08-08 22:55:08 首次发布

其他专栏收录该内容

174 篇文章

订阅专栏

本文介绍了PyTorch中张量的view和permute方法。view用于调整张量形状，须确保元素总数不变，否则会引发错误。permute则重新排列张量的维度顺序。当张量不是contiguous时，使用view前需先转为contiguous。示例展示了这两个方法的用法及错误处理。

部署运行你感兴趣的模型镜像

查看大小

Tensor.shape 
Tensor.size()

AttributeError: ‘list’ object has no attribute ‘shape’

len(list)
或
np.array(a).shape 
或
print(list[0].shape)

view

if __name__ == '__main__':
    x = torch.tensor([[1, 2,3,4],
                      [5, 6,7,8],
                      [9, 10,11,12],
                      [13, 14,15,16],]).unsqueeze(dim=0).unsqueeze(dim=0)
	x = x.view(x.shape[0], x.shape[1]*4, int(x.shape[2]/2), int(x.shape[3]/2))

tensor([[[[ 1,  2,  3,  4],
          [ 5,  6,  7,  8],
          [ 9, 10, 11, 12],
          [13, 14, 15, 16]]]])
tensor([[[[ 1,  2],
          [ 3,  4]],

         [[ 5,  6],
          [ 7,  8]],

         [[ 9, 10],
          [11, 12]],

         [[13, 14],
          [15, 16]]]])

报错原因

$\\ \color{red}但必须保证调整前后元素总数一致$
$v i e w 不会修改自身的数据，返回的新 t e n s o r 与原 t e n s o r 共享内存，即更改一个，另一个也随之改变。$

X.view(X.shape[2:])# 在y的前两个维度不是1的情况下，是会报错的

其他错误

RuntimeError: view size is not compatible with input tensor’s size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(…) instead.

tensor不是contiguous 引起的错误。查看targets.is_contiguous()为False。

两种解决办法：1）按照提示使用reshape代替；2）将变量先转为contiguous ，再进行view:

x.contiguous().view(b,-1)

view VS permute

在这里插入图片描述

import torch
x = torch.randn(2, 3)
y = torch.permute(x, (1, 0))
print(x)
print(y)

tensor([[ 1.0945, -0.0853, -0.3914],
        [-1.1281, -0.0633, -1.0864]])
tensor([[ 1.0945, -1.1281],
        [-0.0853, -0.0633],
        [-0.3914, -1.0864]])

torch.Size([3, 5, 2])
tensor([[[ 0.2059, -0.7165],
         [-1.1305,  0.5886],
         [-0.1247, -0.4969],
         [-0.5788,  0.0159],
         [ 1.4304,  0.6014]],

        [[ 2.4882, -0.3910],
         [-0.5558,  0.6903],
         [-0.4219, -0.5498],
         [-0.5346, -0.0703],
         [ 1.1497, -0.3252]],

        [[-0.5075,  0.5752],
         [ 1.3738, -0.3321],
         [-0.3317, -0.9209],
         [-1.6677, -1.1471],
         [-0.9269, -0.6493]]])
torch.Size([2, 3, 5])
tensor([[[ 0.2059, -1.1305, -0.1247, -0.5788,  1.4304],
         [ 2.4882, -0.5558, -0.4219, -0.5346,  1.1497],
         [-0.5075,  1.3738, -0.3317, -1.6677, -0.9269]],

        [[-0.7165,  0.5886, -0.4969,  0.0159,  0.6014],
         [-0.3910,  0.6903, -0.5498, -0.0703, -0.3252],
         [ 0.5752, -0.3321, -0.9209, -1.1471, -0.6493]]])

[官方链接](https://pytorch.org/docs/stable/generated/torch.permute.html)
[添加链接描述](https://www.geeksforgeeks.org/python-pytorch-permute-method/#:~:text=PyTorch%20torch.permute%20%28%29%20rearranges%20the%20original%20tensor%20according,as%20that%20of%20the%20original.%20Syntax:%20torch.permute%20%28*dims%29)
更多：
[python：pytorch维度变换，爱因斯坦求和约定enisum，einops](https://blog.youkuaiyun.com/ResumeProject/article/details/121354010)