快速入门
Tensor
Tensor是PyTorch中重要的数据结构，可以当成是一个高维数组。本身和numpy的ndarrays很像，但是两者的区别在于，前者是可以使用GPU加速的。除此以外两者的基本操作是没有太大区别的。 现在在ipyhton中演示一下基本的操作
In [1]: import torch
In [2]: import torch as t
In [3]: x = t.Tensor(5,5) # 为初始化的5*5的矩阵
In [4]: x
Out[4]:
tensor([[ 1.6907e-01, 1.7743e+28, 1.1673e-32, 6.7685e+22, 2.9185e+32],
[ 5.3723e+19, 4.4721e+21, 4.8365e+30, 3.0135e+29, 6.6135e+19],
[ 4.3845e+31, 1.8886e+31, 7.1849e+22, 1.0086e-32, 1.7443e+28],
[ 1.8970e+31, 1.2424e+22, 2.5072e-12, 4.5144e+27, 9.1041e-12],
[ 6.2609e+22, 4.7428e+30, 1.5555e+28, 2.7259e+20, 8.1330e+32]])
In [5]: x = t.rand(5,5) #使用[0,1]均匀分布 t.randn()生成正态分布的数组
In [6]: x
Out[6]:
tensor([[ 0.5446, 0.5416, 0.0336, 0.6157, 0.0828],
[ 0.9987, 0.2257, 0.6390, 0.4904, 0.9439],
[ 0.7656, 0.6790, 0.3703, 0.4811, 0.8022],
[ 0.0563, 0.5616, 0.7834, 0.6812, 0.5832],
[ 0.7723, 0.1685, 0.7272, 0.0102, 0.3450]])
有关于加法有几种形式
In [7]: y = t.rand(5,5)
In [8]: x+y #直接用+进行相加
Out[8]:
tensor([[ 0.6177, 0.6424, 0.3739, 0.6916, 0.2209],
[ 1.3712, 0.5577, 1.3626, 1.0993, 1.5942],
[ 1.5684, 0.7533, 1.0628, 1.1381, 1.2293],
[ 0.1066, 1.5505, 1.0467, 1.5348, 1.4244],
[ 1.0477, 0.2311, 1.0946, 0.6907, 0.8061]])
In [9]: t.add(x,y) #需要用变量接收 可以直接以参数的形式 t.add(x, y, out=res)
Out[9]:
tensor([[ 0.6177, 0.6424, 0.3739, 0.6916, 0.2209],
[ 1.3712, 0.5577, 1.3626, 1.0993, 1.5942],
[ 1.5684, 0.7533, 1.0628, 1.1381, 1.2293],
[ 0.1066, 1.5505, 1.0467, 1.5348, 1.4244],
[ 1.0477, 0.2311, 1.0946, 0.6907, 0.8061]])
In [10]: y.add(x) #不修改y本身的值
Out[10]:
tensor([[ 0.6177, 0.6424, 0.3739, 0.6916, 0.2209],
[ 1.3712, 0.5577, 1.3626, 1.0993, 1.5942],
[ 1.5684, 0.7533, 1.0628, 1.1381, 1.2293],
[ 0.1066, 1.5505, 1.0467, 1.5348, 1.4244],
[ 1.0477, 0.2311, 1.0946, 0.6907, 0.8061]])
In [11]: y
Out[11]:
tensor([[ 0.0731, 0.1009, 0.3402, 0.0759, 0.1381],
[ 0.3724, 0.3321, 0.7236, 0.6089, 0.6503],
[ 0.8028, 0.0743, 0.6925, 0.6570, 0.4271],
[ 0.0503, 0.9889, 0.2633, 0.8536, 0.8411],
[ 0.2754, 0.0626, 0.3674, 0.6805, 0.4611]])
In [12]: y.add_(x) #会修改y的本身
Out[12]:
tensor([[ 0.6177, 0.6424, 0.3739, 0.6916, 0.2209],
[ 1.3712, 0.5577, 1.3626, 1.0993, 1.5942],
[ 1.5684, 0.7533, 1.0628, 1.1381, 1.2293],
[ 0.1066, 1.5505, 1.0467, 1.5348, 1.4244],
[ 1.0477, 0.2311, 1.0946, 0.6907, 0.8061]])
In [13]: y
Out[13]:
tensor([[ 0.6177, 0.6424, 0.3739, 0.6916, 0.2209],
[ 1.3712, 0.5577, 1.3626, 1.0993, 1.5942],
[ 1.5684, 0.7533, 1.0628, 1.1381, 1.2293],
[ 0.1066, 1.5505, 1.0467, 1.5348, 1.4244],
[ 1.0477, 0.2311, 1.0946, 0.6907, 0.8061]])
其他的一些生成操作
In [15]: a = t.ones(5) #生成一个全是1的Tensor
In [16]: a
Out[16]: tensor([ 1., 1., 1., 1., 1.])
In [17]: b = t.zeros(5) #生成一个全是0的Tensor
In [18]: b
Out[18]: tensor([ 0., 0., 0., 0., 0.])
支持numpy的操作
In [19]: x[:, 1] #访问一列的数据
Out[19]: tensor([ 0.5416, 0.2257, 0.6790, 0.5616, 0.1685])
In [21]: c = np.ones(5)
In [22]: c
Out[22]: array([ 1., 1., 1., 1., 1.])
In [20]: import numpy as np
In [21]: c = np.ones(5)
In [22]: c
Out[22]: array([ 1., 1., 1., 1., 1.])
In [23]: d = t.from_numpy(c)
In [24]: d
Out[24]: tensor([ 1., 1., 1., 1., 1.], dtype=torch.float64)
由于电脑本身的问题无法使用GPU,这里只是简单的介绍
通过.cuda方法转换为GPU的Tensor，从而加速运算，这里注意当加速的Tensor比较小的时候，不建议使用GPU，使用GPU的优势是在大规模数据和复杂运算才能体现。
