pytorch的索引，view和mean方法的维度问题

1. shape与dim方法

 a = torch.randn(2,3,4)
print(a)
tensor([[[-1.2323, -0.5237,  0.0081, -0.7642],
         [ 1.2312, -1.2463, -1.2034,  1.4166],
         [ 1.5238,  0.1296,  0.7698,  0.4130]],
        [[ 0.7246,  0.3782,  0.9749, -1.3180],
         [ 0.4092, -0.9307,  0.8133, -0.8961],
         [ 0.5478,  0.9748, -0.2082,  1.0809]]])
a.shape
torch.Size([2, 3, 4])
a.dim
3 # 有三个维度，每个维度分别为2,3,4

2. view
   Note： 1. view简单来说就是把原tensor的元素按顺序取出，填入新的tensor中。
   2. view中的参数乘积必须为24否则会报错

b = a.view(3,2,-1)
print(b)
tensor([[[-1.2323, -0.5237,  0.0081, -0.7642],
         [ 1.2312, -1.2463, -1.2034,  1.4166]],
        [[ 1.5238,  0.1296,  0.7698,  0.4130],
         [ 0.7246,  0.3782,  0.9749, -1.3180]],
        [[ 0.4092, -0.9307,  0.8133, -0.8961],
         [ 0.5478,  0.9748, -0.2082,  1.0809]]])
c = a.view(4,3,-1)
print(c)
tensor([[[-1.2323, -0.5237],
         [ 0.0081, -0.7642],
         [ 1.2312, -1.2463]],
        [[-1.2034,  1.4166],
         [ 1.5238,  0.1296],
         [ 0.7698,  0.4130]],
        [[ 0.7246,  0.3782],
         [ 0.9749, -1.3180],
         [ 0.4092, -0.9307]],
        [[ 0.8133, -0.8961],
         [ 0.5478,  0.9748],
         [-0.2082,  1.0809]]])
c = a.view(4,3)
Traceback (most recent call last):
  File "D:\anaconda\lib\site-packages\IPython\core\interactiveshell.py", line 3296, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-68-4f6ba6dedc05>", line 1, in <module>
    c = a.view(4,3)
RuntimeError: shape '[4, 3]' is invalid for input of size 24
c = a.view(5,3,-1)
Traceback (most recent call last):
  File "D:\anaconda\lib\site-packages\IPython\core\interactiveshell.py", line 3296, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-69-5d309fd757fa>", line 1, in <module>
    c = a.view(5,3,-1)
RuntimeError: shape '[5, 3, -1]' is invalid for input of size 24

3. mean
   当不指定参数时，输出为所有元素的平均值，为一个标量;
   当指定dim时,假设dim=i。如果第i个维度为Ni，则从Ni个分量中各取一个元素，求平均值,所得tensor的shape为N1N2…N(i-1)N(i+1)…Nn
   若keepdim=True，则所得tensor的shape为N1N2*…N(i-1)1N(i+1)…Nn

	a.shape
Out[70]: torch.Size([2, 3, 4])
a.mean()
Out[71]: tensor(0.1280)
a.mean(0)
Out[72]: 
tensor([[-0.2539, -0.0727,  0.4915, -1.0411],
        [ 0.8202, -1.0885, -0.1951,  0.2603],
        [ 1.0358,  0.5522,  0.2808,  0.7469]])
a.mean(0).shape
Out[5]: torch.Size([3, 4])
a.mean(0,keepdim=True).shape
Out[6]: torch.Size([1, 3, 4])

相当于

m,b,c = a.shape
mean0_a = torch.zeros(b,c)
for i in range(b):
    for j in range(c):
        sum = 0
        for k in range(a):
            sum += a[k,i,j]
        sum /= m
        mean0_a[i,j] = sum

4. 索引
   引用元素或切片有两种格式a[i][j]或a[i,j]
   切片为a[:][[j]或a[:,j]
   推荐第二种，更简洁