Broadcast 机制——numpy

几种情况：
这里只讲加减，矩阵乘法严格按照矩阵乘法规则（尾首同）。

• 注意比较维度数时从右至左。
• (2， 3， 5）维数为3， 各维维度数为2， 3， 5

1.维数相等时。

• 各维度数均相等。这个不需多说。按位运算即可。
• 至少有一维维度数不等，不可广播。
  Exception：上条中，不等的维度数，有为1者，可广播。
  看例子：

a = np.arange(9).reshape( 3, 3) 
b = np.arange(6).reshape( 2, 3)
print(a, "\n", b, "\n", a+b)  
# ValueError: operands could not be broadcast together with shapes (3,3) (2,3)

a = np.arange(9).reshape( 3, 3) 
b = np.arange(6).reshape( 3, 2)
print(a, "\n", b, "\n", a+b)
# ValueError: operands could not be broadcast together with shapes (3,3) (3,2) 

a = np.arange(9).reshape( 3, 3) 
b = np.arange(3).reshape( 3, 1)
print(a, "\n", b, "\n", a+b)  # 可打印

a = np.arange(9).reshape( 3, 3) 
b = np.arange(2).reshape( 2, 1)
print(a, "\n", b, "\n", a+b)
# ValueError: operands could not be broadcast together with shapes (3,3) (2,1) 

2.维度不等时。

• 从右至左，至少有一维维度数不等时不可广播。（同上：不同处，为空或维度数为1者除外）

a = np.arange(24).reshape(4, 2, 3) 
b = np.arange(9).reshape(3, 3)
print(a, "\n", b, "\n", a+b)
# ValueError: operands could not be broadcast together with shapes (4,2,3) (3,3) 

a = np.arange(18).reshape(2, 3, 3) 
b = np.arange(2).reshape(2)
print(a, "\n", b, "\n", a+b)
# ValueError: operands could not be broadcast together with shapes (2,3,3) (2,) 

a = np.arange(12).reshape(4,1,3) 
b = np.arange(9).reshape(3, 3)
print(a, "\n", b, "\n", a+b)  # 可打印

a = np.arange(18).reshape(2,3,3) 
b = np.arange(3).reshape(3)
print(a, "\n", b, "\n", a+b) # 可打印

其实抓住第二点。就可以全部概括这个BroadCast。