8. Mathematical operation and attribute statistics

Mathematical operations

One .torch.add()/minus()/multply()/divide()

It is recommended to use the operator directly
add

a=torch.rand(3,4)
b=torch.rand(4)
print(a)
print(b)
print(torch.add(a,b))

minus

multply

divide

Two .matmul

matrix multiplication A[a,b], ride B[b,c], Finally, the matrix C[a,c],C The value of the first row and the first column of is equal to A The first line of is multiplied by B The sum of the values in the first column of .

1.torch.mm

Only suitable for 2D

2.torch.matmul recommend

a=torch.ones(2,2)
b=torch.full([2,2],6,dtype=torch.float32)
c=torch.matmul(a,b)
print(a)
print(b)
print(c)

[4,3,28,64] multiply [4,3,64,32], The first two dimensions remain unchanged , The latter two dimensions follow 2D The operation rules of [4,3,28,32]
[4,3,28,64] multiply [4,1,64,32], The second satisfaction broadcast, become [4,3,64,32], And then multiply it .

a=torch.rand(4,3,28,64)
b=torch.rand(4,3,64,32)
c=torch.rand(4,1,64,32)
d=torch.matmul(a,b)
e=torch.matmul(a,c)
print(d.shape)
print(e.shape)

[email protected]

A more concise method , The same principle as the previous one

a=torch.rand(4,3,28,64)
b=torch.rand(4,3,64,32)
c=[email protected]
print(c.shape)

4.example

x=torch.rand(4,784)
w=torch.rand(512,784)
print(([email protected].t()).shape)

When we reduce the dimension , take [4,784] Multiplication matrix [784,512] available [4,512],w Not written [784,512], Because pytorch Fixed writing , First channe-out, Again channel-in.

3、 ... and .pow/exp/log

.pow()

a=torch.full([2,2],3)
print(a.pow(2))

torch.exp()

exp(a), seek e Of a Power .

a=torch.ones([2,2])
a=torch.exp(a)
print(a)

torch.log()

The log In order to e Base number

a=torch.ones([2,2])
a=torch.exp(a)
print(torch.log(a))

Four .sqrt/rsqrt

.sqrt()

square root

a=torch.full([2,2],3)
aa=a.pow(2)
print(aa.sqrt())

.rsqrt()

The reciprocal of the square root

a=torch.full([2,2],3)
aa=a.pow(2)
print(aa.rsqrt())

5、 ... and .approximation

1…floor()/.ceil()

.floor() Rounding down ,.ceil() Rounding up

a=torch.tensor(3.14)
print(a.floor(),a.ceil())

2…trunc()/.frac()

.trunc() Crop integer parts ,.frac() Cut the decimal part

a=torch.tensor(3.14)
print(a.trunc())
print(a.frac())

3…round()

.round() rounding

a=torch.tensor(3.54)
print(a.round())

6、 ... and …clamp() Use more

gradient clipping Gradient cut

a=torch.rand(2,3)*15
print(a)
print(a.clamp(8))
print(a.clamp(5,8))

a.clamp(8), Less than 8 To become 8
a.clamp(5,8) Count in 5 To 8 Between , Greater than 8 To become 8, Less than 5 To become 5.

Attribute Statistics

.norm()

1-norm:

P-norm:

.mean()/.sum()/.min()/.max()/.prod()

product, Multiplicative multiplication

a=torch.arange(8).view(2,4).float()
print(a)
print(a.mean())
print(a.sum())
print(a.min())
print(a.max())
print(a.prod())

dim/keepdim

a[4,10]
a.max(dim=1)z
tensor（0.8,）,tensor（3,）
first tensor yes 4 Probability values , the second tensor It shows the location of each probability value （ Indexes ）, The first one is 3 Is the probability that 0.8
To which dim Conduct max, Take one dim Equivalent to being eliminated .keepdim yes dim Be consistent with the original , return [4,1]

.argmax()/.argmin()

argmax() Returns the index of the maximum value
a[4,10]
a.argmax(dim=1), Then return to 4 One is the location of each probability value （ Indexes ）

.argmin() Returns the index of the minimum value , Will be able to tensor Strike a level , Then look at the index , If you don't want to draw , Look directly at a dimension , The parameter (dim=)

a=torch.arange(8).view(2,4).float()
print(a)
print(a.argmax())
print(a.argmin())

.kthvalue()/.topk()

kthvalue() Return to the smallest k Number , amount to [4,10] obtain [4,k]
topk() Return to the biggest k Number , amount to [4,10] obtain [4,k]

compare

>,>= No true and false, It is 0 and 1.

torch.eq(a,b) Judge a and b Whether each position is equal
torch.equal(a,b) return true