Dive into deep learning - 2.1 data operation & Exercise

Catalog

1.reshape

2、randn

3. Operator

4. Broadcast mechanism

5. practice

1.reshape

effect ： Change the shape of the tensor

X = x.reshape(3, 4)
X

tensor([[ 0,  1,  2,  3],
        [ 4,  5,  6,  7],
        [ 8,  9, 10, 11]])

In the example above , In order to get a 3 Matrix of rows , We manually specified that it has 3 Row sum 4 Column . Fortunately, , We can go through -1 To call this function of automatically calculating dimensions . That is, we can use x.reshape(-1,4) or x.reshape(3,-1) To replace x.reshape(3,4).

2、randn

When we construct an array as a parameter in a neural network , We usually initialize the value of the parameter randomly . The following code creates a shape of （3,4） Tensor . Each of these elements has a mean value of 0、 The standard deviation is 1 Of Standard Gaussian distribution （ Normal distribution ） Medium random sampling .

torch.randn(3, 4)

tensor([[-0.5979,  0.6042, -1.1094, -0.9001],
        [ 1.3877,  0.1119,  0.2164,  0.7615],
        [-0.6081, -0.2789, -0.7933,  0.7229]])

3. Operator

x = torch.tensor([1.0, 2, 4, 8])
y = torch.tensor([2, 2, 2, 2])
x + y, x - y, x * y, x / y, x ** y  # ** Operators are exponentiation operations 

(tensor([ 3.,  4.,  6., 10.]),
 tensor([-1.,  0.,  2.,  6.]),
 tensor([ 2.,  4.,  8., 16.]),
 tensor([0.5000, 1.0000, 2.0000, 4.0000]),
 tensor([ 1.,  4., 16., 64.]))

Be careful ： Here, only one element in the tensor is a floating point number , Then all the elements of the calculated result tensor are floating-point numbers .

x**y Express x Of y Power

torch.exp(x)

tensor([2.7183e+00, 7.3891e+00, 5.4598e+01, 2.9810e+03])

For tensors x Index each element of ,x by 【1,2,4,8】, So the result is 【e Of 1 Power ,e Of 2 Power ,e Of 4 Power ,e Of 8 Power 】

4. Broadcast mechanism

a = torch.arange(3).reshape((3, 1))
b = torch.arange(2).reshape((1, 2))
a, b

(tensor([[0],
         [1],
         [2]]),
 tensor([[0, 1]]))

a + b
tensor([[0, 1],
        [1, 2],
        [2, 3]])

because a and b Namely 3×1 and 1×2 matrix , If you add them together , Their shapes don't match . We will use two matrices radio broadcast For a bigger 3×2 matrix , As shown below ： matrix a Columns will be copied , matrix b The row will be copied , Then add by element .

5. practice

1. Run the code in this section . The conditional statements in this section X == Y Change to X < Y or X > Y, Then see what kind of tensor you can get .

X = torch.tensor([1,2,3,4])
Y = torch.tensor([2,3,3,3])
print(X > Y)
print(X < Y)

tensor([False, False, False,  True])
tensor([ True,  True, False, False])

2. Use other shapes （ For example, three-dimensional tensors ） Replace two tensors operated by elements in the broadcast mechanism . Whether the result is the same as expected ？ 

# practice 2
a = torch.arange(12).reshape(3,1,4)
b = torch.ones(1,2,1)
print(a)
print(b)
print(a + b)

tensor([[[ 0,  1,  2,  3]],

        [[ 4,  5,  6,  7]],

        [[ 8,  9, 10, 11]]])
tensor([[[1.],
         [1.]]])
tensor([[[ 1.,  2.,  3.,  4.],
         [ 1.,  2.,  3.,  4.]],

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

        [[ 9., 10., 11., 12.],
         [ 9., 10., 11., 12.]]])