It's a Python based scientific computing package targeted at two sets of audiences:
- A replacement for
numpyto use the power of GPUs - A deep learning research platform that provides maximum flexibility and speed
In [1]:
import torchIn [2]:
# Construct a 5x3 matrix
x = torch.Tensor(5, 3)
print(x)1.00000e-29 *
0.0000 0.0000 3.9260
0.0000 0.0000 0.0000
0.0000 0.0000 0.0000
0.0000 0.0000 0.0000
0.0000 0.0000 0.0000
[torch.FloatTensor of size 5x3]
In [3]:
# Construct a 5x3 randomly initialized matrix
x = torch.rand(5, 3)
print(x) 0.9152 0.4046 0.7690
0.6653 0.6564 0.9553
0.4627 0.0858 0.8089
0.2361 0.9341 0.0296
0.0496 0.9051 0.3340
[torch.FloatTensor of size 5x3]
In [4]:
# get it's size
x.size()torch.Size([5, 3])In [5]:
y = torch.FloatTensor(3, 5)
print(y) 0.0000e+00 0.0000e+00 2.6258e-29 3.6893e+19 5.6052e-45
0.0000e+00 0.0000e+00 0.0000e+00 0.0000e+00 0.0000e+00
0.0000e+00 0.0000e+00 0.0000e+00 0.0000e+00 2.6372e-29
[torch.FloatTensor of size 3x5]
In [6]:
x = torch.Tensor(5, 3)
y = torch.rand(5, 3)In [7]:
print(x + y) 0.4040 0.5859 0.8088
0.2688 0.1638 0.1029
0.0107 0.7332 0.3672
0.2479 0.1942 0.1848
0.3706 0.5134 0.4678
[torch.FloatTensor of size 5x3]
In [8]:
out = x + y
print(out) 0.4040 0.5859 0.8088
0.2688 0.1638 0.1029
0.0107 0.7332 0.3672
0.2479 0.1942 0.1848
0.3706 0.5134 0.4678
[torch.FloatTensor of size 5x3]
In [9]:
# syntax 2:
print(torch.add(x, y)) 0.4040 0.5859 0.8088
0.2688 0.1638 0.1029
0.0107 0.7332 0.3672
0.2479 0.1942 0.1848
0.3706 0.5134 0.4678
[torch.FloatTensor of size 5x3]
In [10]:
out = torch.add(x, y)
print(out) 0.4040 0.5859 0.8088
0.2688 0.1638 0.1029
0.0107 0.7332 0.3672
0.2479 0.1942 0.1848
0.3706 0.5134 0.4678
[torch.FloatTensor of size 5x3]
In [11]:
result = torch.Tensor(5, 3)
torch.add(x, y, out=result)
print(result) 0.4040 0.5859 0.8088
0.2688 0.1638 0.1029
0.0107 0.7332 0.3672
0.2479 0.1942 0.1848
0.3706 0.5134 0.4678
[torch.FloatTensor of size 5x3]
In [12]:
# in-place
y.add_(x)
print(y) 0.4040 0.5859 0.8088
0.2688 0.1638 0.1029
0.0107 0.7332 0.3672
0.2479 0.1942 0.1848
0.3706 0.5134 0.4678
[torch.FloatTensor of size 5x3]
NOTE: Any operation that mutates a tensor in-pace is post-fixed with an _. For example: x.copy_(y), x.t_(), will change x.
In [13]:
# You can use standard NumPy-like indexing
print(x[:, 1])1.00000e-45 *
0.0000
7.0065
0.0000
0.0000
0.0000
[torch.FloatTensor of size 5]
In [14]:
x = torch.rand(4, 4)
y = x.view(16)
z = x.view(-1, 8) # the size -1 is inferred from other dimensions
print('x.size() = {}'.format(x.size()))
print('y.size() = {}'.format(y.size()))
print('z.size() = {}'.format(z.size()))x.size() = torch.Size([4, 4])
y.size() = torch.Size([16])
z.size() = torch.Size([2, 8])
In [15]:
# Tensor.view_as(Tensor)
a = x.view_as(z) # inferred from a Tensor shape
print('a.size() = {}'.format(a.size()))a.size() = torch.Size([2, 8])
In [16]:
# Taking a dot product of x with y using the Python 3.5 syntax
product = x@y
print(product) 0.0000e+00 0.0000e+00 2.4032e-29 3.3767e+19 2.0280e-29
0.0000e+00 0.0000e+00 1.7468e-29 2.4544e+19 2.5193e-29
0.0000e+00 0.0000e+00 1.2150e-29 1.7072e+19 2.1332e-29
0.0000e+00 0.0000e+00 6.1984e-30 8.7092e+18 7.8053e-31
0.0000e+00 0.0000e+00 1.3022e-30 1.8297e+18 8.8081e-30
[torch.FloatTensor of size 5x5]
Converting a Torch Tensor to a NumPy array and vice versa is a breeze.
NOTE: The Torch Tensor and NumPy array will share their underlying memory locations, and changing one will change the other.
In [17]:
import numpy as npIn [18]:
a = torch.ones(5)
print(a)
print(b) 1
1
1
1
1
[torch.FloatTensor of size 5]
In [19]:
b = a.numpy()
print(b)[1. 1. 1. 1. 1.]
See how the numpy array changed in value
In [20]:
a.add_(1)
print(a)
print(b) 2
2
2
2
2
[torch.FloatTensor of size 5]
[2. 2. 2. 2. 2.]
In [21]:
# b is a numpy array from the torch Tensor, a
# c is a torch tensor created from the numpy array, b
# a, b & c shares the same memory location
c = torch.from_numpy(b)
print(c) 2
2
2
2
2
[torch.FloatTensor of size 5]
In [22]:
a.add_(3)
print(f'a: {a} -Torch')
print(f'b: {b} -Numpy')
print(f'c: {c} -Torch')a:
5
5
5
5
5
[torch.FloatTensor of size 5]
-Torch
b: [5. 5. 5. 5. 5.] -Numpy
c:
5
5
5
5
5
[torch.FloatTensor of size 5]
-Torch
See how the change in the np array changed the Torch Tensor automatically
In [23]:
a = np.ones(5)
b = torch.from_numpy(a)
# Add 1 to a and store the result in a
# i.e. Update the value of a by adding 1 to it.
np.add(a, 1, out=a)
print(a)[2. 2. 2. 2. 2.]
2
2
2
2
2
[torch.DoubleTensor of size 5]
Tensors can be moved into the GPU using the .cuda method.
In [24]:
# let us run this cell only if CUDA is available
if torch.cuda.is_available():
# Create `x` as a Float tensor on the GPU
x = torch.cuda.FloatTensor(5, 3)
# Randomly initialize y ont the GPU
y = torch.rand(5, 3).cuda()
result = torch.Tensor(5, 3).cuda()
# Add a & b on the GPU
torch.cuda.torch.add(a, b, out=result)
# PRINT THE RESULT
print(x + y)
else:
print("You don't have a GPU or CUDA isn't enabled!")You don't have a GPU or CUDA isn't enabled!