[Pytorch] 설치부터 기본 사용법
Pytorch 주피터 노트북 설치하기
pip install torch torchvision torchaudio
Collecting torch
Downloading torch-1.12.1-cp39-cp39-win_amd64.whl (161.8 MB)
Collecting torchvision
Downloading torchvision-0.13.1-cp39-cp39-win_amd64.whl (1.1 MB)
Collecting torchaudio
Downloading torchaudio-0.12.1-cp39-cp39-win_amd64.whl (969 kB)
Requirement already satisfied: typing-extensions in c:\anaconda\lib\site-packages (from torch) (4.1.1)
Requirement already satisfied: numpy in c:\anaconda\lib\site-packages (from torchvision) (1.21.6)
Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in c:\anaconda\lib\site-packages (from torchvision) (9.0.1)
Requirement already satisfied: requests in c:\anaconda\lib\site-packages (from torchvision) (2.27.1)
Requirement already satisfied: certifi>=2017.4.17 in c:\anaconda\lib\site-packages (from requests->torchvision) (2022.9.24)
Requirement already satisfied: idna<4,>=2.5 in c:\anaconda\lib\site-packages (from requests->torchvision) (3.3)
Requirement already satisfied: urllib3<1.27,>=1.21.1 in c:\anaconda\lib\site-packages (from requests->torchvision) (1.26.9)
Requirement already satisfied: charset-normalizer~=2.0.0 in c:\anaconda\lib\site-packages (from requests->torchvision) (2.0.4)
Installing collected packages: torch, torchvision, torchaudio
Successfully installed torch-1.12.1 torchaudio-0.12.1 torchvision-0.13.1
Note: you may need to restart the kernel to use updated packages.
기본 연산 및 동작
- numpy와 유사함
import numpy as np
import torch
n_array = np.arange(10).reshape(2, 5)
print(n_array)
print('dim: ', n_array.ndim, ', shape: ', n_array.shape)
[[0 1 2 3 4]
[5 6 7 8 9]]
dim: 2 , shape: (2, 5)
# Array to Tensor
data =[[3, 5], [10, 5]]
x_data = torch.tensor(data)
x_data
#ndarray to Tensor
array = np.array(data)
tensor_array = torch.from_numpy(array)
tensor_array
tensor([[ 3, 5],
[10, 5]], dtype=torch.int32)
# Operations like numpy generally
torch.ones_like(x_data)
tensor([[1, 1],
[1, 1]])
n_array = np.random.random((2, 2))
t = torch.from_numpy(n_array)
t
tensor([[0.8497, 0.9555],
[0.9499, 0.4401]], dtype=torch.float64)
t[1:]
tensor([[0.9499, 0.4401]], dtype=torch.float64)
t[:1]
tensor([[0.8497, 0.9555]], dtype=torch.float64)
n = np.random.random((3, 5))
t2 = torch.from_numpy(n)
t2
tensor([[0.0011, 0.7782, 0.9890, 0.3174, 0.9397],
[0.9175, 0.6137, 0.1440, 0.6878, 0.7808],
[0.8496, 0.1207, 0.6045, 0.6222, 0.9297]], dtype=torch.float64)
t2[1:2, 2:]
tensor([[0.1440, 0.6878, 0.7808]], dtype=torch.float64)
t2[1:3]
tensor([[0.9175, 0.6137, 0.1440, 0.6878, 0.7808],
[0.8496, 0.1207, 0.6045, 0.6222, 0.9297]], dtype=torch.float64)
t2.flatten()
tensor([0.0011, 0.7782, 0.9890, 0.3174, 0.9397, 0.9175, 0.6137, 0.1440, 0.6878,
0.7808, 0.8496, 0.1207, 0.6045, 0.6222, 0.9297], dtype=torch.float64)
t2.shape
torch.Size([3, 5])
t2.dtype
torch.float64
t2.numpy()
array([[0.00113862, 0.77818119, 0.98904411, 0.31744628, 0.93970248],
[0.9174607 , 0.61374777, 0.14404259, 0.68778041, 0.78075312],
[0.84956723, 0.12072346, 0.60445122, 0.62224636, 0.9297356 ]])
torch.ones_like(t2)
tensor([[1., 1., 1., 1., 1.],
[1., 1., 1., 1., 1.],
[1., 1., 1., 1., 1.]], dtype=torch.float64)
t2.device
device(type='cpu')
x_data = t2
if torch.cuda.is_available():
x_data_cuda = x_data.to('cuda')
x_data.device
device(type='cpu')
형변환, 차원 축소추가
- view: reshape 과 동일하게 tensor의 shape 변환 (강의에서는 shape보다 view 권장; 메모리 저장)
- squeeze: 차원 개수가 1인 차원 삭제 (압축)
- unsqueeze: 차원 개수가 1인 차원 추가 (사이즈 맞출 때 종종 사용)
tensor_ex = torch.rand(size=(2, 3, 2))
tensor_ex
tensor([[[0.2809, 0.2419],
[0.6830, 0.6607],
[0.8493, 0.3878]],
[[0.2236, 0.8167],
[0.1403, 0.5048],
[0.6939, 0.0065]]])
tensor_ex.view([-1, 6]) # 1 by 6 2개로 변환
tensor([[0.2809, 0.2419, 0.6830, 0.6607, 0.8493, 0.3878],
[0.2236, 0.8167, 0.1403, 0.5048, 0.6939, 0.0065]])
tensor_ex.reshape([-1, 6])
tensor([[0.2809, 0.2419, 0.6830, 0.6607, 0.8493, 0.3878],
[0.2236, 0.8167, 0.1403, 0.5048, 0.6939, 0.0065]])
# view, reshape 차이점 데이터 유지 측면에서 다름
a = torch.zeros(3, 2)
b = a.view(2, 3)
a.fill_(1)
tensor([[1., 1.],
[1., 1.],
[1., 1.]])
c= torch.zeros(3, 2)
d= c.t().reshape(6)
d.fill_(1)
tensor([1., 1., 1., 1., 1., 1.])
data = [[1, 2], [3, 4]]
t = torch.tensor(data)
t
tensor([[1, 2],
[3, 4]])
t.unsqueeze(0)
tensor([[[1, 2],
[3, 4]]])
t.unsqueeze(1)
tensor([[[1, 2]],
[[3, 4]]])
t.unsqueeze(2)
tensor([[[1],
[2]],
[[3],
[4]]])
t.squeeze(0)
tensor([[1, 2],
[3, 4]])
tensor_ex = torch.rand(size=(2,2))
tensor_ex.unsqueeze(0).shape
torch.Size([1, 2, 2])
tensor_ex.unsqueeze(1).shape
torch.Size([2, 1, 2])
tensor_ex.unsqueeze(2).shape
torch.Size([2, 2, 1])
n1 = np.arange(10).reshape(2, 5)
t1 = torch.FloatTensor(n1)
t1
tensor([[0., 1., 2., 3., 4.],
[5., 6., 7., 8., 9.]])
t1 + t1
tensor([[ 0., 2., 4., 6., 8.],
[10., 12., 14., 16., 18.]])
t1+ 10
tensor([[10., 11., 12., 13., 14.],
[15., 16., 17., 18., 19.]])
t1*5
tensor([[ 0., 5., 10., 15., 20.],
[25., 30., 35., 40., 45.]])
t1 - t1
tensor([[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.]])
n2 = np.arange(10).reshape(5, 2)
t2 = torch.FloatTensor(n2)
t2
tensor([[0., 1.],
[2., 3.],
[4., 5.],
[6., 7.],
[8., 9.]])
- 행렬계산:
mm사용!!
t1.mm(t2)
tensor([[ 60., 70.],
[160., 195.]])
t1.dot(t2)
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
Input In [64], in <cell line: 1>()
----> 1 t1.dot(t2)
RuntimeError: 1D tensors expected, but got 2D and 2D tensors
t1.matmul(t2) # broadcasting 처리
tensor([[ 60., 70.],
[160., 195.]])
a = torch.rand(5, 2, 3)
b = torch.rand(5)
a.mm(b)
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
Input In [73], in <cell line: 4>()
1 a = torch.rand(5, 2, 3)
2 b = torch.rand(5)
----> 4 a.mm(b)
RuntimeError: self must be a matrix
a = torch.rand(5, 2, 3)
b = torch.rand(3)
a.matmul(b)
tensor([[1.5056, 0.8395],
[1.4356, 0.5943],
[0.6893, 0.5864],
[0.4444, 1.0078],
[0.7802, 0.7125]])
a
tensor([[[0.8785, 0.7743, 0.8483],
[0.1465, 0.6375, 0.7575]],
[[0.7441, 0.8085, 0.8749],
[0.4431, 0.3386, 0.1816]],
[[0.2375, 0.8554, 0.1997],
[0.1116, 0.4353, 0.5250]],
[[0.0756, 0.3286, 0.4108],
[0.6210, 0.2485, 0.7427]],
[[0.6676, 0.2477, 0.2849],
[0.3762, 0.0232, 0.7308]]])
b
tensor([0.7524, 0.4632, 0.5729])
a[0].mm(torch.unsqueeze(b,1))
a[1].mm(torch.unsqueeze(b,1))
a[2].mm(torch.unsqueeze(b,1))
a[3].mm(torch.unsqueeze(b,1))
a[4].mm(torch.unsqueeze(b,1))
tensor([[0.7802],
[0.7125]])
수식 변환지원 nn.functional
import torch as F
tensor = torch.FloatTensor([0.5, 0.7, 0.1])
h_tensor = F.softmax(tensor, dim=0)
h_tensor
tensor([0.3458, 0.4224, 0.2318])
y = torch.randint(5, (10, 5))
y
tensor([[4, 2, 3, 1, 3],
[2, 3, 1, 1, 1],
[1, 3, 3, 2, 3],
[3, 4, 4, 1, 0],
[3, 2, 3, 3, 1],
[2, 4, 2, 2, 0],
[3, 2, 1, 4, 2],
[3, 3, 0, 0, 0],
[2, 0, 4, 0, 2],
[0, 3, 2, 4, 2]])
y_label = y.argmax(dim=1)
y_label
tensor([0, 1, 1, 1, 0, 1, 3, 0, 2, 3])
torch.nn.functional.one_hot(y_label)
tensor([[1, 0, 0, 0],
[0, 1, 0, 0],
[0, 1, 0, 0],
[0, 1, 0, 0],
[1, 0, 0, 0],
[0, 1, 0, 0],
[0, 0, 0, 1],
[1, 0, 0, 0],
[0, 0, 1, 0],
[0, 0, 0, 1]])
자동미분 -> backward 함수 사용
https://tutorials.pytorch.kr/beginner/blitz/autograd_tutorial.html
w = torch.tensor(2.0, requires_grad=True)
y = w**2
z = 10*y+2
z.backward()
z
tensor(42., grad_fn=<AddBackward0>)
w.grad
tensor(40.)
# 변화도 gradient 확인
a = torch.tensor([2., 3.], requires_grad=True)
b = torch.tensor([6., 4.], requires_grad=True)
Q = 3*a**3 - b**2
external_grad = torch.tensor([1, 1])
Q.backward(gradient=external_grad)
a.grad
tensor([36., 81.])
b.grad
tensor([-12., -8.])
# 수집된 변화도가 올바른지 확인
print(9*a**2 == a.grad)
print(-2*b == b.grad)
tensor([True, True])
tensor([True, True])
댓글남기기