[PyTorch] 다중 선형 회귀

Posted: 2024.12.10 Updated: 2024.12.10

AI 스터디를 하며 ‘파이토치 트랜스포머를 활용한 자연어 처리와 컴퓨터비전 심층학습’ 교재를 정리한 글입니다.

Dataset과 DataLoader

데이터셋(Dataset) 클래스

class Dataset:
    def __init__(self, data, *arg, **kwargs):
        self.data = data

    def __getitem__(self, index):
        return tuple(data[index] for data in self.data.tensors)

    def __len__(self):
        return self.data[0].size(0)

데이터로더(DataLoader) 클래스

batch_size, shuffle, num_workers 등 기능 제공
batch_size: 전체 데이터 셋에서 배치 크기만큼 데이터 샘플을 나누어 모든 배치 대상으로 학습
- 배치 크기가 1인 배치를 불완전한 배치(Incomplete Batch)라고 함
shuffle: 데이터 순서 학습을 막기 위해 행 순서 변경
num_workers: 데이터를 불러올 때 사용하는 프로세스 수

다중 선형 회귀

\[\begin{align} y_1 = w_1 x_1 + w_2 x_2 + b_1 \\ y_2 = w_3 x_1 + w_4 x_2 + b_2 \end{align}\]

x1, x2: 독립 변수
y1, y2: 종속 변수
모델 매개변수 형태

\[\text{Weight} = \begin{bmatrix} w_1 & w_2 \\ w_3 & w_4 \end{bmatrix} = \begin{bmatrix} 1.7 & -0.8 \\ 1.1 & 0.2 \end{bmatrix}\] \[\text{Bias} = \begin{bmatrix} b_1 \\ b_2 \end{bmatrix} = \begin{bmatrix} 0 \\ 0 \end{bmatrix}\]

10,000회 학습 결과

\[\text{Weight} = \begin{bmatrix} w_1 & w_2 \\ w_3 & w_4 \end{bmatrix} = \begin{bmatrix} 0.8990 & -0.0235 \\ 1.0535 & 0.2315 \end{bmatrix}\] \[\text{Bias} = \begin{bmatrix} b_1 \\ b_2 \end{bmatrix} = \begin{bmatrix} -0.6733 \\ 0.0315 \end{bmatrix}\]

import torch
from torch import nn
from torch import optim
from torch.utils.data import TensorDataset, DataLoader

train_x = torch.FloatTensor([
    [1, 2], [2, 3], [3, 4], [4, 5], [5, 6], [6, 7]
])

train_y = torch.FloatTensor([
    [0.1, 1.5], [1, 2.8], [1.9, 4.1], [2.8, 5.4], [3.7, 6.7], [4.6, 8]
])

train_dataset = TensorDataset(train_x, train_y)  # TensorDataset: 기본 데이터세트 클래스를 상속받아 재정의한 클래스
train_dataloader = DataLoader(train_dataset, batch_size=2, shuffle=True, drop_last=True)

model = nn.Linear(2, 2, bias=True)
criterion = nn.MSELoss()
optimizer = optim.SGD(model.parameters(), lr=0.001)

# 편향 제거
# model = nn.Linear(2, 2, bias=False)
# criterion = nn.MSELoss()
# optimizer = optim.SGD(model.parameters(), lr=0.001)

for epoch in range(20000):
    cost = 0.0
    
    for batch in train_dataloader:  # 데이터로더를 반복하여 배치 반환
        x, y = batch
        output = model(x)
        loss = criterion(output, y)
    
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
    
        cost += loss

    cost = cost / len(train_dataloader)  # 오차 평균 계산

    if (epoch + 1) % 1000 == 0:
        print(f"Epoch : {epoch + 1:4d}, Model : {list(model.parameters())}, Cost : {cost:.3f}")

출력

Epoch : 1000, Model : [Parameter containing:
tensor([[0.3475, 0.3747],
        [1.2057, 0.0328]], requires_grad=True), Parameter containing:
tensor([-0.4251,  0.4266], requires_grad=True)], Cost : 0.061
Epoch : 2000, Model : [Parameter containing:
tensor([[5.2809e-01, 2.8108e-01],
        [1.2682e+00, 3.9314e-04]], requires_grad=True), Parameter containing:
tensor([-0.6994,  0.3317], requires_grad=True)], Cost : 0.016
Epoch : 3000, Model : [Parameter containing:
tensor([[ 0.6202,  0.2335],
        [ 1.3000, -0.0161]], requires_grad=True), Parameter containing:
tensor([-0.8391,  0.2834], requires_grad=True)], Cost : 0.004
Epoch : 4000, Model : [Parameter containing:
tensor([[ 0.6671,  0.2093],
        [ 1.3163, -0.0244]], requires_grad=True), Parameter containing:
tensor([-0.9102,  0.2587], requires_grad=True)], Cost : 0.001
Epoch : 5000, Model : [Parameter containing:
tensor([[ 0.6911,  0.1970],
        [ 1.3246, -0.0287]], requires_grad=True), Parameter containing:
tensor([-0.9465,  0.2462], requires_grad=True)], Cost : 0.000
Epoch : 6000, Model : [Parameter containing:
tensor([[ 0.7032,  0.1907],
        [ 1.3288, -0.0309]], requires_grad=True), Parameter containing:
tensor([-0.9649,  0.2398], requires_grad=True)], Cost : 0.000
Epoch : 7000, Model : [Parameter containing:
tensor([[ 0.7094,  0.1875],
        [ 1.3309, -0.0320]], requires_grad=True), Parameter containing:
tensor([-0.9743,  0.2365], requires_grad=True)], Cost : 0.000
Epoch : 8000, Model : [Parameter containing:
tensor([[ 0.7126,  0.1858],
        [ 1.3320, -0.0326]], requires_grad=True), Parameter containing:
tensor([-0.9791,  0.2349], requires_grad=True)], Cost : 0.000
Epoch : 9000, Model : [Parameter containing:
tensor([[ 0.7142,  0.1850],
        [ 1.3326, -0.0329]], requires_grad=True), Parameter containing:
tensor([-0.9816,  0.2340], requires_grad=True)], Cost : 0.000
Epoch : 10000, Model : [Parameter containing:
tensor([[ 0.7150,  0.1846],
        [ 1.3329, -0.0330]], requires_grad=True), Parameter containing:
tensor([-0.9828,  0.2336], requires_grad=True)], Cost : 0.000
Epoch : 11000, Model : [Parameter containing:
tensor([[ 0.7154,  0.1843],
        [ 1.3330, -0.0331]], requires_grad=True), Parameter containing:
tensor([-0.9835,  0.2334], requires_grad=True)], Cost : 0.000
Epoch : 12000, Model : [Parameter containing:
tensor([[ 0.7157,  0.1842],
        [ 1.3331, -0.0331]], requires_grad=True), Parameter containing:
tensor([-0.9838,  0.2333], requires_grad=True)], Cost : 0.000
Epoch : 13000, Model : [Parameter containing:
tensor([[ 0.7158,  0.1842],
        [ 1.3331, -0.0331]], requires_grad=True), Parameter containing:
tensor([-0.9840,  0.2332], requires_grad=True)], Cost : 0.000
Epoch : 14000, Model : [Parameter containing:
tensor([[ 0.7158,  0.1842],
        [ 1.3331, -0.0331]], requires_grad=True), Parameter containing:
tensor([-0.9840,  0.2332], requires_grad=True)], Cost : 0.000
Epoch : 15000, Model : [Parameter containing:
tensor([[ 0.7158,  0.1841],
        [ 1.3331, -0.0332]], requires_grad=True), Parameter containing:
tensor([-0.9841,  0.2332], requires_grad=True)], Cost : 0.000
Epoch : 16000, Model : [Parameter containing:
tensor([[ 0.7159,  0.1841],
        [ 1.3331, -0.0331]], requires_grad=True), Parameter containing:
tensor([-0.9841,  0.2332], requires_grad=True)], Cost : 0.000
Epoch : 17000, Model : [Parameter containing:
tensor([[ 0.7159,  0.1841],
        [ 1.3331, -0.0331]], requires_grad=True), Parameter containing:
tensor([-0.9841,  0.2332], requires_grad=True)], Cost : 0.000
Epoch : 18000, Model : [Parameter containing:
tensor([[ 0.7159,  0.1841],
        [ 1.3331, -0.0331]], requires_grad=True), Parameter containing:
tensor([-0.9841,  0.2332], requires_grad=True)], Cost : 0.000
Epoch : 19000, Model : [Parameter containing:
tensor([[ 0.7159,  0.1841],
        [ 1.3331, -0.0331]], requires_grad=True), Parameter containing:
tensor([-0.9841,  0.2332], requires_grad=True)], Cost : 0.000
Epoch : 20000, Model : [Parameter containing:
tensor([[ 0.7159,  0.1841],
        [ 1.3331, -0.0331]], requires_grad=True), Parameter containing:
tensor([-0.9841,  0.2332], requires_grad=True)], Cost : 0.000

Better Jeong

[PyTorch] 다중 선형 회귀

Dataset과 DataLoader

데이터셋(Dataset) 클래스

데이터로더(DataLoader) 클래스

다중 선형 회귀

출력

댓글남기기