| 일 | 월 | 화 | 수 | 목 | 금 | 토 |
|---|---|---|---|---|---|---|
| 1 | ||||||
| 2 | 3 | 4 | 5 | 6 | 7 | 8 |
| 9 | 10 | 11 | 12 | 13 | 14 | 15 |
| 16 | 17 | 18 | 19 | 20 | 21 | 22 |
| 23 | 24 | 25 | 26 | 27 | 28 | 29 |
| 30 | 31 |
- BoundaryCondition
- Higher-order Derivatives
- PINN
- 이메일마케팅
- Efficient Training
- PINNs
- LossFunctions
- completenessaxiom
- Adaptive Optimization
- infimum
- Multiscale Problems
- Dynamic Weighting
- Stiff PDEs
- Adapitveweighting
- 소셜미디어마케팅
- VS-PINN
- densityofrationals
- Computational Efficiency
- GradientConstraint
- Binary Tree
- archimedeanprinciple
- Loss Landscape Flattening
- SG-PINN
- BoundaryConditionLearning
- Gradient Imbalance
- GaussianProcess
- Variable Scaling
- NTK
- supremum
- DiffusionEquation
- Today
- Total
목록전체 글 (40)
Mathematics & Computation
model = PINN([2, 50, 50, 50, 1]).to(device)step = 20a1, a2 = 1,1 LAMBDA_BC = 1.0LAMBDA_PDE = 1.0LAMBDA_IC = 5.0 # anchor를 쓰는 경우 # 손실 총 스케일(항상 고정): 예시 60.0SUM_W = 60.0 1.K_SCHEDULE = [ (0, 1.0), (5000, 3.0), (10000, 6.0), (15000, 10.0),] L2 abs: 1.566e-02 | L2 rel: 1.566e-02 H1 abs: 1.632e-01 | H1 rel: 3.531e-02 L_inf abs:2.810e-02 2.K_SCHEDULE = [ (0, 1.0), (10000, 10.0),] L2 abs: 2.551e-02 |..
epoch = 20000model = PINN([3, 50, 50, 50,50, 1]).to(device)a1, a2 = 1, 1 k=1k=2k=3k=4k=5k=6step = 10Relative L2 error4.8791e-026.7382e-02 1.0356e-01 9.9476e-01 9.9869e-01 4.8525e-01Relative H1 error 5.2810e-024.2188e-029.2454e-029.9470e-01 9.9657e-01 5.2326e-01Time 2m2m2m3m3m3mstep = 20Relative L2 error1.0876e-01 4.5514e-02 2.4261e-01 9.0612e-01 6.0647e-01Relative H1 error 1.1269e-014.0538e..
#Try1: Max error: 2.1042e+00 Relative Error: 1.4976e+00 #Try2: #Try3:L-BGFS추가파일
model = PINN([2, 50,50,50, 1]).to(device)step = 50a1,a2 = 1,1k = pi * torch.sqrt(torch.tensor(a1**2 + a2**2, dtype=torch.float32, device=device)optimizer lr = 1e-3threshold loss = 1e-9 #Try1:epoch = 40000max error: 1.0238e-01 Relative Error: 8.9404e-02 #Try2:epoch 40000L-BFGS사용 Max error: 1.8156e-03 Relative Error: 7.2430e-04 #Try3:epoch = 30000 Max error: 1.2876e-02 Relative Error: 7.9915e-0..
class PINN(nn.Module): def __init__(self, in_dim=16, hidden=64, out_dim=1): super().__init__() self.net = nn.Sequential( nn.Linear(in_dim, hidden), nn.Tanh(), nn.Linear(hidden, hidden), nn.Tanh(), nn.Linear(hidden, out_dim) SA-PINN + Fourier #Try 1:collocation point = 100bounday point = 40 epochs = 10000 max error: 4.2092e-01 Relative Error: 2.6617e-014m (시간은 정확하지 않음) => Prediction은 Exact에 비해 파..
기본 설정a1 =2 , a2=2Model = [2,50,50,50,1]Lambda 1 = 100, Lambda 2 = 1a1,a2 = 2Epochstep4105010030050010000max abs Error3.99403e+002.0125e+001.9990e+00 2.0042e+001.99772+002.0072e+00relative Error9.3241e-01 4.9915e-014.9974e-015.0022e-014.9919e-015.0067e-01Time1m 1m1m1m9m 30000max abs Error4.0350e+00 2.0064e+00 1.9998e+002.0036e+002.0041e+00relative Error9.1490e-015.0030e-014.9995e-014.9990e-015.00..
1) SG-PINN과 NTK 기반 PINN은 PINN의 구조자체는 유지하면서, 학습과정 (loss 구성, loss weight 조절)을 수정하여 결과를 개선 2) SG-PINN : Loss weight를 Gradient Norm으로 soft하게 조절기존 PINN은 람다 값을 사용자가 사전에 정하고, 시 비교 포인트 NTK 기반 PINN SG-PINN이론 정당성매우 강함 (NTK 분석 기반)약함 (경험적 효과 위주)구현 난이도복잡 (NTK 행렬 계산)간단 (gradient norm만 추적)계산량크다매우 적다λ 조정 방식NTK 해석 또는 정적 λ∇Loss 기반 자동 동적 조정적합한 경우이론 분석, 고주파 학습 중요할 때다양한 문제에 실용적 적용 시 구분SG-PINNNTK PINN목표경계 조건에서의 정확도 ..