01_PINN_pytorch.py

import sys
from sklearn.model_selection import train_test_split
sys.path.append(".")
import numpy as np
import torch
from torch.autograd import grad
from network import DNN
from scipy.io import loadmat
import pandas as pd
import torch
import torch.nn as nn
import numpy as np
import os

current_path = os.getcwd()

torch.backends.cuda.matmul.allow_tf32 = (
    False  # This is for Nvidia Ampere GPU Architechture
)
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
torch.manual_seed(1234)
np.random.seed(1234)


# current_path = os.path.dirname(os.path.abspath(__file__))
# print(current_path)
"""
Burgers Eqn.
f = u_t + lambda_1 * u * u_x - lambda_2 * u_xx = 0, x ~ [-1, 1], t ~ [0, 1]
lambda_1 = 1
lambda_2 = 0.01/pi = 0.0031831
f = w*u_t+w*u_x+
"""

N_u = 2000

data = loadmat(r"D:\02_github\PINNs-torch\Burgers\burgers_shock.mat")

x = data["x"]
t = data["t"]
u = data["usol"].T

ub = np.array([x.max(), t.max()])    ###上下限制
lb = np.array([x.min(), t.min()])

# Clean Data Preparation
x_, t_ = np.meshgrid(x, t)
x_ = x_.reshape(-1, 1)
t_ = t_.reshape(-1, 1)
u_ = u.reshape(-1, 1)

rand_idx = np.random.choice(len(u_), N_u, replace=False)

x = torch.tensor(x_[rand_idx], dtype=torch.float32).to(device)
t = torch.tensor(t_[rand_idx], dtype=torch.float32).to(device)
xt = torch.cat((x, t), dim=1)
u = torch.tensor(u_[rand_idx], dtype=torch.float32).to(device)

# 1% Noisy Data Preparation
noise = 0.01
noisy_u = u_ + noise * np.std(u_) * np.random.randn(*u_.shape)
noisy_u = torch.tensor(noisy_u[rand_idx], dtype=torch.float32).to(device)

print(u_.shape,u.shape)


class PINN:
    def __init__(self, u):
        self.u = u
        self.lambda_1 = torch.tensor([0.0], requires_grad=True).to(device)
        self.lambda_2 = torch.tensor([-6.0], requires_grad=True).to(device)
        self.lambda_1 = torch.nn.Parameter(self.lambda_1)
        self.lambda_2 = torch.nn.Parameter(self.lambda_2)
        self.net = DNN(dim_in=2, dim_out=1, n_layer=7, n_node=20, ub=ub, lb=lb,).to(
            device
        )
        self.net.register_parameter("lambda_1", self.lambda_1)
        self.net.register_parameter("lambda_2", self.lambda_2)

        self.optimizer = torch.optim.LBFGS(
            self.net.parameters(),
            lr=1.0,
            max_iter=50000,
            max_eval=50000,
            history_size=50,
            tolerance_grad=1e-5,
            tolerance_change=1.0 * np.finfo(float).eps,
            line_search_fn="strong_wolfe",
        )
        self.iter = 0

    def f(self, xt):
        # lambda_1 = self.lambda_1
        # lambda_2 = torch.exp(self.lambda_2)
        xt = xt.clone()
        xt.requires_grad = True

        u = self.net(xt)

        f = grad(u.sum(), xt, create_graph=True)[0]
        # u_x = u_xt[:, 0:1]
        # u_t = u_xt[:, 1:2]
        #
        # u_xx = grad(u_x.sum(), xt, create_graph=True)[0][:, 0:1]
        #
        # f = u_t + lambda_1 * u * u_x - lambda_2 * u_xx
        # Apply output bounds
        f = torch.clamp(f, 0, 365)
        return f

    def closure(self):
        self.optimizer.zero_grad()

        u_pred = self.net(xt)
        f_pred = self.f(xt)

        mse_u = torch.mean(torch.square(u_pred - self.u))
        mse_f = torch.mean(torch.square(f_pred))

        loss = mse_u + mse_f
        loss.backward()

        self.iter += 1
        print(
            f"\r{self.iter} loss : {loss.item():.3e} l1 : {self.lambda_1.item():.5f}, l2 : {torch.exp(self.lambda_2).item():.5f}",
            end="",
        )
        if self.iter % 500 == 0:
            print("")
        return loss


def calcError(pinn):
    u_pred = pinn.net(torch.hstack((x, t)))
    u_pred = u_pred.detach().cpu().numpy()
    u_ = u.detach().cpu().numpy()
    error_u = np.linalg.norm(u_ - u_pred, 2) / np.linalg.norm(u_, 2)
    lambda1 = pinn.lambda_1.detach().cpu().item()
    lambda2 = np.exp(pinn.lambda_2.detach().cpu().item())
    error_lambda1 = np.abs(lambda1 - 1.0) * 100
    error_lambda2 = np.abs(lambda2 - 0.01 / np.pi) * 100
    print(
        f"\nError u  : {error_u:.5e}",
        f"\nError l1 : {error_lambda1:.5f}%",
        f"\nError l2 : {error_lambda2:.5f}%",
    )
    return (error_u, error_lambda1, error_lambda2)


if __name__ == "__main__":
    pinn = PINN(u)
    pinn.optimizer.step(pinn.closure)
    torch.save(pinn.net.state_dict(), r"D:\02_github\PINNs-torch\Burgers\Identification\weight_clean3.pt")
    pinn.net.load_state_dict(torch.load(r"D:\02_github\PINNs-torch\Burgers\Identification\weight_clean3.pt"))
    calcError(pinn)
    #
    # pinn = PINN(noisy_u)
    # pinn.optimizer.step(pinn.closure)
    # torch.save(pinn.net.state_dict(), r"D:\02_github\PINNs-torch\Burgers\Identification\weight_noisy2.pt")
    # calcError(pinn)
    #