PINN学习记录（3.3）

最新推荐文章于 2024-08-11 20:49:18 发布

每天进步一点点吧

最新推荐文章于 2024-08-11 20:49:18 发布

阅读量850

点赞数

文章标签： pycharm python ide

本文链接：https://blog.csdn.net/weixin_45805559/article/details/121670980

版权

本文记录了PINN学习过程中的PhyCRNet模型构建，详细介绍了Loss部分，包括Conv2dDerivative、Conv1dDerivative的实现，以及loss_generator的计算过程。同时，提到了数据处理流程，数据经过encoder、convlstm、pixelshuffle和outputlayer的处理。

摘要由CSDN通过智能技术生成

PhyCRNet模型构建完成

$\rightarrow Train Loss$

part5 Loss

1、Conv2dDerivative

class Conv2dDerivative(nn.Module):
    def __init__(self, DerFilter, resol, kernel_size=3, name=''):
        super(Conv2dDerivative, self).__init__()

        self.resol = resol  # constant in the finite difference
        self.name = name
        self.input_channels = 1
        self.output_channels = 1
        self.kernel_size = kernel_size

        self.padding = int((kernel_size - 1) / 2)
        self.filter = nn.Conv2d(self.input_channels, self.output_channels, self.kernel_size, 
            1, padding=0, bias=False)

        # Fixed gradient operator
        self.filter.weight = nn.Parameter(torch.FloatTensor(DerFilter), requires_grad=False)  

    def forward(self, input):
        derivative = self.filter(input)
        return derivative / self.resol

filter
filter=conv2d(1, 1, 3x3, 1, padding=0, bias=False)

通过给filter传参，得到derivative
定义了filter的权重，weight
nn.Parameter()表示对weight会进行优化

derivative
derivative = self.filter(input)
输出Conv2d网络的output
resol
有限差分中的常数

laplace_operator：(dx**2)
dx_operator：(dx*1)
dy_operator：(dy*1)

最后Conv2dDerivative返回 $\frac {derivative}{resol}$
2、Conv1dDerivative

class Conv1dDerivative(nn.Module):
    def __init__(self, DerFilter, resol, kernel_size=3, name=''):
        super(Conv1dDerivative, self).__init__()

        self.resol = resol  # $\delta$*constant in the finite difference
        self.name = name
        self.input_channels = 1
        self.output_channels = 1
        self.kernel_size = kernel_size

        self.padding = int((kernel_size - 1) / 2)
        self.filter = nn.Conv1d(self.input_channels, self.output_channels, self.kernel_size, 
            1, padding