ENGLISH | 简体中文
PDE-Net is a feedforward deep network proposed by Zichao Long et al. to learn partial differential equations from data, predict the dynamic characteristics of complex systems accurately and uncover potential PDE models. The basic idea of PDE-Net is to approximate differential operators by learning convolution kernels (filters). Neural networks or other machine learning methods are applied to fit unknown nonlinear responses. Numerical experiments show that the model can identify the observed dynamical equations and predict the dynamical behavior over a relatively long period of time, even in noisy environments. More information can be found in PDE-Net: Learning PDEs from Data.
train.py
script on the command linepython train.py --config_file_path ./configs/pde_net.yaml --device_target Ascend --device_id 0 --mode GRAPH
Among them,
--config_file_path
represents the parameter and path control file, default './configs/pde_net.yaml'
--device_target
represents the type of computing platform used, which can be selected as 'Ascend' or 'GPU', default 'Ascend';
--device_id
represents the calculation card number used, which can be filled in according to the actual situation, default 0;
--mode
represents the running mode, 'GRAPH' indicates the static Graphical model, 'PYNATIVE' indicates the dynamic Graphical model, default 'GRAPH';
You can run training and validation code line by line using both the Chinese version and the English version of Jupyter Notebook.
Parameter | Ascend | GPU |
---|---|---|
Hardware | Ascend 32G | NVIDIA V100 32G |
MindSpore version | >=2.1.0 | >=2.1.0 |
Parameters | 3.5e4 | 3.5e4 |
Train Config | batch_size=16, steps_per_epoch=70, epochs=500 | batch_size=16, steps_per_epoch=70, epochs=500 |
Evaluation Config | batch_size=16 | batch_size=16 |
Optimizer | Adam | Adam |
Train Loss(MSE) | 0.9 | 0.6 |
Evaluation Error(RMSE) | 0.06 | 0.04 |
Speed(ms/step) | 45 | 105 |
gitee id:liulei277
email: liulei2770919@163.com
此处可能存在不合适展示的内容,页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容,可点击提交进行申诉,我们将尽快为您处理。