ENGLISH | 简体中文

Contents

• DeepBSDE Description
• Environment Requirements
• Quick Start
• Script Description
  • Script and Sample Code
  • Script Parameters
  • Training Process
  • Evaluation Process
• Model Description
  • Evaluation Performance
  • Inference Performance
• Description of Random Situation

DeepBSDE Description

DeepBSDE is a power of deep neural networks by developing a strategy for solving a large class of high-dimensional nonlinear PDEs using deep learning. The class of PDEs that we deal with is (nonlinear) parabolic PDEs.

paper: Han J , Arnulf J , Weinan E . Solving high-dimensional partial differential equations using deep learning[J]. Proceedings of the National Academy of Sciences, 2018:201718942-.

HJB equation

Hamilton–Jacobi–Bellman Equation which is the term curse of dimensionality was first used explicitly by Richard Bellman in the context of dynamic programming, which has now become the cornerstone in many areas such as economics, behavioral science, computer science, and even biology, where intelligent decision making is the main issue.

Environment Requirements

• Hardware(GPU)
  • Prepare hardware environment with GPU processor.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Quick Start

After installing MindSpore via the official website, you can start training and evaluation as follows:

• running on GPU

Default:

python train.py

Full command is as follows:

python train.py \
    --save_ckpt false \
    --save_ckpt_path ./checkpoints \
    --load_ckpt_path ./checkpoints/deepbsde_HJBLQ_end.ckpt \
    --log_path ./logs \
    --print_interval 100 \
    --total_time 1.0 \
    --dim 100 \
    --num_time_interval 20 \
    --y_init_range 0 1 \
    --num_hiddens 110 110 \
    --lr_values 0.01 0.01 \
    --lr_boundaries 1000 \
    --num_iterations 2000 \
    --batch_size 64 \
    --valid_size 256 \
    --sink_size 100 \
    --file_format MINDIR \
    --epochs 1001 \
    --amp_level O0 \
    --device_id 0 \
    --mode 0

Script Description

Script and Sample Code

.
├── config
│     └── HJBLQ_config.yaml    # default config for HJB equation.
├── src
│     ├── config.py            # config parse script.
│     ├── equation.py          # equation definition and dataset helper.
│     ├── eval_utils.py        # evaluation callback and evaluation utils.
│     └── net.py               # DeepBSDE network structure.
├── eval.py                    # evaluation API entry.
├── export.py                  # export models API entry.
├── README_CN.md
├── README.md
├── requirements.txt           # requirements of third party package.
└── train.py                   # train API entry.

Script Parameters

Parameters for both training and evaluation can be set in config.yaml

• config for HBJ

Training Process

• Running on GPU

  python train.py

• The python command above will run in the background, you can view the results through the file train.log。

  The loss value can be achieved as follows:

  epoch: 1 step: 100, loss is 245.3738
  epoch time: 26883.370 ms, per step time: 268.834 ms
  total step: 100, eval loss: 1179.300, Y0: 1.400, elapsed time: 34
  epcoh: 2 step: 100, loss is 149.6593
  epoch time: 3184.401 ms, per step time: 32.877 ms
  total step: 200, eval loss: 659.457, Y0: 1.693, elapsed time: 37
  ...

  After training, you'll get the last checkpoint file under the folder log_dir in config.

Evaluation Process

• Evaluation on GPU

  Before running the command below, please check the checkpoint path used for evaluation. Such as ./log/deepbsde_HJBLQ_end.ckpt

  python eval.py --config_path=./config/HJBLQ_config.yaml > eval.log  2>&1 &

  The above python command will run in the background. You can view the results through the file "eval.log". The error of evaluation is as follows:

  eval loss: 5.146923065185527, Y0: 4.59813117980957

Model Description

Evaluation Performance

Inference Performance

Description of Random Situation

We use random sampling in equation.py, which can be set seed to fixed randomness.