veya2ztn / Metasurface

MetaSurface

This is open source code for our metasurface machine learning series

1. Metasurface Machine Learning performance Benchmark
2. SUTDPRT Dataset and Network Architect Search
3. Symmetry Inspired Metasurface Machine Learning

---

pip install tensorboardX
pip install scipy
pip install matplotlib
pip install optuna
pip install PyWavelets
pip install pandas
pip install tensorboard

git clone https://github.com/veya2ztn/mltool.git
git clone https://github.com/veya2ztn/SUTD_PRCM_dataset.git

echo "{\"DATAROOT\": \"[the path for your data]",\"SAVEROOT\": \"[the path for your checkpoint]"}" > .DATARoot.json

---

1. run python makeproject.py to generate task files under projects/undo

2. run python multi_task_train.py to start training.

   use -m or --mode assign train mode

   1. The default mode will train task in projects/undo one by one.
   2. The test mode will run only 1 epoch and exit, meanwhile it will give the maximum batch size for current GPU. ( may fail after PyTorch 1.10)
   3. The parallel flag will active data parallel. Notice the official torch.optim.Adam optimizer is not compatible. Recommend use torch.optim.SGD.
   4. The multiTask flag should work with --gpu NUM. It will auto move task file to project/muliti_set{gpu} and run task on its individual GPU part.
   5. The assign flag should work with --path , which make script run assigned task.

   python makeproject.py
   python multi_task_train.py
   

---

Normal machine learning train

1. add model under ./model
2. now model avaible:
   • Modified 2D Resnet18/34/50/101 with different tail layer
   • Modified Resnet18/34/50/101 Complex number version (Half/Full)
   • Modified Resnet+NLP
   • Modified DenseNet/SqueezeNet
   • 1D Resnet for directly inverse problem
   • Deep Multilayer Perceptron(MLP)
   • DARTS model (with geotype input)
3. Task-json-templete see makeproject.py
4. The front script is train_base.py

Network Architecture Search (DARTS-GAEA)

1. Task-json-templete see makeNASproject.py
2. The code is rewrite and unit weight now is split
   • call architect_darts.ArchitectDARTS for share weight evolution.
   • call architect_msdarts.ArchitectMSDARTS for full weight evolution. (which may allocate much more memory and may not stable [under my test] )
3. Notice the drop phase only work on train.
4. The front script is darts_search.py

Generative Adversarial Network

1. There are two way to train the GAN model

   • original Deep Convolutional GAN (DCGAN) at ./GANproject/DCGAN*.py
   • Wasserstein GAN (WGAN) at ./GANproject/GAN*.py . However, Wasserstein distance in this scenario would make train unstable.

2. There are DCGANTRAIN and DCGANCURVE two part, link to pattern confidence and pattern-curve map accordingly.

3. The whole GAN flow for practice implement is