veya2ztn / SUTD_PRCM_dataset

This is the SUTD PRT dataset for Metasurface Machine Learning

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

SUTD_PRT_dataset

This is the SUTD PRT dataset for Metasurface Machine Learning.

The full dataset is a 8G file and can be download here:

Introduction

This dataset is consist of 30k+60k+60k+110k 「image($16\times16)$」- 「complex value curve $2\times1001$」pair.

have_a_look_for_all_image_dataset.curvehave_a_look_for_all_image_dataset.image

Divide along the pattern symmetry, we have four main image classes:

Name Volume Freedom Description
PLG 30000 $2^{152}$ The unit assembles as a polygon image, which must be a connected topo
PLR 60000 $2^{???}$ The unit assembles as a polygon image, which another polygon inside
PTN 60000 $2^{102}$ Combination of square, cross, triangle, U-shape, H-shape
RDN 120000 $2^{256}$ All the units are randomly set 0 or 1

image-20210507014350020

Usage

Completely use the dataset

  1. download the dataset into dataset_path and unzip.

    The dataset structure tree is like:

    SUTDPRTDATASET
    ├── PLGDATASET
    │   ├── full_data_list
    │   │   ├── Data_001
    │   │   │   ├── Integrate_curve_1.npy
    │   │   │   └── Integrate_image_1.npy
    │   │   ├── Data_002
    │   │   │   ├── Integrate_curve_2.npy
    │   │   │   └── Integrate_image_2.npy
    │   ........
    │   ........
    ├── RDNDATASET
    │   ├── full_data_list
    │   │   ├── Data_001
    │   │   │   ├── Integrate_curve_1.npy
    │   │   │   └── Integrate_image_1.npy
    │   │   ├── Data_002
    │   │   │   ├── Integrate_curve_2.npy
    │   │   │   └── Integrate_image_2.npy
    │  ........
    │  ........
  2. Use ln -s real_data_path load_data_path put the data what you want to use into a new file.

    For example,

    PLGDATASET
    ├── full_data_list
    │     ........
    │     ........
    ├── train_data_list                            
    │   ├── Data_001 -> ../full_data_list/Data_001 
    │   └── Data_002 -> ../full_data_list/Data_002 
    └── valid_data_list                            
        ├── Data_001 -> ../full_data_list/Data_003 
        └── Data_002 -> ../full_data_list/Data_004 
  3. use the SMSDatasetN or SMSDatasetC module load the dataset path. This module will automatedly load the data in the assigned path. For example.

    dataset_train = SMSDatasetN("data/PLGDATASET/train_data_list",None)
    dataset_train = SMSDatasetN("data/PLGDATASET/valid_data_list",None)
  4. More option please see the class SMSDataset in dataset_module.py

Fast use this dataset.

We provide fast train/test dataset script.

  • B1NE class: The fast task for
    • Transmission curve norm: so the value now is real
    • Precision is 0.001

For example,

from dataset import get_FAST_B1NE_dataset
dataset_path = "data"
dataset_class= "RDN" # can choose ["RDN","PTN","PLR","PLG","PLG250"]
curve_feature=  128  # assign the reduced dimenstion, fast mode only support [32,128,1001]
# for simple norm-curve  
dataset_train,dataset_valid = get_FAST_B1NE_dataset(dataset_path,
                                                    dataset=dataset_class,FeatureNum=curve_feature)

# for simple binary classifation problem: MPJ  
dataset_train,dataset_valid = get_FAST_B1NE_dataset(dataset_path,
                                         dataset=dataset_class,FeatureNum=curve_feature,                                                              type_predicted="onehot",target_predicted="balance_leftorright")

# for peak parameter prediction  
dataset_train,dataset_valid = get_FAST_B1NE_dataset(dataset_path,
                                         dataset=dataset_class,FeatureNum=curve_feature,                                                              type_predicted="combination",target_predicted="peakparameters")

If you use this dataset in your research, please use below citation:

@article{zhang2022symmetry,
  title={Symmetry Enhanced Network Architecture Search for Complex Metasurface Design},
  author={Zhang, Tianning and Kee, Chun Yun and Ang, Yee Sin and Li, Erping and Ang, Lay Kee},
  journal={IEEE Access},
  volume={10},
  pages={73533--73547},
  year={2022},
  publisher={IEEE}
}

About

This is the SUTD PRT dataset for Metasurface Machine Learning

License:MIT License


Languages

Language:Python 100.0%