bilepeng/risnet

RISnet

This repository is the source code and data for the paper

B. Peng, F. Siegismund-Poschmann, E. Jorswieck, "RISnet: a dedicated scalable neural network architecture for optimization of reconfigurable intelligent surfaces", International ITG 26th Workshop on Smart Antennas and 13th Conference on Systems, Communications, and Coding, Braunschweig, 2023.

To train the neural network, run train.py with the following arguments:

• --tsnr: the transmit SNR with default value 1e11.
• --lr: the learning rate with default value 8e-4.
• --ris_shape: the RIS shape with default value 32, 32.
• --weights: the user weights in the weighted sum-rate with default value 0.25, 0.25, 0.25, 0.25.
• --record: True if you want to save the tensorboard log and trained models in a folder named after date and time of the beginning of training, False otherwise.
• --device: cpu or cuda.

To test the saved neural network, run test.py with the path to the saved model in line 46.

risnet's Issues

my results is not same your results in your paper

when I retrain the model risnet , my results is higher your results in your paper, and gain good results at 10 iteration. I cannot explain my results.

Reason for choosing Convolution 1d layer

Why you choose the Convolution 1d layer with input feature channel(16,1024) and output(1,1024)?
What is your relevant formula?.
I look forward to hearing from you soon. Thank you so much