This is the simulation code for the paper "Beamforming Design for Large-scale Antenna Arrays with the Help of Deep Learning Approach" and we highly respect reproducible research.

Due to the limitation of file size, we only upload the executable .py files in the github repository:

• BFNN_Model.py : Architecture of the proposed BFNN model.
• load_samples.py : used to load samples of train set or test set.
• main_imperfect_CSI.py : simulate with imperfect CSI
• main_perfect_CSI.py : simulate with perfect CSI

After download the samples (see following), you can directly run the main_imperfect_CSI.py or main_perfect_CSI.py and then their are three option steps: First, you will see

Please type 0 or 1 in the python console to choose Mode: 
 0: train  1: test 

As you can see, type 0 for training and 1 for testing.

If you choose the train mode, you need to further provide the path of samples. You can easily copy the absolute path of the package (Noticed is the path of the package), then the program will load the sample data in the package. For example, type C:\OneDrive\BFNN\sample_set in the python console.

If you choose the test mode, you need to further provide the path of the trained model. For example, type C:\OneDrive\BFNN\\trained_models\trained_20dB.h5 in the python console.

We have shared our generated samples and trained models by Google Driver. Please kindly refer to https://drive.google.com/drive/folders/1nSk9TftoCMA5iRUqC9GSK5g5a67Q8FRG?usp=sharing After download the files, you can straightforwardly run main_erfect_CSI.py or main_imperfect_CSI.py. Just easily copy the packages path in the python console, the program can load the data for training or testing. Similar operation can help you load our trained model, which is corresponding to the results in our paper. In addition, we also provide the matlab codes on onedrive so you can generate samples by yourself. If you still need more information, you can contact me via e-mail.

My email is lint17@fudan.edu.cn and we welcome discussions. As I'm new to programing, I have tried my best to add more comments in .py files for ease of comprehension. If you have any questions, you can contact me by email. Sorry for my English, hopefully this work can be helpful to you.

Since the loss function is associated with perfect CSI. i.e., real(h) and imag(h), to utilize the automatically allocation of batchs in keras, in the code implementation we regard the perfect CSI as the label so that can be easily put into the loss function and compute the loss value. Noticed that it is just for code implementation simplication, the perfect CSI is completely not the real label. Besides, the lambda layer that transfer the output of Dense Unit into a complex-valued vector, is also realized in the loss function. It is equivalent to the discribed BFNN in our paper, but since complex operation is not supported well, so we use this trick to implement it simply.