Some questions on this interesting project

[planewave]

Hello, I am working on physical lay signal processing and new to deep learning and Python. I have read the paper and found your code here. I managed to run several tests and have some questions.

• Your code is an implementation of the paper's Section III-A autoencoder for end to end communication system (figure 2), not exactly the Section III-C radio transformer networks (figure 8). Are you going to realize that in the future?
• I spent some time learn to use PyTorch and then realized that half of the code is using TorchNet. Why do you use TorchNet? Can we realize it with PyTorch only? I only know how to use Matlab, running python code is really painful to me...
• You mentioned it can achieve 100% within a span of ~30 epochs. However, I seldom achieved this result, in fact, most of the time, it went to 50% or 75% and stoped improving. Any suggestions?
• I am interested in the encoder's output. I tried to plot it after each epoch, but because I am new to Python, I can only draw the final result. I was hoping to have something like figure 4 (a), but never succeed, even when the accuracy is 100%.

[ghost]

@planewave is the implementation of normalisation layer in this repository correct? can you please explain me if it's wrong or what needs to be done? thanks

[planewave]

@vsag sorry, not sure about that. it seems not important (?) so i basically ignored it...

[ghost]

@planewave Even I am not sure, In the paper they say they use it to make the transmitter follow Gaussian channel properties. There's another repository which has a different implementation of that layer basically he multiplies the l2 norm with square root of n, in this repo he multiplies it with square of n.