search

TianLin0509 / DNN_detection_via_keras

This is the simplest implementation of Power of Deep Learning for Channel Estimation and Signal Detection in OFDM Systems using keras.
165 stars 67 forks source link

training label issue #3

Open zhgjia opened 5 years ago

zhgjia commented 5 years ago

why is the size of label 16 instead of payloadBits_per_OFDM（128）？Is there a difference between the two?

TianLin0509 commented 5 years ago

sorry, I'm not understanding your problem

zhgjia commented 5 years ago

What I mean is that there are 128 Numbers of a OFDM symbol (64-IFFT) with real and image parts. Why is the predicted label only 16 of them? Will this affect the simulation results of bit error rate?

------------------ 原始邮件 ------------------ 发件人: "Lin Tian"notifications@github.com; 发送时间: 2019年7月6日(星期六) 中午12:51 收件人: "TianLin0509/DNN_detection_via_keras"DNN_detection_via_keras@noreply.github.com; 抄送: "张佳"2673224202@qq.com;"Author"author@noreply.github.com; 主题: Re: [TianLin0509/DNN_detection_via_keras] training label issue (#3)

sorry, I'm not understanding your problem

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

TianLin0509 commented 5 years ago

This is because that the author said: he used totally 8 networks to predict the 128bits. And therefore, every network only need to predict 16bits. Please kindly refer to the paper. Since the 8 network intutively have similar performance, in the program, we can only consider one network and predict 16bits. You can choose to predict 128bits, it will work, but results in some performance loss, as expected.

zhgjia commented 5 years ago

Ok, I get it. Thank you very much!

------------------ 原始邮件 ------------------ 发件人: "Lin Tian"notifications@github.com; 发送时间: 2019年7月6日(星期六) 下午2:22 收件人: "TianLin0509/DNN_detection_via_keras"DNN_detection_via_keras@noreply.github.com; 抄送: "张佳"2673224202@qq.com;"Author"author@noreply.github.com; 主题: Re: [TianLin0509/DNN_detection_via_keras] training label issue (#3)

This is because that the author said: he used totally 8 networks to predict the 128bits. And therefore, every network only need to predict 16bits. Please kindly refer to the paper. Since the 8 network intutively have similar performance, in the program, we can only consider one network and predict 16bits. You can choose to predict 128bits, it will work, but results in some performance loss, as expected.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

BassantTolba1234 commented 4 years ago

Dear Sir, please I need to know which type of modulation scheme is used here QPSK as mentioned in the paper or QAM as written in the code ? what should I do to make it 16 QAM ? the second question , can you kindly share the code required to get the same results figures ? thanks in advance

haide666 commented 4 years ago

Since some subcarriers locate in the channel with low SNR, I might not agree with the point that every network predicting 16bits had the similar performance. For example, in the low pass channel with AWGN, last subcarriers will have lower SNRs than those of the first subcarriers. So the bits modulated in the last subcarriers will be more difficult to be predicted.

BassantTolba1234 commented 4 years ago

Dear Sir, I really appreciate your hard work..and please I have a question, in main file line 48, why do you make (model.evaluate) on function called (validation_gen) which exactly contains the generation of training datasets not the test datasets??..

why did not you apply evaluation the model on testing datasets you provided , instead of training datasets ? I'm waiting for your reply.. thanks in advance.

BassantTolba1234 commented 4 years ago

Dear sir, please what is the value of SNR (Signal Noise to ratio used in training the model ) ??