solaaa
commented
6 years ago In your result, what is the mean of 500 in filename resnet26_500k.txt ? epochs? By the way,your model can't classify QAM16 and QAM64, the acc I test is 50%, I have been puzzled for many months, could you give me some advice? Thanks
hi, Ostnie pretty sorry for not checking for long time. here i will show some info. about what i think first, this is a developing research by myself, thus i'm sorry for making u much doubts:( here, i will tell u some details: I train CNNs (like ResNet) to classify MPSK(2,4,8), MFSK(2,4,8), MQAM(16,64,128 or 16, 128), and GFSK. All signals are under different channels (Rayleigh+AWGN). the simple CNN with 2 or 3 layers, performs well on those signals without channels (also well on high SNR), but not satisfactory on Rayleigh channel. so i try to find different ways to make this better. the ways include: A. deeper CNN, B. ensemble CNN, C. denoise signal first and than implement CNN, D. use equalization method to remove channel interfere. second, the dataset is simulated by MATLAB 2016b, here, i test different size of dataset including 50k, 100k, 200k, 500k, 1M. the '500K' means i use 500,000 samples to train the model, each sample is 2 1024 size (I/Q). Keras does not support complex64, so i divide 1 1024(complex) into 2 1024(float). with the dataset size bigger, the ACC is increasing. when it is bigger than 500k, the ACC stop increasing. so i choose 500k to conduct next experiment(like change size of each sample, e.g. 2 128, 2 512, 2 2048 etc.). third, the result shows that QAMs are hard to classified as u mentioned, but not always 50%, in my experiment, i use 500k 2* 1024 signals with 9 types(bpsk, qpsk, 8psk, 2fsk, 4fsk, 8fsk, 16qam, 128qam, gfsk) and a simple ResNet26 to implement the experiment. the ACC of 16QAM and 128QAM are 70%~80% or so in high SNR (Rayleigh channel). anyway, it's not a satisfactory result :)
i see u r from Beijing, u can use CN to chart with me, and if necessary, my qq:410671760 :)
In your result, what is the mean of 500 in filename resnet26_500k.txt ? epochs? By the way,your model can't classify QAM16 and QAM64, the acc I test is 50%, I have been puzzled for many months, could you give me some advice? Thanks