search

lorenlugosch / neural-min-sum-decoding

Neural min-sum decoding in TensorFlow
Other
44 stars 22 forks source link

didn't see loss decreasing for BCH (63, 45) #2

Open liaocs2008 opened 4 years ago

liaocs2008 commented 4 years ago

Hi,

I tried to run the script on BCH (63, 45) using your default setting. However, I didn't find loss decreasing for even 100K iterations. The final BER is very close to original belief propagation. Could you help me fix it?

The output is following:

0 minibatches completed 0.26780248 20 minibatches completed 0.2812021 40 minibatches completed 0.27524856 60 minibatches completed 0.2865098 80 minibatches completed 0.2665105 ....... 99800 minibatches completed 0.26420146 99820 minibatches completed 0.33095878 99840 minibatches completed 0.30447066 99860 minibatches completed 0.2416447 99880 minibatches completed 0.26412797 99900 minibatches completed 0.2476056 99920 minibatches completed 0.27065468 99940 minibatches completed 0.23565353 99960 minibatches completed 0.2468922 99980 minibatches completed 0.27805367 Trained decoder on 100000 minibatches. .... SNR: 1.0 frame count: 106080 bit errors: 640124 BER: 0.09578335607747372 FER: 0.9434106334841629 SNR: 2.0 frame count: 124800 bit errors: 545150 BER: 0.06933633496133496 FER: 0.8014663461538462 SNR: 3.0 frame count: 186960 bit errors: 495371 BER: 0.04205729431981037 FER: 0.5349326059050065 SNR: 4.0 frame count: 399000 bit errors: 483991 BER: 0.019254127381946932 FER: 0.25065413533834585 SNR: 5.0 frame count: 1255680 bit errors: 508151 BER: 0.006423522624306263 FER: 0.07963971712538226 SNR: 6.0 frame count: 5370600 bit errors: 560241 BER: 0.0016558139287443277 FER: 0.018620079693144154

lorenlugosch commented 4 years ago

Hm, unfortunately I'm not sure what your problem might be. What's the exact command you ran?

liaocs2008 commented 4 years ago

The command I used is: python main.py 0 1 6 1 100 10000000000000000 5 codes/BCH_63_45.alist codes/BCH_63_45.gmat laskdjhf 0.5 100000 FNOMS

lorenlugosch commented 4 years ago

Oh, I think the problem is you used L=0.5. That enables the syndrome loss. I found that the syndrome loss can help FER but hurt BER, so maybe that's the problem.

Change the 0.5 in your command to 1.0 and the 100000 to 20000, and you should get the results in the NOMS paper.