search

chunbaobao / Deep-JSCC-PyTorch

A implement of Deep JSCC for wireless image transmission by PyTorch
34 stars 6 forks source link

About the results #7

Open homesomele opened 1 month ago

homesomele commented 1 month ago

Hello, thank u so much for providing the codes of jscc pytorch, I found that the results are worse than the figure showed in the paper, can u restruct the same results as the paper?

chunbaobao commented 1 month ago

I found the same problem you mentioned. The models' performance is approximately 5dB worse than reported in the paper, as shown in the README.md results. I believe this discrepancy is due to system differences:

  1. The complex communication system has twice as many parameters as the real-valued communication system implemented in the repository.
  2. The noise power in the complex communication system is half that of the real-valued communication system.
homesomele commented 1 month ago

I found that after chaging the decoder transconv structure, the results are similar to the results showed in the paper _GF%DI60(}336}CDO$HOSFK

chunbaobao commented 1 month ago

Thanks for sharing. I noticed that you used the output_padding parameter with kernel_size=5 in tconv4-5. However, I encountered a mismatch error when running your code pasted above. Could you please provide the whole modified code? Thanks in advance.