I use your code to train the model successfully. But the MSE(mean square error) is so high. For example, when SNR=20dB and compression ratio= 0.09, the MSE is about 0.009. But in paper named 'Deep Joint Source-Channel Coding for
Wireless Image Transmission", the MSE is about 0.001 when SNR=20dB and compression ratio= 0.09, as fig3 shows in paper. That's a big gap. I want to know that when you train the model, could you reach the performance the same as paper shows?

I'm a beginner so can you tell me how to peform the program?

I'm studying paper, so found your code.
I have a question, is normalizationnoise layer used as a non-trainable layer?
Or trained with other NN?