search

netabecker / Stegastamp_pytorch_version

12 stars 4 forks source link

Failed to decode #14

Open swyuan521 opened 3 months ago

swyuan521 commented 3 months ago

Hello,

  1. the model I trained has a valid residual image generated by the encoder, but it cannot be successfully decoded. As follows:

Im1552_hidden. png Failed to decode

2.What are the differences between different "borders"? Will it affect the decoder?

Have you ever encountered this situation? Thank you.

netabecker commented 2 months ago

Hi! I have. Try setting a seed while training (I explained a bit more about it in another answer). Make sure to follow the secret loss and the image loss while training, sometimes the model reaches overfitting while training and then the decoder fails.

Alex17swim commented 2 months ago

Hi! I have. Try setting a seed while training (I explained a bit more about it in another answer). Make sure to follow the secret loss and the image loss while training, sometimes the model reaches overfitting while training and then the decoder fails.

Dear netabecker, thanks for your contribution to the code! I have a similar issue where the bit accuracy is always around 0.5 and the str accuracy is 0.0. Can we get in touch via email, WeChat, Twitter, etc., please?

swyuan521 commented 2 months ago

Hi! I have. Try setting a seed while training (I explained a bit more about it in another answer). Make sure to follow the secret loss and the image loss while training, sometimes the model reaches overfitting while training and then the decoder fails.

Thank you very much, but the residual image I trained on has speckled blur in the middle, Should lpips_loss_scale and G_loss_scale be set to 0? Which one does Borders choose to make the residual effect better?