alexandremuzio/deep-steg

Global NIPS Paper Implementation Challenge of "Hiding Images in Plain Sight: Deep Steganography"

deep-steg's Issues

Decouple the decoder

Can the decoder be decoupled after training to be used independently (i.e. use the encoder and the decoder separately)?

Running the script gives an error of 'list' object is not callable

the script used to run correctly, when i checked again lately i got this error
TypeError: 'list' object is not callable
on line

ae_loss.append(autoencoder_model.train_on_batch(x=[batch_S, batch_C],
                                                   y=np.concatenate((batch_S, batch_C),axis=3)))

in training section

NB_EPOCHS = 1000
BATCH_SIZE = 32

m = input_S.shape[0]
loss_history = []
for epoch in range(NB_EPOCHS):
    np.random.shuffle(input_S)
    np.random.shuffle(input_C)
    
    t = tqdm(range(0, input_S.shape[0], BATCH_SIZE),mininterval=0)
    ae_loss = []
    rev_loss = []
    for idx in t:
        
        batch_S = input_S[idx:min(idx + BATCH_SIZE, m)]
        batch_C = input_C[idx:min(idx + BATCH_SIZE, m)]
        
        C_prime = encoder_model.predict([batch_S, batch_C])
        
        ae_loss.append(autoencoder_model.train_on_batch(x=[batch_S, batch_C],
                                                   y=np.concatenate((batch_S, batch_C),axis=3)))
        rev_loss.append(reveal_model.train_on_batch(x=C_prime,
                                              y=batch_S))
        
        # Update learning rate
        K.set_value(autoencoder_model.optimizer.lr, lr_schedule(epoch))
        K.set_value(reveal_model.optimizer.lr, lr_schedule(epoch))
        
        t.set_description('Epoch {} | Batch: {:3} of {}. Loss AE {:10.2f} | Loss Rev {:10.2f}'.format(epoch + 1, idx, m, np.mean(ae_loss), np.mean(rev_loss)))
    loss_history.append(np.mean(ae_loss))

im not sure what's wrong with it

Question about using the model files and log files

There are only model files here, maybe share the model calling code?

Besides, when I open the log file, I find input_1, input_2 ... input_42. What are these inputs? What format of data should I feed into the model?

Question about the dataset

I would like to ask that at the end of the paper, "The proposed embeddings described in this paper are not intended for use with lossy image files", but here uses the jpeg image in tiny imagenet. Why choose this data set?