Set "d_container.trainable = True" after training all_model

[bis-carbon]

Thank you for the great work. Don't you think d_container.trainable should be set to True after training all_model. Something like this:

d_container.trainable = True 
all_model.compile(loss=['mse', 'binary_crossentropy'], loss_weights=[1.0, alpha], optimizer=optimizer) 

[neka-nat]

Thank you for your opinion. Probably I think that trainable is correct with false.

I attached the pseudo code of the paper. The place where all_model is used is line 10. Here the discriminator is fixed and learning only the completion network. The discriminator update is done on line 8, and it uses d_model in my code.

[bis-carbon]

I agree that d_container.trainable = False but once you make the discriminator non trainable then you wouldn't able to train it on the following batches. The discriminator has to be trained as long as t>Tc and in order to do that I guess we need to set d_container.trainable = True after the completion network is trained. Correct me if I am wrong and thank you for your quick response.

The algorithm on the paper is something like this:

            if n < tc:
                ''' Train completion network '''
            elif n<tc+td:
                ''' Train discriminator network '''
            else:
                ''' Train both completion and discriminator '''

What I am suggesting is something like this:

if n >= tc + td:
     d_container.trainable = False
    all_model = Model([org_img, mask, in_pts], [cmp_out, d_container([cmp_out, in_pts])])
    all_model.compile(loss=['mse', 'binary_crossentropy'], loss_weights=[1.0, alpha], optimizer=optimizer)
    g_loss = all_model.train_on_batch([inputs, masks, points], [inputs, valid])
    g_loss = g_loss[0] + alpha * g_loss[1]

    "" the following codes makes the discriminator trainable again on the following batch "" 
     d_container.trainable = True 
     all_model.compile(loss=['mse', 'binary_crossentropy'], loss_weights=[1.0, alpha], optimizer=optimizer) ```

[neka-nat]

The following links may be helpful for your point. https://stackoverflow.com/questions/45154180/how-to-dynamically-freeze-weights-after-compiling-model-in-keras

The trainable flag is fixed in the model at compile time. So, changes to the flag after compile will not affect the compiled model.

[bis-carbon]

Thank you, that clarifies my question.