Stop training the context during train_G/train_D? about dialogwae HOT 3 CLOSED

We considered this issue and tried such training strategy. However, the performance looks worse.
I think if we shut down context encoder, the auto-encoder is just independent to the GAN module thus cannot be regularized by the WAE.
You can still try it and let us know any better strategies.

from dialogwae.

Thanks for the quick reply.

So how about turning both utterance encoder and context encoder on during training generator and discriminator.

I assume you had tried such strategy and got worse performance. I would like to ask how bad it became when utterance encoder was turned on. Was the drop significant?

This paper proposed an interesting thought, nice work.

from dialogwae.

I did that! Please be noted that the context encoder contains the utterance encoder,

self.context_encoder = ContextEncoder(self.utt_encoder, config['n_hidden']*2+2, config['n_hidden'], 1, config['noise_radius'])

so whenever we update the context encoder, the utterance encoder will be updated as well.

from dialogwae.