Meeting Notes CS224s

Meeting Notes CS224s

  1. We want a trained Conv layer:

  2. Anthoy wrote an autoencoder & a GAN to train the Conv layer

  3. We need to work on style transfer component that will do the style transfer.

  4. Incorporate a discriminatory GAN loss into the style transfer thing. 

  5. Different from training loss from training the Conv layer

Archa:     How are you training the pretrained networks?

        Currently, trying to overfit on the content and style 

Size does not matter for the Conv layer. But we do need a fixed length for the autoencoder. 

Include a lot of conv layers at the start. Functions return those layers. 

Next step: setting up the style transfer mechanism using these      Returning some intermediate layers. 

Anthony involved in pretraining models. Archa working on setting up style architecture. 

To run another experiment, Chris: TRY a convolutional GAN     Try DC GAN architecture. 

     Radford, A., Metz, L., & Chintala, S. (2015). Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434.

Modify WGAN loss to use my call to the discriminator