>>12576645while those fancy new generative models are nice from a theoretical perspective, I only care about real-world performance (de-novo drug design is what I do for the company I work for). RNN's are my favorite because they are extremely fast to train and generalize the best, from what I've produced. They're a little inflexible though for specific problems by themselves, but I more or less just stack RNNs to produce better results.
VAE perform well, my second favorite probs. GANs are still being figured out in my domain, and don't seem to perform amazingly well, but I'm definitely keeping my eye on them as they have potential (VAE + GAN on the latent space is an okay-performing model, but things like starGAN seem to be onto something with their abstraction of input labels into latent space)
