generative models experimentations

in 2023, i’ve participated in a statistical machine learning course and, as part of my assignments, i explored some models to play with. since we are in a generative/llm era, why not to play with them? developed a deep convolutional generative adversarial network, using tensorflow and fed with celeba dataset.

striving to find time to explore, but other generative implementations are on the roadmap. More on the post.

i have a special interest in images as they represent somewhat complex data to work with. don’t you think there is some marvel in a group of pixels being able to bring back meaning? we are so used to correlate 3D and 2D that we don’t even realize how amazing is to be able to generalize and identify objects and scenarios. since i’ve got other experiences with convolutional networks, exploring with famous architectures as YOLO and inception, i wanted to try a deep conv net applying an generative adversarial framework.

the whole concept is not hard to understand: you’ll train two models simultaneously: a generative model(G) to model stuff like your data and a discriminative model(D) to classify original data or G generated material. thinking in a ideal world, were everything works as expected and there is no overfitting, D reaching near perfection and being deceived by G would imply that G capturated the essence of the original data, being able to create new samples that would perfectly impersonate anyone. mathematics are not that permissive so it doesn’t, usually, gets to this point. but results are pretty okay.

eventually this code will be available on gitlab, my main coding platform. right now, i’ve a dirty jupyter document which contains a lot more of text i want to publish. this is a ongoing effort, as i still want to implement other generative models to compare with, such as one of those amazing variational autoencoders ( have you ever understood how this thing works? makes total sense, but little fiddlyto implement )