Pythae: Unifying Generative Autoencoder Implementations in PyTorch

Key Takeaways

so hi everyone I'm Kim master of deck and today we represent you my python library pite that you need files generator to encourage implementations in pytosh well first what's the generality auto encoder well you may know what the encoders the class of different models that try to map complex input data into a much lower dimensional space called the Latin space and so the main idea is to learn a Latin representation so lower dimensional representation of those input data generative using a sampling distribution and that something distribution can be used either during training as for the variational training course or after training as for the regularized auto encoders but well why did we create p-day so well first as you may know existing implementations of the original papers may be difficult to adapt over use cases be different Frameworks or longer maintained and so pite has a brick like structure that allows users to use different models with different sampling techniques that neural network architectures and training schemes on their own data pretty easily second pitay provides you with a reproducible research environment so most of the models that are implemented in pite were actually able to reproduce the original paper results finally the code was thought to be as flexible as possible and as accessible as possible and so we will find on the GitHub an online documentation and the library is also Illustrated through different tutorials this is the structure of the code so basically this is a really brick light structure so on the left you you have the models that you can choose and you can Define your encoder and decoder architecture in the middle you've got the trainers so trainers also referred to be as flexible as possible because you can Define your optimizers schedulers and callbacks and on the right you've got the samples that you can choose to generate data from a trend uh General to encoder and you may find also other features such as experiment monitoring tools such as 1db Comet ML and ml flow that has been integrated to uh Pita and you can also share your train model through the organ phase app so now let's have a look to the API of the code so let's say that I want to train a version to encoder with pite on my own data so the first thing that I need to do is to Define architecture so it can be done pretty easily using the base encoder and base decoder class so I import them and I create two other classes my encoder and my decoder where basically I specify my neural network architecture then I instantiate those classes and I can launch training to run to training and different things so the first thing that I need is to define a training configuration this is basically what I specified the number of epoch the learning rate and the batch size that I want to use then I Define a model configuration here a VA configuration and I can instant set my VA so to instantiate it I pass the model configuration as argument but I also passed the encoder and decoder that I've just built before then I can instantiate a training Pipeline with the training configuration and the model and I can launch the training Pipeline on my own data in the end once the model is trained I can reload it using the auto modern class pretty easily and then I can use whatever sampler I like to generate new data from that train model so in that example this is a quotient mixture uh sampler so basically all you have to do is to Define your sampling configuration and you can instant set the generation Pipeline and you can launch a pipeline by specifying the number of samples that you want to generate so here you've got the list of the models so actually implemented in Pita so there are 25 models that are currently implemented in the library and most of which why both reproduce the original results uh provided in the papers finally those are different resources that you may need to get ranks on the code so basically the code is available in giftable touch link you will also find an online documentation the library is available in buy buy and can be installed with that command line and I want to stress that that library is also open to contributors so if you want to contribute to your new model a new sampler or simply fix a bug I would be very happy to help you in that direction thank you very much for attention I would be very happy to answer any of your questions if you want