MVA Generative Models project on Order Agnostic Autoregressive Diffusion Models.
Face generation (CelebA)
Run TinyCelebA.ipynb to train and evaluate a UNet model on our resized Tiny CelebA dataset. The custom dataset is composed of 20.000 60×73 images with 32 grey-levels. This model was trained for 10 hours on a single GPU.
Character Generation (binary MNIST)
Run MNIST.ipynb to train and evaluate a UNet model on the binary MNIST dataset. The dataset is composed of 60.000 28x28 binary images. This model was trained for 1 hour.
We used parts from UNet, oardm and pytorch-fid