Conditional VAE As stated in my first post, it can useful sometimes to have a generative model that can generate data knowing some prior information about the data to generate. By slightly modifying the Vanilla VAE and its underlying probabilistic assumptions, we can obtain a conditional generative model. More concretely, in our case, we would … Continue reading Stage 2 : Conditional VAE

Convolutional/deconvolutional L2 reconstruction networks The models I developed for this first stage of the project have architectures very close to Autoencoders ones. They consist in two stacked "submodels" : the encoder and the decoder. The encoder is used to produce a compact and low-dimensional representation of the images borders. The decoder is used to reconstruct … Continue reading Stage 1 : conv/decon reconstructor