Inspired by Zalando Fashion DNA encoding, I wanted to try to recreate it by also using the opportunity offered by their released Feidegger dataset.

I implemented the code to optimally download the dataset, and I firstly implemented a Linear Autoencoder (with disastrous results). The outcome was way better with a Convolutional autoencoder. I then used the learned encoding and a KDtree to search similarities between articles, and the queries results are shown in the following pictures.

I then moved to develop a DCGAN architecture both for generation and encoding, but without success for now, as highlighted in this issue.