A simple linear autoencoder using ocaml-torch and bimage. I used the MNIST fashion dataset but regular MNIST will work just fine too. As few as 10 dimensions in the bottleneck still works surprisingly well from a subjective evaluation.
First you need to extract the MNIST files into a folder called =data= in the repository root. Then you can build the project with
dune build
and start training with
dune exec ./train.exe
which should generate a =weights.ot= file that saves the learnt weights. To preview a comparison of a sample before and after passing through the AE, run
dune exec ./compare.exe
and you should see two new files called =image.jpg= and =processed_image.jpg=.
The following (tiny!) examples were generated when training with the defaults.
