OCaml-neural-style-transfer

This is the final project for Cornell CS3110 - Data Structures and Functional Programming. This repo belongs to Canal Li, Thomas Cui and Canwen Zhang. This project partially adapts from examples in PyTorch Ocaml binding: https://github.com/LaurentMazare/ocaml-torch

Step 0: update system

sudo apt update sudo apt upgrade

Step 1: install required packages

sudo apt install pkg-config libffi-dev zlib1g-dev imagemagick

Step 2: install ocaml pytorch

opam install torch ANSITerminal

Step 3:

For MS3, we now enable to use different VGG models. Download the pretrained weights from:

vgg16: https://github.com/LaurentMazare/ocaml-torch/releases/download/v0.1-unstable/vgg16.ot

vgg19: https://github.com/LaurentMazare/ocaml-torch/releases/download/v0.1-unstable/vgg19.ot

and save the file in folder /resources

To run the engine,

make launch Content image can be any image (should be smaller than 1k otherwise likely out of memory) - Here use cornell as an example Style image should be any image (should be smaller than 1k otherwise likely out of memory) - Here use starry as an example

Pre-trained model can be any vgg models, but currently Ocaml Torch only provides vgg16.ot and vgg19.ot for downloads

For both VGG19 and VGG16, the default flags that work (for both of them):

-layers_style_loss [2,10,14,21,28]

-layers_content_loss [21]

-style_weight 9e5

-learning_rate 8e-2

-total_steps 80

We will update this list as soon as we add more pretrained models

The default flags for filtering/resizing are:

-k 5

-sigma 0.75

All of the flags are optional. To specify a flag value, use -flag value. The flags not specified will use the default values. What we generated in MS1 is an "artwork". What we generated in MS2 is a "picture". The differences are explained in our progress report.

The output artwork is in /data/output The gif for correspond artwork is in /data/output