vseledkin / fast-neural-doodle Goto Github PK

This is my try on drawing with neural networks, which is faster than Alex J. Champandard's version, and similar in quality. This approach is based on neural artistic style method (L. Gatys), whereas Alex's version uses CNN+MRF approach of Chuan Li.

It takes several minutes to redraw Renoir example using GPU and it will easily fit in 4GB GPUs. If you were able to work with Justin Johnson's code for artistic style then this code should work for you too.

• torch
• torch.cudnn (optional)
• torch-hdf5
• python + numpy + scipy + h5py + sklearn

Tested with python2.7 and latest conda packages.

First download VGG-19.

cd data/pretrained && bash download_models.sh && cd ../..

Use this script to get intermediate representations for masks.

python get_mask_hdf5.py --n_colors=4 --style_image=data/Renoir/style.png --style_mask=data/Renoir/style_mask.png --target_mask=data/Renoir/target_mask.png

Now run doodle.

th fast_neural_doodle.lua -masks_hdf5 masks.hdf5 -vgg_no_pad

And here is the result. First row: original, second -- result.

And Monet.

Processing the image at low resolution first can provide a significant speed-up. You can pass a list of resolutions to use when processing. Passing 256 means that the images and masks should be resized to 256x256 resolution. With 0 passed no resizing is done. Here is an example for cmd parameters:

• -num_iterations 450,100 -resolutions 256,0 Which means: work for 450 iterations at 256x256 resolution and 100 iterations at original.

Monet and Renoir examples take ~1.5 min to process with these options.

• Supported backends:

  • nn (CPU/GPU mode)
  • cudnn
  • clnn (not tested yet..)

• When using -backend cudnn do not forget to switch -cudnn_autotune.

The code is heavily based on Justin Johnson's great code for artistic style.