MXNet-scala module implementation of fast neural style[1].

Based on:

https://github.com/jcjohnson/fast-neural-style and https://github.com/dmlc/mxnet/tree/master/example/neural-style/end_to_end

The results are not as good as the torch version but not bad anyway :)

Time: 0.075 seconds in average on NVIDIA GTX 1070 GPU for resolution of 712 x 414.

   

 

Tested on Ubuntu 14.04

• sbt 0.13 https://www.scala-sbt.org/
• Mxnet https://github.com/dmlc/mxnet

1, compile Mxnet with CUDA, then compile the scala-pkg，doc： https://github.com/dmlc/mxnet/tree/master/scala-package

2, under the Mxnet-Scala/FastNeuralStyle folder

 mkdir lib;

3, copy the compiled mxnet-full_2.11-INTERNAL.jar into lib folder;

4, run sbt then compile the project

The script run_fast_neural_style.sh under the scripts folder, lets you use a pretrained model to stylize new images:

java -Xmx1G -cp $CLASS_PATH \
	FastNeuralStyle \
	--model-path  $PREAREIN_MODEL \
	--input-image $INPUT_IMAGE \
	--output-path $OUTPUT_PATH \
	--gpu $GPU

You can run this script on CPU or GPU,

For cpu set the GPU flag to -1;

For gpu plesase specifying the GPU number on which to run.

Some pretrained style models are in datas/pretrain_models

1, download the vgg model

cd scripts
bash download_vgg_model.sh

This will download the pretrained vgg model file (~80MB) to the folder datas/vggmodel.

2, download the Coco dataset：https://msvocds.blob.core.windows.net/coco2015/test2015.zip

3, use the script train_fast_nueral_style.sh under the scripts folder to train on new style image.

you just need to set the TRAIN_DATA_PATH flag to your coco dataset path and STYLE_IMAGE flag to your own style images.

# path to the coco dataset,
# you can download it by : https://msvocds.blob.core.windows.net/coco2015/test2015.zip
TRAIN_DATA_PATH=

VGG_MODEL=$ROOT/datas/vggmodel/vgg19.params

SAVE_MODEL_PATH=$ROOT/datas/models

STYLE_IMAGE=$ROOT/datas/images/the_scream.jpg

LEARNING_RATE=0.0001

# resume the training progress
# by adding the commamd line parameter: 
# --resume-model-path $RESUME_MODEL_PATH
RESUME_MODEL_PATH=

# -1 for cpu
GPU=0

java -Xmx1G -cp $CLASS_PATH \
	Train \
	--data-path $TRAIN_DATA_PATH \
	--vgg-model-path $VGG_MODEL \
	--save-model-path $SAVE_MODEL_PATH \
	--style-image $STYLE_IMAGE \
	--lr $LEARNING_RATE \
	--gpu $GPU

For now, the training program only support training batch size 1, and the program may crash during the training process, can't figure out the reason. If you encounter this situasion, you can set the RESUME_MODEL_PATH to your lastest saved model (a mdoel will be saved every 1000 iters udring training), then use the --resume-model-path command line parameter to resume the training.

MIT

[1] Johnson, Justin, Alexandre Alahi, and Li Fei-Fei. "Perceptual losses for real-time style transfer and super-resolution." arXiv preprint arXiv:1603.08155.