MXNET-Scala Pix2Pix

MXNet-scala module implementation of Pix2Pix[1].

Based on: https://github.com/phillipi/pix2pix

I reproduced the results of the 'edge2shoes' and 'edges2handbags' datasets successfully. But the results on the 'cityscapes' and 'facades' datasets were not good enough, so I did not show the results of them.

Results:

Requirements

• sbt 0.13 http://www.scala-sbt.org/

• Mxnet https://github.com/dmlc/mxnet

steps

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

2, under the Mxnet-Scala/Pix2Pix folder

 mkdir lib;

3, copy your compiled mxnet-full_2.11-linux-x86_64-gpu-0.10.0-SNAPSHOT.jar into lib folder;

4, run sbt then compile the project

Datasets

You can download the datasets with the datasets/download_dataset.sh. you can refer to https://github.com/phillipi/pix2pix for how to use this script.

Training new models

Use the train_pix2pix.sh script under scripts folder. If you keep all the default settings, you just need to provide the dataset path, and set the direction 'AtoB' or 'BtoA'

DATA_PATH=DIRECTION=

Testing

I did not release the pretrained models because the model size is too large ~200M.

But you can reproduce the results by using the train_pix2pix.sh script to train your own model.

Use the test_pix2pix.sh script under scripts folder.

You need to provide the input image path, pretrain model path and set the corresponding direction.

PREAREIN_G_MODEL=INPUT_IMAGE=# -1 for cpuGPU=0# "AtoB" or "BtoA"DIRECTION="BtoA"

License

MIT

Reference

[1] Isola, Phillip, et al. "Image-to-image translation with conditional adversarial networks." 2016.

链接：https://github.com/Ldpe2G/DeepLearningForFun/tree/master/Mxnet-Scala/Pix2Pix