DC-GAN Training

This codebase is a slightly modified version of ganshowcase. View the web version of the original here to see what's possible.

Requirements

• A GPU is (ideally) required for training. If you have an NVIDIA GPU on your machine, ensure cuda and cudnn are installed correctly. Or you can use a service like Spell or Google Colab to access a remote GPU. Training is possible on a CPU, but will be much slower — taking weeks instead of hours. Instructions for training on Spell or NYU HPC (for NYU students) are here
• Set up a Python environment, running Python 3.6 or higher

Usage

1) Download this repository

git clone https://github.com/noisyneuron/training-dcgan
cd training-dcgan

2) Collect training data

Collect a dataset of images and store them in the folder images. These could be images of the same type — eg. pictures of bedrooms or cityscapes — which would generate recognizable images of the same type, or you could use a more varied dataset that would probably generate more abstract output.

The code will resize and (center) crop the images to prepare for training, as 64px or 128px or 256px squares.

3) Training

There are 4 steps to training, the method of execution differs slightly based on where/how you are training. (Instructions for training on Spell or NYU HPC (for NYU students) are here):

1. Install dependencies

2. Convert image dataset to numpy arrays

3. Training (takes a few hours to days)

4. Converting the resulting model to a web-compatible model

• Install dependencies:

  pip install -r requirements.txt
  

• Convert to numpy array: Specify the cropped image size you would like to use for training (larger will take longer and require more memory) — 64, 128 or 256. The following code will generate a dataset.npz file.

  ./datatool.py --size IMAGE_SIZE
  

  (If the images are stored elsewhere, use --dir_path to specify the path. You can change the path of the generated .npz file with the --npz_path flag)

• Train: Depending on the image size you used, specify the architecture to train with --

  64px : ./chainer_dcgan.py --arch dcgan64 --image_size 64

  128px : ./chainer_dcgan.py --arch resnet128 --image_size 128

  256px : ./chainer_dcgan.py --arch resnet256 --image_size 256

  (If you used a custom path for the .npz file, you can use the --npz_path flag to specify this. Output is saved to the result directory by default, this can be changed with the --out flag. See chainer_dcgan.py for more parameters)

• Convert model: After the training is complete, examine the result directory — you should see a collection of files generated at different points of the training — SmoothedGenerator_XXXXX.npz.

  You can see sample images from different iterations in the result/preview_smoothed directory — examine these to choose the iteration that produced the best results, and use that iteration number to convert the model to a Tensorflow.js compatible format. Specify the architecture used.

  ./dcgan_chainer_to_keras.py --arch ARCHITECTURE_YOU_USED --chainer_model_path result/SmoothedGenerator_XXXXX.npz
  

  eg. using the dcgan64 architecture with the model generated at the 50000th iteration:

  ./dcgan_chainer_to_keras.py --arch dcgan64 --chainer_model_path result/SmoothedGenerator_50000.npz
  

  This will generate a folder SmoothedGenerator_50000_tfjs which holds all the files you need to use the model with ml5 — take note of the model.json file.

4) Using the model with ml5

See this example for using DCGAN with ml5. To use your newly created model, make your own manifest.json file, that contains the path to your model.json file (inside the folder generated in the previous step).

The contents of your manifest.json file should be : (modify as appropriate)

{
    "description": "DESCRIPTION OF YOUR MODEL",
    "model": "PATH/TO/YOUR/MODEL.JSON/FILE",
    "modelSize": THE_IMAGE_SIZE_YOU_USED,
    "modelLatentDim": 128
}

For eg.

{
    "description": "Simpsons dataset from Kaggle",
    "model": "SmoothedGenerator_50000_tfjs/model.json",
    "modelSize": 256,
    "modelLatentDim": 128
}