GAN Showcase

https://alantian.github.io/ganshowcase/ is available as a web showcase where a deep GAN (Generative Adversarial Network) that generates (or dreams) images. This repo contains all code for it. Feedback are welcome!

Technically, the network architecture is similar to the residual network (ResNet) based generator (Gulrajani et al.), as well as the classical DCGAN generator Radford et al. and the GAN training uses DRAGAN Kodali et al. style gradient penalty for better stability.

Training code is written inChainer. The trained model is then manually converted to a Keras model, which in turn is converted to a web-runnable TensorFlow.js model.

The dataset used for training is CelebAHQ, an dataset for Karras et al. which can be obtained by consulting its GitHub repo (https://github.com/tkarras/progressive_growing_of_gans).

What is in This Repo

This repo consists of code that

Prepares data
trains deep GAN.
Converts the saved model to a web-runnable one.
Presents the deep neural network, running completely in the browser.

How to use the code

Step 1 through 3 are done offline and are covered by scripts in ./code/ directory and their discussing assumes that you are in ./code/ directory. Also note several bash variables in UPPERCASE should be adjusted accordingly.

Step 1 - Prepare data

Dataset is stored as an npz file, and can be converted from either a folder containing images or the CelebAHQ dataset.

For using the folder of images, use

DIR_PATH=...
DATA_FILE=...
SIZE=... # can be 64, 128 or 256
./datatool.py --task dir_to_npz \
  --dir_path $DIR_PATH --npz_path $DATA_FILE --size $SIZE

For using the CelebAHQ dataset which can be obtained by consulting its GitHub repo:

CELEBAHQ_PATH=...  # should be an h5 file
DATA_FILE=...
SIZE=... # can be 64, 128 or 256

../scripts/run_docker.sh \
./datatool.py --task multisize_h5_to_npz \
  --multisize_h5_path $CELEBAHQ_PATH  --npz_path $DATA_FILE --size $SIZE

Step 2 - Training the model


# Training DCGAN64 model
DATA_FILE_SIZE_64=...  # Data file of size 64
DCGAN64_OUT=... # Output directory
./chainer_dcgan.py \
  --arch dcgan64 \
  --image_size 64 \
  --adam_alpha 0.0001 --adam_beta1 0.5 --adam_beta2 0.999 --lambda_gp 1.0 --learning_rate_anneal 0.9 --learning_rate_anneal_trigger 0 --learning_rate_anneal_interval 5000 --max_iter 100000 --snapshot_interval 5000 --evaluation_sample_interval 100 --display_interval 10 \
  --npz_path $DATA_FILE_SIZE_64 \
  --out $DCGAN64_OUT \
  ;


# Training ReSNet128 model
DATA_FILE_SIZE_128=...  # Data file of size 64
RESNET128_OUT=... # Output directory
./chainer_dcgan.py \
  --arch dcgan64 \
  --image_size 64 \
  --adam_alpha 0.0001 --adam_beta1 0.5 --adam_beta2 0.999 --lambda_gp 1.0 --learning_rate_anneal 0.9 --learning_rate_anneal_trigger 0 --learning_rate_anneal_interval 5000 --max_iter 100000 --snapshot_interval 5000 --evaluation_sample_interval 100 --display_interval 10 \
  --npz_path $DATA_FILE_SIZE_128 \
  --out $RESNET128_OUT \
  ;


# Training ReSNet256 model
DATA_FILE_SIZE_256=...  # Data file of size 64
RESNET256_OUT=... # Output directory
./chainer_dcgan.py \
  --arch dcgan64 \
  --image_size 64 \
  --adam_alpha 0.0001 --adam_beta1 0.5 --adam_beta2 0.999 --lambda_gp 1.0 --learning_rate_anneal 0.9 --learning_rate_anneal_trigger 0 --learning_rate_anneal_interval 5000 --max_iter 100000 --snapshot_interval 5000 --evaluation_sample_interval 100 --display_interval 10 \
  --npz_path $DATA_FILE_SIZE_256 \
  --out $RESNET256_OUT \
  ;

Step 3 - Convert from Chainer model to Keras/Tensorflow.js model

Note that due to difficulty in training GANs, you may want to select a proper snapshot by specifying ITER below. This script also samples a few images serving as a sanity check and providing clue for picking the correct snapshot.

# DCGAN64

ITER=50000
./dcgan_chainer_to_keras.py \
  --arch dcgan64 \
  --chainer_model_path $DCGAN64_OUT/SmoothedGenerator_${ITER}.npz \
  --keras_model_path $DCGAN64_OUT/Keras_SmoothedGenerator_${ITER}.h5 \
  --tfjs_model_path $DCGAN64_OUT/tfjs_SmoothedGenerator_${ITER} \
  ;

# ResNet128

ITER=20000
./dcgan_chainer_to_keras.py \
  --arch resnet128 \
  --chainer_model_path $RESNET128_OUT/SmoothedGenerator_${ITER}.npz \
  --keras_model_path $RESNET128_OUT/Keras_SmoothedGenerator_${ITER}.npz.h5 \
  --tfjs_model_path $RESNET128_OUT/tfjs_SmoothedGenerator_${ITER} \
  ;

# ResNet256

ITER=45000
./dcgan_chainer_to_keras.py \
  --arch resnet256 \
  --chainer_model_path $RESNET256_OUT/SmoothedGenerator_${ITER}.npz \
  --keras_model_path $RESNET256_OUT/Keras_SmoothedGenerator_${ITER}.npz.h5 \
  --tfjs_model_path $RESNET256_OUT/tfjs_SmoothedGenerator_${ITER} \
  ;

Step 4 - Present the generative as a web page

This step is covered by a web project under ./webcode/ganshowcase. Now it is assumeed that you are in ./webcode/ganshowcase directory.

First you need to copy TensorFlow.js model (specified as argument to --tfjs_model_path in previous step) to a public accessible place, and modify model_url in all_model_info which is in the beginning of index.js.

Then run the following:

yarn
yarn build

Finally, copy ./dist/, which is the built web page and js file, to whatever suitable for web hosting.

As an example, deploy.sh does the compilation and put everthing to docs, since the GitHub Pages site is currently being built from the /docs folder in the master branch.