Training GANs with Optimism

Code for reproducing results in Training GANs with Optimism

Hereinafter, REPO_HOME is used to denote the path to this repository.

Prerequisites

• Keras (2.0.8)
• Theano
• Pillow
• Numpy, Scipy, cPickle, Matplotlib

DNA generation

Unzip the dataset

cd data/
tar -zxvf motif_spikein_ATAGGC_50runs.tar.gz

Trainng

For each of the optimizatin strategies with each of 3 different learning rates, train 50 WGAN models for 100 epochs. The resulting models are saved under $REPO_HOME/script/DNA.

DATADIR=../data/motif_spikein_ATAGGC_50runs
RUN_CMD=''
cd script/

for RUN in {0..49}
do
	for LR in '5e-02' '5e-03' '5e-04'
	do
		
		# SGD
        	KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/SGD_lr$LR \
        		--optimizer SGD --lr $LR
        
		# SGD with Adagrad
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/adagrad_lr$LR \
			--optimizer SGD  -s adagrad --lr $LR

		# SGD with momentum
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/momentum_lr$LR \
			--optimizer SGD  --momentum 0.9 --lr $LR

		# SGD with Nesterov momentum
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/nesterov_lr$LR \
			--optimizer SGD  --momentum 0.9 --nesterov  --lr $LR

		# SGD with Adam
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/adam_lr$LR \
		    --optimizer SGD  -s adam --lr $LR

		# optimAdam
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/optimAdam_lr$LR \
		    --optimizer optimAdam  --lr $LR

		# optimAdam with 1:1 training ratio
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/optimAdam_ratio1_lr$LR \
		    --optimizer optimAdam  --g_interval 1 --lr $LR
		
		# SOMD (3 different versions)
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/SOMDv1_lr$LR \
			--optimizer OMDA  -v 1 --lr $LR
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/SOMDv2_lr$LR \
			--optimizer OMDA  -v 2 --lr $LR
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/SOMDv3_lr$LR \
			--optimizer OMDA  -v 3 --lr $LR

		# SOMD with 1:1 training ratio (3 different versions)
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/SOMDv1_ratio1_lr$LR \
			--optimizer OMDA  --g_interval 1 -v 1 --lr $LR
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/SOMDv2_ratio1_lr$LR \
			--optimizer OMDA  --g_interval 1 -v 2 --lr $LR
		KERAS_BACKEND=theano python wgan_train.py -d $DATADIR/run$RUN -o DNA/run$RUN/SOMDv3_ratio1_lr$LR \
			--optimizer OMDA  --g_interval 1 -v 3 --lr $LR

	done
done

KL divergence Evaluation

Change the rundir argument in this Notebook to $REPO_HOME/script/DNA/ to reproduce the figures.

CIFAR 10

Training

Run the following codes to train a WGAN model for each of the optimization strategies compared in this section. The resulting models are saved under $REPO_HOME/script/cifar10.

cd script/

## Adam
python cifar10.py -o cifar10/adam --optimizer SGD -v 1 --schedule adam

## Adam with ratio=1
python cifar10.py -o cifar10/adam_ratio1 --optimizer SGD -v 1 --schedule adam --training_ratio  1

## optimAdam
python cifar10.py -o cifar10/optimAdam --optimizer optimAdam

## optimAdam with ratio=1
python cifar10.py -o cifar10/optimAdam_ratio1 --optimizer optimAdam  --training_ratio 1

Evaluation

Change the rundir argument in this Notebook to $REPO_HOME/script/cifar10/ to reproduce the figures.

License

MIT