Skip to content

Latest commit

 

History

History
63 lines (41 loc) · 3.49 KB

readme.md

File metadata and controls

63 lines (41 loc) · 3.49 KB
Raw

This repo is for reproducing our results in "Lipschitz Generative Adversarial Nets".

We use tensorflow 1.5 with python 3.5.

You can refer to setting_cuda9_cudnn7_tensorflow1.5.sh to build up your environment.

sobolev_cifar10.py:

This code can be used to reproduce results in Table 2, Figure4 and Figure 6.

Run the following code with varying fWeightLip and varying sGAN_type, e.g.,:

    python3 sobolev_cifar10.py -sDataSet=cifar10 -fWeightLip=0.01 -sGAN_type=exp
    python3 sobolev_cifar10.py -sDataSet=cifar10 -fWeightLip=0.1 -sGAN_type=exp
    python3 sobolev_cifar10.py -sDataSet=cifar10 -fWeightLip=1.0 -sGAN_type=exp
    python3 sobolev_cifar10.py -sDataSet=cifar10 -fWeightLip=10.0 -sGAN_type=exp

    python3 sobolev_cifar10.py -sDataSet=cifar10 -fWeightLip=0.1 -sGAN_type=x
    python3 sobolev_cifar10.py -sDataSet=cifar10 -fWeightLip=0.1 -sGAN_type=log_sigmoid
    python3 sobolev_cifar10.py -sDataSet=cifar10 -fWeightLip=0.1 -sGAN_type=sqrt
    python3 sobolev_cifar10.py -sDataSet=cifar10 -fWeightLip=0.1 -sGAN_type=lsgan
    python3 sobolev_cifar10.py -sDataSet=cifar10 -fWeightLip=0.1 -sGAN_type=hinge

By default we use the MaxGP, to switch to GP:

    python3 sobolev_cifar10.py -sDataSet=cifar10 -sGAN_type=exp -fWeightLip=10.0 -bMaxGP=False -sGP_type=gp

To try the buffered MaxGP, add the following flag: 

    -fBufferBatch=0.25
    or
    -fBufferBatch=-0.25

    The buffer size equals to fBufferBatch * iBatchSize; 
    The sign of fBufferBatch indicates the way we use the buffer: 
        postive -> extend the batch: batch size for maxgp becomes iBatchSize * (1+fBufferBatch) 
        negative -> insert into the batch: keep the batchsize of gp unchanged.

gan_synthetic4.py:

This code can be used to reproduce results in Figure 1, Figure 2 and Appendix B.6.

To reproduce Figure 1, run the following code: 

    python3 gan_synthetic4.py -iBaseNumFilterD=128 -iBlockPerLayerD=64 -oActD=selu -oOptD=sgd -fLrIniD=1e-4 -sGAN_Type=lsgan -bLip=False -sResultTag=case1
    python3 gan_synthetic4.py -iBaseNumFilterD=128 -iBlockPerLayerD=64 -oActD=selu -oOptD=sgd -fLrIniD=1e-4 -sGAN_Type=lsgan -bLip=False -sResultTag=case2
    python3 gan_synthetic4.py -iBaseNumFilterD=128 -iBlockPerLayerD=64 -oActD=selu -oOptD=sgd -fLrIniD=1e-4 -sGAN_Type=lsgan -bLip=False -sResultTag=case3       

To reproduce Figure 2, run the following code:

    python3 gan_synthetic4.py -iBaseNumFilterD=1024 -iBlockPerLayerD=4 -oActD=relu -oOptD=adam -fLrIniD=1e-4 -bLip=True -sGAN_Type=x -sResultTag=case3
    python3 gan_synthetic4.py -iBaseNumFilterD=1024 -iBlockPerLayerD=4 -oActD=relu -oOptD=adam -fLrIniD=1e-4 -bLip=True -sGAN_Type=log_sigmoid -sResultTag=case3
    python3 gan_synthetic4.py -iBaseNumFilterD=1024 -iBlockPerLayerD=4 -oActD=relu -oOptD=adam -fLrIniD=1e-4 -bLip=True -sGAN_Type=sqrt -sResultTag=case3
    python3 gan_synthetic4.py -iBaseNumFilterD=1024 -iBlockPerLayerD=4 -oActD=relu -oOptD=adam -fLrIniD=1e-4 -bLip=True -sGAN_Type=exp -sResultTag=case3

gan_synthetic4.3.py

This code can be used to reproduce Figure 3. 

    python3 gan_synthetic4.3.py -n=2  # using two cifar10 images as P_r, which can get a quick result.
    python3 gan_synthetic4.3.py -n=10 # using ten cifar10 images as P_r, which is the same setting as Figure 3.

We use training_curve_draw.py to draw the training curves and select random samples.