GAN-Cifar10-PyTorch

A graduation design project: Collected GAN's improved models, and innovation introduces Contrastive Learning to help GAN improve the generated image effect.

This work adopts the evaluation index calculation method of IS and FID in Project(https://github.com/w86763777/pytorch-inception-score-fid)

Requirements

• Python 3
• matplotlib
• tensorboardX
• torchsummary
• dataclasses
• numpy
• Pillow
• scipy
• torch
• torchvision
• tqdm
• typing-extensions
• tensorboard

pip install -r requirements.txt

• Dataset

  • Cifar-10 dataset

• InceptionV3

  • if you have download this weight file, you can add it to data/weight/inceptionv3_weight_name,
  • or you can discard arguments: --inception_weight_path when you train model(the default value of this arguments is None)

• cifar-10 fid stats

CUDA_VISIBLE_DEVICES=0 \
python calc_stats.py --output=cifar_10_fid_stats.npz --use_torch --data_path=your_cifar-10_dataset_path

Usage

To train an DCGAN(CNN) on Cifar10

CUDA_VISIBLE_DEVICES=0 \
python train.py --n_d=1 --gpu=True --data_save_root=output --inception_weight_path=data/weight/inceptionv3_weight_name --ms_file_name=cifar_10_fid_stats.npz --experiment_name=dcgan_cnn --total_steps=100000 --latent_dim=100 --mode='dcgan' --model='dcgan_cnn' --batch_size=128

To train an WGAN(CNN) on Cifar10

CUDA_VISIBLE_DEVICES=0 \
python train.py --n_d=1 --gpu=True --data_save_root=output --inception_weight_path=data/weight/inceptionv3_weight_name --ms_file_name=cifar_10_fid_stats.npz --experiment_name=wgan_gp_cnn --total_steps=100000 --latent_dim=100 --mode='wgan' --model='wgan_gp_cnn' --batch_size=128	--b1=0 --b2=0.9

To train an WGAN(Resnet) on Cifar10

CUDA_VISIBLE_DEVICES=0 \
python train.py --n_d=5 --gpu=True --data_save_root=output --inception_weight_path=data/weight/pt_inception-2015-12-05-6726825d.pth --ms_file_name=cifar_10_fid_stats.npz --experiment_name=wgan_gp_resnet --total_steps=100000 --latent_dim=128 --mode='wgan' --model='wgan_gp_resnet' --batch_size=64 --b1=0 --b2=0.9

To train an WGAN(Resnet) with Contrastive Loss on Cifar10

CUDA_VISIBLE_DEVICES=0 \
python train.py --n_d=5 --gpu=True --data_save_root=output --weight_path=/root/wbw/zhangl/dataset/InceptionV3/pt_inception-2015-12-05-6726825d.pth --data_path=/root/wbw/zhangl/dataset/cifar10 --ms_file_name=m1s1_np.npz --experiment_name=con_gan_5_27 --total_steps=100000 --latent_dim=128 --mode='wgan' --model='con_gan' --batch_size=64 --b1=0 --b2=0.9 --t=0.5