DCGAN

PyTorch implementation of DCGAN.

UNSUPERVISED REPRESENTATION LEARNING WITH DEEP CONVOLUTIONAL GENERATIVE ADVERSARIAL NETWORKS.

Modifications for Implementation

To ensure output size (64 x 64), I set kernel size, stride, padding size to 4, 2, 1, respectively.

Run Code

$ python train.py

DDP MULTI-GPUs Training

if you have multi-gpus using it to train more faster! I use it for imagenet.

$ python -m torch.distributed.launch --nproc_per_node={num gpus} --master_port={port} train_ddp.py --dataset_name 'imagenet' --dataset_path {your dataset path} --n_batch {batchsize}

Generated Images

CIFAR10

Epochs : 10 - 50

Epochs : 100, 200, 300, 500, 1000

Epochs : 1000

LSUN

Epochs : 200

IMAGENET (32 x 32 min-resized center crops)

Epochs : 200

Walking in the latent space

TODO

• model code
• train code
• add model save code
• add load check point
• add infer code : generate samples
• add other datasets (ImageNet)
• add other datasets (LSUN)
• vis results
• add argparse for hyperparams
• update wandb to log losses
• update wandb to log samples
• tqdm : progress bar
• ddp : support multi-gpu training