Zeolite GAN

Tensorflow 1.12.0 implementation of ZeoGAN.

Generates energy and material shapes (in voxel format).

Input Arguments

--dataset_path /path/to/input_shapes              
--device       1                     # gpu device number
--logdir       /path/to/logdir       # path for log, checkpoint
--z_size       1024
--voxel_size   32
--rate         0.5
--move         True
--rotate       True
--invert       True
--energy_limit -4000 5000
--energy_scale -4000 5000
--cell_length_scale 0.0 150.0
--save_every     1000                # frequency of saving checkpoints
--batch_size     32
--bottom_size    4
--bottom_filters 256
--top_size       4
--filter_unit    32
--d_learning_rate 1e-4
--g_learning_rate 1e-4
--minibatch                False
--minibatch_kernel_size    256
--minibatch_dim_per_kernel 5
--l2_loss False
--train_gen_per_disc 1
--n_critics 5
--gp_lambda 10
--feature_matching True
--in_temper 298.0
--user_desired False                 # non-user-desired generation
--user_range 18 22
#--restore_ckpt /path/to/checkpoint  # if use pre-trained checkpoints

Training of Model

$ python main.py @input_example.in

In input_example.in,

--dataset_path /path/to/input_shapes              
--device       1                     # gpu device number
--logdir       /path/to/logdir       # path for log, checkpoint

...

--user_range 18 22
#--restore_ckpt /path/to/checkpoint  # if use pre-trained checkpoints

Generation of Material Shapes from Trained Model

$ python gen.py --checkpoint {} --n_samples {} --savedir={} --device {} --batch_size {} --type normal

checkpoint: checkpoint path in log_dir

e.g. ./test_log/save-2020-06-17T13:52:51.090310-100000

n_samples: number of generated shapes

e.g. 100000

savedir: directory for generated shapes

e.g. ./test_generation

device: GPU device index

batch_size: batch size for generation

batch_size < of n_samples / 100

$ python gen.py --checkpoint ./test_log/save-2020-06-17T13:52:51.090310-100000 --n_samples 10000 --savedir=./test_generation --device 0 --batch_size 100 --type normal