def build(self):
logger.info("loading model {} from cache at {}".format(
self.model_name, self.weight_path))
# Load config
config = make_unconditioned_big_gan(self.config_path, self.weight_path)
self.config = config
logger.info("Model config {}".format(config))
self.net = BigGAN.Generator(**config)
def make_big_gan(weights_root, target_class):
config = make_biggan_config(weights_root)
config['resolution'] = utils.imsize_dict[config['dataset']]
config['n_classes'] = utils.nclass_dict[config['dataset']]
config['G_activation'] = utils.activation_dict[config['G_nl']]
config['D_activation'] = utils.activation_dict[config['D_nl']]
config['skip_init'] = True
config['no_optim'] = True
G = BigGAN.Generator(**config)
G.load_state_dict(torch.load(config['weights_root'], map_location='cpu'),
strict=True)
return ConditionedBigGAN(G, target_class).cuda()
def make_big_bi_gan_as_big_gan(weights_root, resolution=128):
attn_dict = {128: '64', 256: '128', 512: '64'}
dim_z_dict = {128: 120, 256: 140, 512: 128}
config = {
'G_param': 'SN',
'D_param': 'SN',
'G_ch': 96,
'D_ch': 96,
'D_wide': True,
'G_shared': True,
'shared_dim': 128,
'dim_z': dim_z_dict[resolution],
'hier': True,
'cross_replica': False,
'mybn': False,
'G_activation': nn.ReLU(inplace=True),
'G_attn': attn_dict[resolution],
'norm_style': 'bn',
'G_init': 'ortho',
'skip_init': True,
'no_optim': True,
'G_fp16': False,
'G_mixed_precision': False,
'accumulate_stats': False,
'num_standing_accumulations': 16,
'G_eval_mode': True,
'BN_eps': 1e-04,
'SN_eps': 1e-04,
'num_G_SVs': 1,
'num_G_SV_itrs': 1,
'resolution': resolution,
'n_classes': 1000
}
G = BigGAN.Generator(**config)
G.load_state_dict(torch.load(weights_root,
map_location=torch.device('cpu')),
strict=False)
return UnconditionalBigGAN(G).cuda()