def _create_discriminator(self):
return NetworksFactory.get_by_name('global_local',
input_nc=3 + self._D_cond_nc // 2,
norm_type=self._opt.norm_type,
ndf=64,
n_layers=4,
use_sigmoid=False)
def _init_create_networks(self):
# features network
self._net = NetworksFactory.get_by_name('prob_map_net3', conv_dim=16)
self._net.init_weights()
self._net = self._move_net_to_gpu(self._net)
if len(self._gpu_ids) > 0:
summary(self._net,
(3, self._opt.net_image_size, self._opt.net_image_size))
def _init_create_networks(self):
# features network
self._net = NetworksFactory.get_by_name('uv_prob_net2', num_nc=32)
self._net.init_weights()
self._net = self._move_net_to_gpu(self._net)
if len(self._gpu_ids) > 0:
summary(self._net,
(3, self._opt.net_image_size, self._opt.net_image_size))
def _init_create_networks(self):
# features network
self._net_features = NetworksFactory.get_by_name('vgg_features',
freeze=True)
self._net_features = self._move_net_to_gpu(self._net_features)
# bb network
if self._opt.lambda_bb > 0:
self._net_bb = NetworksFactory.get_by_name('bb_net_values',
freeze=False)
self._net_bb = self._move_net_to_gpu(self._net_bb,
output_gpu=self._gpu_bb)
# prob network
if self._opt.lambda_prob > 0:
self._net_prob = NetworksFactory.get_by_name('prob_net',
freeze=False)
self._net_prob = self._move_net_to_gpu(self._net_prob)
def _init_create_networks(self):
self._G = NetworksFactory.get_by_name(
'Ours_DeblurOnly',
eventbins_between_frames=self.eventbins_between_frames,
if_RGB=self._opt.channel,
inter_num=self.inter_num)
if len(self._opt.load_G) > 0:
self._load_network(self._G, self._opt.load_G)
else:
raise ValueError("Weights file not found.")
self._G.cuda()
def _init_create_networks(self):
self._G = NetworksFactory.get_by_name(
'DeblurOnly',
eventbins_between_frames=self.eventbins_between_frames,
if_RGB=self._opt.channel,
inter_num=self.inter_num)
self._load_network(self._G, self._opt.load_G)
self._G.cuda()
n_parameters = sum(p.numel() for p in self._G.parameters()
if p.requires_grad)
print('number of parameters of G:', n_parameters)
def _create_generator(self):
net = NetworksFactory.get_by_name(self._opt.gen_name, bg_dim=4, src_dim=3+self._G_cond_nc,
tsf_dim=3+self._G_cond_nc, repeat_num=self._opt.repeat_num)
if self._opt.load_path:
self._load_params(net, self._opt.load_path)
elif self._opt.load_epoch > 0:
self._load_network(net, 'G', self._opt.load_epoch)
else:
raise ValueError('load_path {} is empty and load_epoch {} is 0'.format(
self._opt.load_path, self._opt.load_epoch))
net.eval()
return net
def _create_generator(self):
return NetworksFactory.get_by_name(
'img_encoder', input_chann=4,
output_dim=33) # 3-rot, 45-PCA, 3-translation
def _create_discriminator(self):
return NetworksFactory.get_by_name('discriminator_wasserstein_gan',
c_dim=self._opt.cond_nc)
def _create_bgnet(self):
net = NetworksFactory.get_by_name('deepfillv2', c_dim=4)
self._load_params(net, self._opt.bg_model, need_module=False)
net.eval()
return net
def _create_network(self, net_name):
return NetworksFactory.get_by_name(net_name)
def _create_discriminator(self):
return NetworksFactory.get_by_name('discriminator_patch_gan', input_nc=3 + self._D_cond_nc,
norm_type=self._opt.norm_type, ndf=64, n_layers=4,
use_sigmoid=False, sn=self._opt.spectral_norm)
def _create_generator(self):
return NetworksFactory.get_by_name(self._opt.gen_name, bg_dim=4, src_dim=3+self._G_cond_nc,
tsf_dim=3+self._G_cond_nc, repeat_num=self._opt.repeat_num)
def create_image_encoder_and_grasp_predictor(self):
return NetworksFactory.get_by_name(
'img_encoder_and_grasp_predictor') # The output is 6 or 7 as we consider 6 or 7 grasp taxonomy classes for the Barrett hand
def _create_fcnet(self):
return NetworksFactory.get_by_name(
'MLP_refine', input_dim=51 + 21, output_dim=51
) # 51 for 3 wrist rotation, 45 PCA for joint rotation, 3 translation
def create_grasp_generator(self):
return NetworksFactory.get_by_name(
'grasp_generator', input_dim=3 + 3 + 7) # 3 for rotation 3 for tranlsation and 7 for hand joints
def _init_create_networks(self):
# features network
# self._net = NetworksFactory.get_by_name('small_net', freeze=False)
self._net = NetworksFactory.get_by_name('vgg_finetune')
self._net = self._move_net_to_gpu(self._net)
def _create_generator(self):
# _opt.cond_nc = 17 by default
return NetworksFactory.get_by_name(self._opt.generator_name, self._opt)
def _create_discriminator(self):
return NetworksFactory.get_by_name('mano_discriminator',
input_size=51 + 21 + 4)
def create_discriminator(self):
return NetworksFactory.get_by_name('discriminator',
input_dim=3 + 3 + 1) # 3 for rotation 3 for tranlsation and 1 for finger spread