def __init__(self, args):
super().__init__()
self.args = args
self.device = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
self.nets, self.nets_ema = build_model(args)
# self.arcface, self.conf = load_arcface()
# self.arcface = load_arcface_2()
self.writer = SummaryWriter(
'log/test_vox_256_smalldata_id_1_20_20_retrain_alldata_id_embedder_vggface_add_noise'
)
# print(self.arcface)
# assert False
# below setattrs are to make networks be children of Solver, e.g., for self.to(self.device)
for name, module in self.nets.items():
utils.print_network(module, name)
setattr(self, name, module)
for name, module in self.nets_ema.items():
setattr(self, name + '_ema', module)
if args.mode == 'train':
self.optims = Munch()
for net in self.nets.keys():
if net == 'fan':
continue
self.optims[net] = torch.optim.Adam(
params=self.nets[net].parameters(),
lr=args.f_lr if net == 'mapping_network' else args.lr,
betas=[args.beta1, args.beta2],
weight_decay=args.weight_decay)
# self.ckptios = [
# CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets.ckpt'), **self.nets),
# CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets_ema.ckpt'), **self.nets_ema),
# CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_optims.ckpt'), **self.optims)]
self.ckptios = [
CheckpointIO(ospj(args.checkpoint_dir, '{}_nets.ckpt'),
**self.nets),
CheckpointIO(ospj(args.checkpoint_dir, '{}_nets_ema.ckpt'),
**self.nets_ema),
CheckpointIO(ospj(args.checkpoint_dir, '{}_optims.ckpt'),
**self.optims)
]
else:
self.ckptios = [
CheckpointIO(
ospj(args.checkpoint_dir,
'{}_nets_ema.ckpt'.format(args.resume_iter)),
**self.nets_ema)
]
# self.ckptios = [CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets_ema.ckpt'), **self.nets_ema)]
self.to(self.device)
for name, network in self.named_children():
# Do not initialize the FAN parameters
if ('ema' not in name) and ('fan' not in name):
print('Initializing %s...' % name)
network.apply(utils.he_init)
def __init__(self, args):
super().__init__()
self.args = args
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.nets, self.nets_ema = build_model(args)
# below setattrs are to make networks be children of Solver, e.g., for self.to(self.device)
for name, module in self.nets.items():
utils.print_network(module, name)
setattr(self, name, module)
for name, module in self.nets_ema.items():
setattr(self, name + "_ema", module)
if args.resume_ckpt is not None:
print("Transfer learning from", args.resume_ckpt)
CheckpointIO(
args.resume_ckpt, **{
k: (n.module if isinstance(n, nn.DataParallel) else n)
for k, n in self.nets.items()
}).load(args.resume_ckpt.split("_")[0], restore_D=False
) # no discriminator included in EMA ckpts :(\
if args.mode == "train":
self.optims = Munch()
for net in self.nets.keys():
if net == "fan":
continue
self.optims[net] = torch.optim.Adam(
params=self.nets[net].parameters(),
lr=args.f_lr if net == "mapping_network" else args.lr,
betas=[args.beta1, args.beta2],
weight_decay=args.weight_decay,
)
self.ckptios = [
CheckpointIO(
ospj(args.checkpoint_dir, "{:06d}_nets.ckpt"), **{
k: (n.module if isinstance(n, nn.DataParallel) else n)
for k, n in self.nets.items()
}),
CheckpointIO(ospj(args.checkpoint_dir, "{:06d}_nets_ema.ckpt"),
**self.nets_ema),
CheckpointIO(ospj(args.checkpoint_dir, "{:06d}_optims.ckpt"),
**self.optims),
]
else:
self.ckptios = [
CheckpointIO(ospj(args.checkpoint_dir, "{:06d}_nets_ema.ckpt"),
**self.nets_ema)
]
self.to(self.device)
for name, network in self.named_children():
# Do not initialize the FAN parameters
if ("ema" not in name) and ("fan" not in name):
print("Initializing %s..." % name)
network.apply(utils.he_init)
def __init__(self, args):
super().__init__()
self.args = args
self.device = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
self.summary_writer = SummaryWriter('./runs/experiment_1')
#self.nets, self.nets_ema = build_model(args)
self.nets = build_model(args)
# below setattrs are to make networks be children of Solver, e.g., for self.to(self.device)
for name, module in self.nets.items():
utils.print_network(module, name)
setattr(self, name, module)
# for name, module in self.nets_ema.items():
# setattr(self, name + '_ema', module)
if args.mode == 'train':
self.optims = Munch()
for net in self.nets.keys():
if net == 'fan':
continue
self.optims[net] = torch.optim.Adam(
params=self.nets[net].parameters(),
lr=args.lr,
betas=[args.beta1, args.beta2],
weight_decay=args.weight_decay)
self.ckptios = [
CheckpointIO(ospj(args.checkpoint_dir, '{0}_nets.ckpt'),
**self.nets),
#CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets_ema.ckpt'), **self.nets_ema),
CheckpointIO(ospj(args.checkpoint_dir, '{0}_optims.ckpt'),
**self.optims)
]
#""" load the pretrained checkpoint """
#self._load_checkpoint(step="", fname='./checkpoints/git_nets_ema.ckpt')
if args.mode == 'eval':
self.ckptios = [
CheckpointIO(ospj(args.checkpoint_dir, '{0}_nets.ckpt'),
**self.nets),
#CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets_ema.ckpt'), **self.nets_ema),
#CheckpointIO(ospj(args.checkpoint_dir, '{0}_optims.ckpt'), **self.optims)]
]
self.to(self.device)
for name, network in self.named_children():
# Do not initialize the FAN parameters
if ('ema' not in name) and ('fan' not in name):
print('Initializing %s...' % name)
network.apply(utils.he_init)
def __init__(self, args):
super().__init__()
self.args = args
self.device = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
self.nets, self.nets_ema = build_model(args)
# below setattrs are to make networks be children of Solver, e.g., for self.to(self.device)
for name, module in self.nets.items():
utils.print_network(module, name)
setattr(self, name, module)
for name, module in self.nets_ema.items():
setattr(self, name + '_ema', module)
if args.mode == 'train':
self.optims = Munch()
for net in self.nets.keys():
if net == 'fan':
continue
self.optims[net] = torch.optim.Adam(
params=self.nets[net].parameters(),
lr=args.f_lr if net == 'mapping_network' else args.lr,
betas=[args.beta1, args.beta2],
weight_decay=args.weight_decay)
self.ckptios = [
CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets.ckpt'),
**self.nets),
CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets_ema.ckpt'),
**self.nets_ema),
CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_optims.ckpt'),
**self.optims)
]
# self.ckptios = [
# CheckpointIO(ospj(args.checkpoint_dir, str(self.nets) + '_nets.ckpt'), **self.nets),
# CheckpointIO(ospj(args.checkpoint_dir, str(self.nets_ema)+ '_nets_ema.ckpt'), **self.nets_ema),
# CheckpointIO(ospj(args.checkpoint_dir, str(self.optims)+ '_optims.ckpt'), **self.optims)]
else:
self.ckptios = [
CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets_ema.ckpt'),
**self.nets_ema)
]
# self.ckptios = [CheckpointIO(ospj(args.checkpoint_dir, str(self.nets_ema) + '_nets_ema.ckpt'), **self.nets_ema)]
self.to(self.device)
for name, network in self.named_children():
# Do not initialize the FAN parameters
if ('ema' not in name) and ('fan' not in name):
print('Initializing %s...' % name)
network.apply(utils.he_init)
def __init__(self, args):
super().__init__()
self.args = args
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.nets, self.nets_ema = build_model(args)
# below setattrs are to make networks be children of Solver, e.g., for self.to(self.device)
for name, module in self.nets.items():
utils.print_network(module, name)
setattr(self, name, module)
for name, module in self.nets_ema.items():
setattr(self, name + "_ema", module)
if args.mode == "train":
self.optims = Munch()
for net in self.nets.keys():
if net == "fan":
continue
self.optims[net] = torch.optim.Adam(
params=self.nets[net].parameters(),
lr=args.f_lr if net == "mapping_network" else args.lr,
betas=[args.beta1, args.beta2],
weight_decay=args.weight_decay,
)
self.ckptios = [
CheckpointIO(ospj(args.checkpoint_dir, "{:06d}_nets.ckpt"),
**self.nets),
CheckpointIO(ospj(args.checkpoint_dir, "{:06d}_nets_ema.ckpt"),
**self.nets_ema),
CheckpointIO(ospj(args.checkpoint_dir, "{:06d}_optims.ckpt"),
**self.optims),
]
else:
self.ckptios = [
CheckpointIO(ospj(args.checkpoint_dir, "{:06d}_nets_ema.ckpt"),
**self.nets_ema)
]
self.to(self.device)
for name, network in self.named_children():
# Do not initialize the FAN parameters
if ("ema" not in name) and ("fan" not in name):
print("Initializing %s..." % name)
network.apply(utils.he_init)
def __init__(self, args):
super().__init__()
self.args = args
self.nets, self.nets_ema = build_model(args)
# below setattrs are to make networks be children of Solver, e.g., for self.to(self.device)
for name, module in self.nets.items():
setattr(self, name, module)
for name, module in self.nets_ema.items():
setattr(self, name + '_ema', module)
if args.mode == 'train':
place = paddle.fluid.CUDAPlace(
self.args.whichgpu) if paddle.fluid.is_compiled_with_cuda(
) else paddle.fluid.CPUPlace()
with fluid.dygraph.guard(place):
self.optims = Munch()
self.ckptios = Munch()
for net in self.nets.keys():
if net == 'fan':
continue
self.optims[net] = fluid.optimizer.AdamOptimizer(
learning_rate=args.f_lr
if net == 'mapping_network' else args.lr,
beta1=args.beta1,
beta2=args.beta2,
parameter_list=self.nets[net].parameters(),
regularization=fluid.regularizer.L2Decay(
regularization_coeff=args.weight_decay))
self.ckptios[net] = [
CheckpointIO(
ospj(args.checkpoint_dir,
'{:06d}_nets_ema_' + net)),
CheckpointIO(
ospj(args.checkpoint_dir, '{:06d}_nets_' + net))
]
else:
self.ckptios = Munch()
for net in self.nets.keys():
self.ckptios[net] = [
CheckpointIO(
ospj(args.checkpoint_dir, '{:06d}_nets_ema_' + net))
]
def load_model(args):
_, nets_ema = build_model(args)
ckptios = [
CheckpointIO(ospj(args.checkpoint_dir, '{0:06d}_nets_ema.ckpt'),
**nets_ema)
] # compatible with Windows
for ckptio in ckptios:
ckptio.load(args.resume_iter)
return nets_ema
def __init__(self, args):
super().__init__()
self.args = args
# self.device = porch.device('cuda' if porch.cuda.is_available() else 'cpu')
print("Solver init....")
self.nets, self.nets_ema = build_model(args)
# below setattrs are to make networks be children of Solver, e.g., for self.to(self.device)
for name, module in self.nets.items():
utils.print_network(module, name)
setattr(self, name, module)
for name, module in self.nets_ema.items():
setattr(self, name + '_ema', module)
if args.mode == 'train':
self.optims = Munch()
for net in self.nets.keys():
if net == 'fan':
continue
self.optims[net] = porch.optim.Adam(
params=self.nets[net].parameters(),
lr=args.f_lr if net == 'mapping_network' else args.lr,
betas=[args.beta1, args.beta2],
weight_decay=args.weight_decay)
self.ckptios = [
CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets_ema.ckpt'),
**self.nets),
CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets_ema.ckpt'),
**self.nets_ema),
CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_optims.ckpt'),
**self.optims)
]
else:
self.ckptios = [
CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets_ema.ckpt'),
**self.nets_ema)
]
self
def __init__(self, args):
super().__init__()
self.args = args
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
_, self.nets_ema = build_model(args)
for name, module in self.nets_ema.items():
setattr(self, name + '_ema', module)
self.ckptios = [CheckpointIO(ospj(args.checkpoint_dir, '{}_nets_ema.ckpt'.format(args.resume_iter)), **self.nets_ema)]
self.to(self.device)
self._load_checkpoint(args.resume_iter)
def __init__(self):
super().__init__()
args = resolver_args()
self.args = args
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.nets, self.nets_ema = build_model(args)
# below setattrs are to make networks be children of Solver, e.g., for self.to(self.device)
for name, module in self.nets.items():
utils.print_network(module, name)
setattr(self, name, module)
for name, module in self.nets_ema.items():
setattr(self, name + '_ema', module)
self.ckptios = [CheckpointIO(ospj(args.checkpoint_dir, '{:06d}_nets_ema.ckpt'), **self.nets_ema)]
self.to(self.device)
for name, network in self.named_children():
# Do not initialize the FAN parameters
if ('ema' not in name) and ('fan' not in name):
print('Initializing %s...' % name)
network.apply(utils.he_init)