def init_running_average_generator(self):
self.running_average_generator = Generator(self.pose_size,
self.start_channel_size,
self.image_channels)
self.running_average_generator = wrap_models(
self.running_average_generator)
to_cuda(self.running_average_generator)
self.running_average_generator = amp.initialize(self.running_average_generator,
None, opt_level=self.opt_level)
def init_generator(config, ckpt):
g = Generator(
config.models.pose_size,
config.models.start_channel_size,
config.models.image_channels
)
g.load_state_dict(ckpt["running_average_generator"])
g.eval()
torch_utils.to_cuda(g)
return g
def init_model(pose_size, start_channel_dim, image_channels,
discriminator_model):
if discriminator_model == "deep":
d = DeepDiscriminator
else:
assert discriminator_model == "normal"
d = Discriminator
discriminator = d(image_channels, start_channel_dim, pose_size)
generator = Generator(pose_size, start_channel_dim, image_channels)
discriminator, generator = wrap_models([discriminator, generator])
return discriminator, generator
def init_model(pose_size, start_channel_dim, image_channels,
discriminator_model):
if discriminator_model == "deep":
d = DeepDiscriminator
else:
assert discriminator_model == "normal"
d = Discriminator
discriminator = d(image_channels, start_channel_dim, pose_size)
generator = Generator(pose_size, start_channel_dim, image_channels)
discriminator, generator = wrap_models([discriminator, generator])
# Freeze discriminator
for parameter in discriminator.parameters():
parameter.requires_grad = False
return discriminator, generator