def _project(self):
input_shape = self._input_shape
print("===Projecting===")
for i in range(self.num_projections):
if self.orthogonal: # project onto orthogonal convolution kernel space
self.weight.data = torch.from_numpy(
conv_clip_2_norm_numpy(self.weight.data.cpu().numpy(),
(self.weight.size(-2) * 2 - 1,
self.weight.size(-1) * 2 - 1),
clip_to=1,
force_same=True)).to(
device=self.weight.device,
dtype=self.weight.dtype)
else: # project onto 1-Lipschitz convolution kernel space
self.weight.data = torch.from_numpy(
conv_clip_2_norm_numpy(self.weight.data.cpu().numpy(),
input_shape,
clip_to=1)).to(
device=self.weight.device,
dtype=self.weight.dtype)
self.svs = torch.from_numpy(
conv_singular_values_numpy(self.weight.data.cpu().numpy(),
input_shape))
print("Projection Result: {}".format(self.lipschitz_constant()))
print("===Finished===")
def singular_values(self):
svs = torch.from_numpy(
conv_singular_values_numpy(
self.buffer_weight.detach().cpu().numpy(), self._input_shape
)
).to(device=self.buffer_weight.device)
return svs
def singular_values(self):
# Implements interface required by LipschitzModuleL2
svs = torch.from_numpy(
conv_singular_values_numpy(
self.buffer_weight.detach().cpu().numpy(), self._input_shape
)
).to(device=self.buffer_weight.device)
return svs
def singular_values(self):
return torch.from_numpy(
conv_singular_values_numpy(self.weight.data.cpu().numpy(),
self._input_shape))