def __call__(self, data):
if data.pos.size(-1) == 2:
degree = math.pi * random.uniform(*self.degrees) / 180.0
sin, cos = math.sin(degree), math.cos(degree)
matrix = [[cos, sin], [-sin, cos]]
else:
m1, m2, m3 = torch.eye(3), torch.eye(3), torch.eye(3)
if 0 in self.axes:
degree = math.pi * random.uniform(*self.degrees) / 180.0
sin, cos = math.sin(degree), math.cos(degree)
m1 = torch.tensor([[1, 0, 0], [0, cos, sin], [0, -sin, cos]])
if 1 in self.axes:
degree = math.pi * random.uniform(*self.degrees) / 180.0
sin, cos = math.sin(degree), math.cos(degree)
m2 = torch.tensor([[cos, 0, -sin], [0, 1, 0], [sin, 0, cos]])
if 2 in self.axes:
degree = math.pi * random.uniform(*self.degrees) / 180.0
sin, cos = math.sin(degree), math.cos(degree)
m3 = torch.tensor([[cos, sin, 0], [-sin, cos, 0], [0, 0, 1]])
matrix = torch.mm(torch.mm(m1, m2), m3)
# LinearTransformation only rotates `.pos`; need to rotate `.cell` too.
data_rotated = LinearTransformation(matrix)(data)
if hasattr(data_rotated, "cell"):
data_rotated.cell = torch.matmul(data_rotated.cell, matrix)
return (
data_rotated,
matrix,
torch.inverse(matrix),
)
def test_cartesian():
matrix = torch.tensor([[2, 0], [0, 2]], dtype=torch.float)
transform = LinearTransformation(matrix)
out = 'LinearTransformation([[2.0, 0.0], [0.0, 2.0]])'
assert transform.__repr__() == out
pos = torch.tensor([[-1, 1], [-3, 0], [2, -1]], dtype=torch.float)
data = Data(pos=pos)
out = transform(data).pos.tolist()
assert out == [[-2, 2], [-6, 0], [4, -2]]
def test_cartesian():
matrix = torch.tensor([[2, 0], [0, 2]], dtype=torch.float)
transform = LinearTransformation(matrix)
assert transform.__repr__() == ('LinearTransformation('
'[[2.0, 0.0], [0.0, 2.0]])')
pos = torch.Tensor([[-1, 1], [-3, 0], [2, -1]])
data = Data(pos=pos)
data = transform(data)
assert len(data) == 1
assert data.pos.tolist() == [[-2, 2], [-6, 0], [4, -2]]
def __call__(self, data):
if data.pos.size(-1) == 2:
degree = math.pi * random.uniform(*self.degrees) / 180.0
sin, cos = math.sin(degree), math.cos(degree)
matrix = [[cos, sin], [-sin, cos]]
else:
m1, m2, m3 = torch.eye(3), torch.eye(3), torch.eye(3)
if 0 in self.axes:
degree = math.pi * random.uniform(*self.degrees) / 180.0
sin, cos = math.sin(degree), math.cos(degree)
m1 = torch.tensor([[1, 0, 0], [0, cos, sin], [0, -sin, cos]])
if 1 in self.axes:
degree = math.pi * random.uniform(*self.degrees) / 180.0
sin, cos = math.sin(degree), math.cos(degree)
m2 = torch.tensor([[cos, 0, -sin], [0, 1, 0], [sin, 0, cos]])
if 2 in self.axes:
degree = math.pi * random.uniform(*self.degrees) / 180.0
sin, cos = math.sin(degree), math.cos(degree)
m3 = torch.tensor([[cos, sin, 0], [-sin, cos, 0], [0, 0, 1]])
matrix = torch.mm(torch.mm(m1, m2), m3)
return (
LinearTransformation(matrix)(data),
matrix,
torch.inverse(matrix),
)
def __call__(self, data):
dim = data.pos.size(-1)
matrix = data.pos.new_empty(dim, dim).uniform_(-self.shear, self.shear)
eye = torch.arange(dim, dtype=torch.long)
matrix[eye, eye] = 1
return LinearTransformation(matrix)(data)
def __call__(self, data):
degree = math.pi * random.uniform(*self.degrees) / 180.0
sin, cos = math.sin(degree), math.cos(degree)
if data.pos.size(1) == 2:
matrix = [[cos, sin], [-sin, cos]]
else:
if self.axis == 0:
matrix = [[1, 0, 0], [0, cos, sin], [0, -sin, cos]]
elif self.axis == 1:
matrix = [[cos, 0, -sin], [0, 1, 0], [sin, 0, cos]]
else:
matrix = [[cos, sin, 0], [-sin, cos, 0], [0, 0, 1]]
return LinearTransformation(torch.tensor(matrix))(data)
def test_linear_transformation():
matrix = torch.tensor([[2, 0], [0, 2]], dtype=torch.float)
transform = LinearTransformation(matrix)
assert str(transform) == ('LinearTransformation(\n'
'[[2. 0.]\n'
' [0. 2.]]\n'
')')
pos = torch.Tensor([[-1, 1], [-3, 0], [2, -1]])
data = Data(pos=pos)
data = transform(data)
assert len(data) == 1
assert data.pos.tolist() == [[-2, 2], [-6, 0], [4, -2]]
def generate_validation_dataset(self):
val_set = ShapeNet(
ShapeNetGenerator.PATH,
categories=self.categories,
include_normals=False,
split='val',
transform=LinearTransformation(
torch.tensor([
[2.0, 0, 0],
[0, 2.0, 0],
[0, 0, 2.0]
])
)
)
self.val_size = len(val_set)
return val_set