def test_exception(self, batch_size, device, dtype):
with pytest.raises(TypeError):
src = torch.rand(batch_size, 5, 2).numpy()
assert kornia.get_tps_transform(src, src)
with pytest.raises(ValueError):
src = torch.rand(batch_size, 5)
assert kornia.get_tps_transform(src, src)
def test_jit(self, batch_size, device, dtype):
src, dst = _sample_points(batch_size, device)
kernel, affine = kornia.get_tps_transform(src, dst)
op = kornia.warp_points_tps
op_jit = torch.jit.script(op)
assert_allclose(op(src, dst, kernel, affine),
op_jit(src, dst, kernel, affine))
def test_affine_only(self, batch_size, device, dtype):
src = torch.tensor(
[[[0.0, 0.0], [0.0, 1.0], [1.0, 0.0], [1.0, 1.0], [0.5, 0.5]]],
device=device).repeat(batch_size, 1, 1)
dst = src.clone() * 2.0
kernel, affine = kornia.get_tps_transform(src, dst)
assert_allclose(kernel, torch.zeros_like(kernel), atol=1e-4, rtol=1e-4)
def test_jit(self, batch_size, device, dtype):
src, dst = _sample_points(batch_size, device)
kernel, affine = kornia.get_tps_transform(src, dst)
image = torch.rand(batch_size, 3, 32, 32, device=device)
op = kornia.warp_image_tps
op_jit = torch.jit.script(op)
assert_close(op(image, dst, kernel, affine), op_jit(image, dst, kernel, affine), rtol=1e-4, atol=1e-4)
def test_gradcheck(self, batch_size, device, dtype, requires_grad):
opts = dict(device=device, dtype=torch.float64)
src, dst = _sample_points(batch_size, **opts)
kernel, affine = kornia.get_tps_transform(src, dst)
kernel.requires_grad_(requires_grad)
affine.requires_grad_(not requires_grad)
assert gradcheck(kornia.warp_points_tps, (src, dst, kernel, affine), raise_exception=True)
def test_no_warp(self, batch_size, device, dtype):
src = torch.rand(batch_size, 5, 2, device=device)
kernel, affine = kornia.get_tps_transform(src, src)
target_kernel = torch.zeros(batch_size, 5, 2, device=device)
target_affine = torch.zeros(batch_size, 3, 2, device=device)
target_affine[:, [1, 2], [0, 1]] = 1.
assert_allclose(kernel, target_kernel, atol=1e-4, rtol=1e-4)
assert_allclose(affine, target_affine, atol=1e-4, rtol=1e-4)
def test_gradcheck(self, batch_size, device, dtype):
opts = dict(device=device, dtype=torch.float64)
src, dst = _sample_points(batch_size, **opts)
kernel, affine = kornia.get_tps_transform(src, dst)
image = torch.rand(batch_size, 3, 8, 8, requires_grad=True, **opts)
assert gradcheck(
kornia.warp_image_tps, (image, dst, kernel, affine), raise_exception=True, atol=1e-4, rtol=1e-4
)
def test_warp(self, batch_size, device, dtype):
src = torch.tensor([[[-1.0, -1.0], [-1.0, 1.0], [1.0, -1.0],
[1.0, -1.0], [0.0, 0.0]]],
device=device).repeat(batch_size, 1, 1)
# zoom in by a factor of 2
dst = src.clone() * 2.0
tensor = torch.zeros(batch_size, 3, 8, 8, device=device)
tensor[:, :, 2:6, 2:6] = 1.0
expected = torch.ones_like(tensor)
# nn.grid_sample interpolates the at the edges it seems, so the boundaries have values < 1
expected[:, :, [0, -1], :] *= 0.5
expected[:, :, :, [0, -1]] *= 0.5
kernel, affine = kornia.get_tps_transform(dst, src)
warp = kornia.warp_image_tps(tensor, src, kernel, affine)
assert_allclose(warp, expected, atol=1e-4, rtol=1e-4)
def test_warp(self, batch_size, device, dtype):
src, dst = _sample_points(batch_size, device)
kernel, affine = kornia.get_tps_transform(src, dst)
warp = kornia.warp_points_tps(src, dst, kernel, affine)
assert_allclose(warp, dst, atol=1e-4, rtol=1e-4)
def test_smoke(self, batch_size, device, dtype):
src, dst = _sample_points(batch_size, device)
kernel, affine = kornia.get_tps_transform(src, dst)
warp = kornia.warp_points_tps(src, dst, kernel, affine)
assert warp.shape == src.shape
def test_smoke(self, batch_size, device, dtype):
src = torch.rand(batch_size, 4, 2, device=device)
out = kornia.get_tps_transform(src, src)
assert len(out) == 2
def test_smoke(self, batch_size, device, dtype):
src, dst = _sample_points(batch_size, device)
tensor = torch.rand(batch_size, 3, 32, 32, device=device)
kernel, affine = kornia.get_tps_transform(src, dst)
warp = kornia.warp_image_tps(tensor, dst, kernel, affine)
assert warp.shape == tensor.shape
