def test_jit(self, device, dtype):
B, C, H, W = 1, 1, 13, 13
inp = torch.zeros(B, C, H, W, device=device)
inp[:, :, 15:-15, 9:-9] = 1
laf = torch.tensor([[[[20.0, 0.0, 16.0], [0.0, 20.0, 16.0]]]],
device=device)
tfeat = LAFAffineShapeEstimator(W).to(inp.device, inp.dtype).eval()
tfeat_jit = torch.jit.script(
LAFAffineShapeEstimator(W).to(inp.device, inp.dtype).eval())
assert_close(tfeat_jit(laf, inp), tfeat(laf, inp))
def test_toy(self, device):
aff = LAFAffineShapeEstimator(32).to(device)
inp = torch.zeros(1, 1, 32, 32, device=device)
inp[:, :, 15:-15, 9:-9] = 1
laf = torch.tensor([[[[20.0, 0.0, 16.0], [0.0, 20.0, 16.0]]]],
device=device)
new_laf = aff(laf, inp)
expected = torch.tensor([[[[36.643, 0.0, 16.0], [0.0, 10.916, 16.0]]]],
device=device)
assert_close(new_laf, expected, atol=1e-4, rtol=1e-4)
def test_toy_preserve(self, device, dtype):
aff = LAFAffineShapeEstimator(32, preserve_orientation=True).to(
device, dtype)
inp = torch.zeros(1, 1, 32, 32, device=device, dtype=dtype)
inp[:, :, 15:-15, 9:-9] = 1
laf = torch.tensor([[[[0.0, 20.0, 16.0], [-20.0, 0.0, 16.0]]]],
device=device,
dtype=dtype)
new_laf = aff(laf, inp)
expected = torch.tensor([[[[0.0, 36.643, 16.0], [-10.916, 0, 16.0]]]],
device=device,
dtype=dtype)
assert_close(new_laf, expected, atol=1e-4, rtol=1e-4)
def test_gradcheck(self, device):
batch_size, channels, height, width = 1, 1, 40, 40
patches = torch.rand(batch_size,
channels,
height,
width,
device=device)
patches = utils.tensor_to_gradcheck_var(patches) # to var
laf = torch.tensor([[[[5.0, 0.0, 26.0], [0.0, 5.0, 26.0]]]],
device=device)
laf = utils.tensor_to_gradcheck_var(laf) # to var
assert gradcheck(
LAFAffineShapeEstimator(11).to(device),
(laf, patches),
raise_exception=True,
rtol=1e-3,
atol=1e-3,
nondet_tol=1e-4,
)
def test_print(self, device):
sift = LAFAffineShapeEstimator()
sift.__repr__()
def test_shape_batch(self, device):
inp = torch.rand(2, 1, 32, 32, device=device)
laf = torch.rand(2, 34, 2, 3, device=device)
ori = LAFAffineShapeEstimator().to(device)
out = ori(laf, inp)
assert out.shape == laf.shape