def test_api_dygraph(self):
paddle.disable_static(self.place)
x_tensor = paddle.to_tensor(self.x)
out = paddle.diagonal(x_tensor)
out_ref = np.diagonal(self.x)
self.assertEqual(np.allclose(out.numpy(), out_ref, rtol=1e-08), True)
paddle.enable_static()
def gwd_loss(self,
pred,
target,
fun='log',
tau=1.0,
alpha=1.0,
normalize=False):
xy_p, R_p, S_p = self.xywhr2xyrs(pred)
xy_t, R_t, S_t = self.xywhr2xyrs(target)
xy_distance = (xy_p - xy_t).square().sum(axis=-1)
Sigma_p = R_p.matmul(S_p.square()).matmul(R_p.transpose([0, 2, 1]))
Sigma_t = R_t.matmul(S_t.square()).matmul(R_t.transpose([0, 2, 1]))
whr_distance = paddle.diagonal(
S_p, axis1=-2, axis2=-1).square().sum(axis=-1)
whr_distance = whr_distance + paddle.diagonal(
S_t, axis1=-2, axis2=-1).square().sum(axis=-1)
_t = Sigma_p.matmul(Sigma_t)
_t_tr = paddle.diagonal(_t, axis1=-2, axis2=-1).sum(axis=-1)
_t_det_sqrt = paddle.diagonal(S_p, axis1=-2, axis2=-1).prod(axis=-1)
_t_det_sqrt = _t_det_sqrt * paddle.diagonal(
S_t, axis1=-2, axis2=-1).prod(axis=-1)
whr_distance = whr_distance + (-2) * (
(_t_tr + 2 * _t_det_sqrt).clip(0).sqrt())
distance = (xy_distance + alpha * alpha * whr_distance).clip(0)
if normalize:
wh_p = pred[..., 2:4].clip(min=1e-7, max=1e7)
wh_t = target[..., 2:4].clip(min=1e-7, max=1e7)
scale = ((wh_p.log() + wh_t.log()).sum(dim=-1) / 4).exp()
distance = distance / scale
if fun == 'log':
distance = paddle.log1p(distance)
if tau >= 1.0:
return 1 - 1 / (tau + distance)
return distance
def test_api_eager(self):
paddle.disable_static(self.place)
with _test_eager_guard():
x_tensor = paddle.to_tensor(self.x)
out = paddle.diagonal(x_tensor)
out2 = paddle.diagonal(x_tensor, offset=0, axis1=2, axis2=1)
out3 = paddle.diagonal(x_tensor, offset=1, axis1=0, axis2=1)
out4 = paddle.diagonal(x_tensor, offset=0, axis1=1, axis2=2)
out_ref = np.diagonal(self.x)
self.assertEqual(np.allclose(out.numpy(), out_ref, rtol=1e-08), True)
out2_ref = np.diagonal(self.x, offset=0, axis1=2, axis2=1)
self.assertEqual(np.allclose(out2.numpy(), out2_ref, rtol=1e-08), True)
out3_ref = np.diagonal(self.x, offset=1, axis1=0, axis2=1)
self.assertEqual(np.allclose(out3.numpy(), out3_ref, rtol=1e-08), True)
out4_ref = np.diagonal(self.x, offset=0, axis1=1, axis2=2)
self.assertEqual(np.allclose(out4.numpy(), out4_ref, rtol=1e-08), True)
paddle.enable_static()
def test_api_static(self):
paddle.enable_static()
with paddle.static.program_guard(paddle.static.Program()):
x = paddle.fluid.data('X', self.shape)
out = paddle.diagonal(x)
exe = paddle.static.Executor(self.place)
res = exe.run(feed={'X': self.x}, fetch_list=[out])
out_ref = np.diagonal(self.x)
for out in res:
self.assertEqual(np.allclose(out, out_ref, rtol=1e-08), True)
def log_det_by_cholesky(self, matrix):
"""
Args:
matrix: matrix must be a positive define matrix.
shape [N, C, D, D].
"""
chol = paddle.cholesky(matrix)
diag = paddle.diagonal(chol, offset=0, axis1=-2, axis2=-1)
chol = paddle.log(diag + 1e-8)
return 2.0 * paddle.sum(chol, axis=-1)
def trace(self, A):
tr = paddle.diagonal(A, axis1=-2, axis2=-1)
tr = paddle.sum(tr, axis=-1)
return tr