def test_always_single_batch_shape(self):
a = ltd.TermDynamic(torch.randn(1, 100, 2))
b = ltd.TermDynamic(torch.randn(1, 100, 2))
con = ltd.Always(ltd.EQ(a, b), 100)
self.assertEqual(con.loss(0).shape, torch.Size([1]))
self.assertEqual(con.satisfy(0).shape, torch.Size([1]))
def test_always_equal(self):
a = ltd.TermDynamic(torch.ones(64, 10000, 7))
b = ltd.TermDynamic(torch.ones(64, 10000, 7))
con = ltd.Always(ltd.EQ(a, b), 10000)
actual_loss = con.loss(0)
self.assertEqual(actual_loss[0], 0.0)
self.assertEqual(con.satisfy(0).all(), True)
def test_always_unequal_huge(self):
a = ltd.TermDynamic(torch.ones(64, 10000, 7))
b = ltd.TermDynamic(torch.ones(64, 10000, 7) + 1000)
con = ltd.Always(ltd.EQ(a, b), 10000)
expected_loss = torch.ones(64) * np.sqrt(7000000)
actual_loss = con.loss(0)
self.assertAlmostEqual(actual_loss[0].item(),
expected_loss[0].item(),
places=1)
self.assertEqual(con.satisfy(0).all(), False)
def test_always_unequal_stress(self):
a = ltd.TermDynamic(torch.ones(64, 10000, 7))
b = ltd.TermDynamic(torch.ones(64, 10000, 7) + 1)
con = ltd.Always(ltd.EQ(a, b), 10000)
expected_loss = torch.ones(64) * np.sqrt(7)
actual_loss = con.loss(0)
# Note that the numerical errors start to mount up
# In the end we can only get it to about 1 decimal place of accuracy
self.assertAlmostEqual(actual_loss[0].item(),
expected_loss[0].item(),
places=4)
self.assertEqual(con.satisfy(0).all(), False)