def setUp(self):
self.c1 = core.Constant([2.1, 3], trainable=False)
self.c2 = core.Constant([4.5, 0.8], trainable=False)
self.n_x = 10
self.x = core.Variable('x', np.random.normal(0., 1., (self.n_x, 2)))
self.n_y = 5
self.y = core.Variable('y', np.random.normal(0., 4., (self.n_y, 2)))
def test_from_trainable_constants(self):
c1 = core.Constant([2.1, 3], trainable=True)
c2 = core.Constant([4.5, 0.8], trainable=True)
with tf.GradientTape() as tape:
x = core.Variable.from_constants('x', [c1, c2], tape=tape)
self.assertTrue(tape.gradient(x.tensor, c1.tensor) is not None)
tape = tf.GradientTape()
with self.assertRaises(ValueError):
x = core.Variable.from_constants('x', [c1, c2], tape=tape)
def test_trainable(self):
for c_val in self.c_vals:
c = core.Constant(c_val, trainable=True)
self.assertTrue(array_allclose(c.tensor.numpy(), c_val))
self.assertTrue(isinstance(c.tensor, tf.Variable))
self.assertTrue(c.tensor.trainable)
self.assertEqual(c.free_vars, [])
def test_init_from_tf(self):
for c_val in self.c_vals:
c_val = tf.constant(c_val)
c = core.Constant(c_val, trainable=False)
self.assertTrue(array_allclose(c.tensor.numpy(), c_val))
self.assertTrue(isinstance(c.tensor, tf.Tensor))
self.assertEqual(c.free_vars, [])
def test_gradients(self):
"""Tracks gradients"""
c_s = [
core.Constant(np.random.rand(2), trainable=True)
for _ in range(self.n_x)
]
with tf.GradientTape() as tape:
x = core.Variable.from_constants('x', c_s, tape)
phi = self.p1([x, self.y])
res = self.Forall(x, phi)
self.assertTrue(tape.gradient(res.tensor, c_s[0].tensor) is not None)
def setUp(self):
self.And = core.Wrapper_Connective(fuzzy_ops.And_Prod())
self.Not = core.Wrapper_Connective(fuzzy_ops.Not_Std())
self.c1 = core.Constant([2.1, 3], trainable=True)
self.n_x = 10
self.x = core.Variable('x', np.random.normal(0., 1., (self.n_x, 2)))
self.n_y = 5
self.y = core.Variable('y', np.random.normal(0., 4., (self.n_y, 2)))
self.p1 = core.Predicate.MLP([2])
self.p2 = core.Predicate.MLP([2, 2])
self.a = core.Proposition(0., trainable=True)
def test_free_dims_wt_broadcast_constant(self):
"""Adds the correct dimensions if a constant is involved"""
c1 = core.Constant([2.1, 3], trainable=False)
exprs = core.broadcast_exprs(list(self.xs.values()) + [c1])
for expr in exprs:
self.assertEqual(sorted(expr.free_vars),
sorted(self.var_settings.keys()))
for label, v_s in self.var_settings.items():
self.assertEqual(
expr._get_dim_of_free_var(label).numpy(),
v_s["n_individuals"])
def setUp(self):
self.x = core.Variable('x', np.random.rand(3, 1))
self.y = core.Variable('y', np.random.rand(4, 1))
self.c = core.Constant([3.], trainable=False)
self.f1 = core.Function.MLP(input_shapes=[1], output_shape=[1])
self.f2 = core.Function.MLP(input_shapes=[1, 1], output_shape=[1])
self.p1 = core.Predicate.MLP(input_shapes=[1])
self.p2 = core.Predicate.MLP(input_shapes=[1, 1])
self.q = core.Proposition(0., trainable=False)
self.And = core.Wrapper_Connective(fuzzy_ops.And_Prod())
self.Not = core.Wrapper_Connective(fuzzy_ops.Not_Std())
self.Exists = core.Wrapper_Quantifier(fuzzy_ops.Aggreg_Mean(),
semantics="exists")
self.mask = core.Formula(tf.constant([[1., 1., 0.,
0.], [0., 1., 1., 0.],
[0., 0., 1., 0.]]),
free_vars=['x', 'y'])
def test_from_non_trainable_constants(self):
c1 = core.Constant([2.1, 3], trainable=False)
c2 = core.Constant([4.5, 0.8], trainable=False)
with tf.GradientTape() as tape:
x = core.Variable.from_constants('x', [c1, c2], tape=None)
self.assertTrue(tape.gradient(x.tensor, c1.tensor) is None)