def test_grads_no_error_correct_pred(self):
"""
Case in which the prediction is correct.
All learnable parameters should have a gradient value
that does not cause errors with an optimizer.
Passes if no errors occurs.
"""
node = gen_node_at(ZERO)
marble = gen_marble_at(ZERO)
model = NenwinModel([node], [marble])
model.make_timestep(1.0)
optim = torch.optim.Adam(model.parameters())
optim.zero_grad()
loss_fun = NenwinLossFunction([node], model, 0, 1)
target_index = 0
result = loss_fun(target_index)
try:
result.backward()
optim.step()
except RuntimeError as e:
self.fail(f"Error occurred during backprop/optim step: {e}")
def test_reset(self):
# Node and Marble have some arbitrary nonzero initial motion vars.
node = Node(torch.tensor([1.]), torch.tensor([2.]), torch.tensor([3.]),
4, NewtonianGravity(), 1, 1, 1, 1)
marble = Marble(torch.tensor([1.1]), torch.tensor([2.2]),
torch.tensor([3.3]), 4.4, NewtonianGravity(), None)
model = NenwinModel([node], [marble])
model.make_timestep(10)
model.make_timestep(10)
model.make_timestep(10)
# Verify that the motion variables have changed
self.assertFalse(check_close(node.pos, node.init_pos))
self.assertFalse(check_close(marble.vel, marble.init_vel))
self.assertFalse(check_close(marble.acc, marble.init_acc))
model.reset()
# Now they should be the original values again.
self.assertTrue(check_close(node.pos, node.init_pos))
self.assertTrue(check_close(node.vel, node.init_vel))
self.assertTrue(check_close(node.acc, node.init_acc))
self.assertTrue(check_close(marble.pos, marble.init_pos))
self.assertTrue(check_close(marble.vel, marble.init_vel))
self.assertTrue(check_close(marble.acc, marble.init_acc))
def test_eaten_marbles_disappear(self):
# Both are generated at the smame pos, so immediately eaten.
node = MarbleEaterNode(ZERO, ZERO, ZERO, 10, NewtonianGravity(), 1, 1,
1, 0, 10)
marble = Marble(ZERO, ZERO, ZERO, 10, NewtonianGravity(), datum=None)
model = NenwinModel([node], [marble])
model.make_timestep(1.0)
self.assertSetEqual(model.marbles, set([]))
self.assertNotIn(marble, model._NenwinModel__all_particles)
def test_make_timestep_1(self):
"""
Base base: single moving Marble, with constant velocity.
"""
marble = Marble(ZERO, np.array([10]), ZERO, 1, None, None, 0, 0, 0, 0)
model = NenwinModel([], [marble])
model.make_timestep(time_passed=1)
expected_new_pos = torch.tensor([10], dtype=torch.float)
self.assertTrue(torch.allclose(marble.pos, expected_new_pos))
def test_make_timestep_2(self):
"""
Base case: initial stationairy Marble, gets attracted and accelerated.
"""
node = Node(pos=ZERO,
vel=ZERO,
acc=ZERO,
mass=1,
attraction_function=ATTRACT_FUNCT,
marble_stiffness=1,
node_stiffness=1,
marble_attraction=1,
node_attraction=0)
marble = Marble(pos=np.array([5]),
vel=ZERO,
acc=ZERO,
mass=1,
attraction_function=ATTRACT_FUNCT,
datum=None)
model = NenwinModel([node], [marble])
time_passed = 1
expected_pos, expected_vel = runge_kutta_4_step(marble.pos,
marble.vel,
-ATTRACT_FUNCT.value,
duration=time_passed)
model.make_timestep(time_passed)
self.assertTrue(torch.allclose(marble.pos, expected_pos, atol=0.01))
self.assertTrue(torch.allclose(marble.vel, expected_vel, atol=0.01))
self.assertTrue(
torch.isclose(marble.acc,
torch.tensor(-ATTRACT_FUNCT.value),
atol=0.01))
self.assertTrue(torch.allclose(node.pos, ZERO))
self.assertTrue(torch.allclose(node.vel, ZERO))
self.assertTrue(torch.allclose(node.acc, ZERO))
def test_grads_value_correct_pred(self):
"""
Case in which the prediction is correct.
All learnable parameters should have a gradient value
that does not affect the values of the weights.
"""
node = gen_node_at(ZERO)
marble = gen_marble_at(torch.tensor([1.1, 1.1]))
model = NenwinModel([node], [marble])
for _ in range(1000):
model.make_timestep(0.1)
assert node.num_marbles_eaten == 1, "Testcase badly desgined."
loss_fun = NenwinLossFunction([node], model, 0, 1)
self.target_index = 0 # Node at index 1 ate the Marble
result = loss_fun(self.target_index)
result.backward(retain_graph=True)
self.assertIsNone(marble.init_pos.grad)
self.assertIsNone(marble.init_vel.grad)
self.assertIsNone(marble.mass.grad)
class ModelBackpropTestCase(unittest.TestCase):
"""
Testcase tests if backpropagation correctly behaves for particle
interaction within a model.
"""
def setUp(self):
self.node = Node(pos=ZERO,
vel=ZERO,
acc=ZERO,
mass=1,
attraction_function=NewtonianGravity(),
marble_stiffness=1,
node_stiffness=1,
marble_attraction=1,
node_attraction=0)
self.marble = Marble(pos=np.array([5]),
vel=ZERO,
acc=ZERO,
mass=1,
attraction_function=ATTRACT_FUNCT,
datum=None)
self.model = NenwinModel([self.node], [self.marble])
def test_particle_gradients(self):
"""
Base case: a Marble being attracted by a stationary Node,
when using Marble's variables to compute loss,
also Node.pos should receive gradients
when computing backprop on the loss.
"""
self.model.make_timestep(1.0)
self.model.make_timestep(1.0)
loss = 2 * self.marble.pos
loss.backward()
self.assertIsNotNone(self.node.init_pos.grad)
def test_particle_gradients_moving_node(self):
"""
Base case: Marble being attracted by a moving Node,
when using Marble's variables to compute loss,
also Node.pos should receive gradients
when computing backprop on the loss.
"""
self.node = Node(pos=ZERO,
vel=torch.tensor([0.1]),
acc=ZERO,
mass=1,
attraction_function=NewtonianGravity(),
marble_stiffness=1,
node_stiffness=1,
marble_attraction=1,
node_attraction=0)
self.model = NenwinModel([self.node], [self.marble])
self.model.make_timestep(1.0)
self.model.make_timestep(1.0)
self.model.make_timestep(1.0)
loss = 2 * self.marble.acc
loss.backward()
self.assertIsNotNone(self.marble.init_pos.grad)
self.assertIsNotNone(self.node.init_pos.grad)
self.assertIsNotNone(self.node.init_vel.grad)
self.assertIsNotNone(self.node._PhysicalParticle__mass.grad)
class LossFunctionMultiplePredAllWrongTestCase(unittest.TestCase):
"""
Testcases for NenwinLossFunction.__call__()
when multiple output-MarbleEaterNodes have eaten a Marble.
The losses should count up.
"""
def setUp(self):
"""
Sketch:
|
↑| M_1 M_2 M_3
y| ↓ ↓ ↓
|
0| N_1 N_2 N_0
|
|
+-------------------------------
-10 0 10 x→
Marbles M_1 and M_2 are moving towards
MarbleEaterNodes N_1 and N_2 respectively.
They arrive at the same moment.
Only N_0 is the target output,
which will not receive any Marble.
M_3 remains stationary.
"""
self.pos_weight = 0.5
self.vel_weight = 0.5
self.nodes = (gen_node_at(torch.tensor([10.0, 0])),
gen_node_at(torch.tensor([-10.0, 0])),
gen_node_at(torch.tensor([0.0, 0]))
)
self.marble_1 = gen_marble_at(torch.tensor([-10.0, 10.0]),
vel=torch.tensor([0, -3.0]),
datum="M_1")
self.marble_2 = gen_marble_at(torch.tensor([0.0, 10.0]),
vel=torch.tensor([0, -3.0]),
datum="M_2")
self.marble_3 = gen_marble_at(torch.tensor([10.0, 10.0]),
vel=torch.tensor([0, 0.0]),
datum="M_3")
marbles = [self.marble_1, self.marble_2, self.marble_3]
self.model = NenwinModel(self.nodes, marbles)
for _ in range(50): # Should be enough for the Marble to be eaten
self.model.make_timestep(0.1)
assert len(self.model.marbles) == 1, "Testcase badly designed."
assert self.nodes[0].num_marbles_eaten == 0, "Testcase badly designed."
self.loss_fun = NenwinLossFunction(self.nodes, self.model,
vel_weight=self.vel_weight,
pos_weight=self.pos_weight)
self.target_index = 0
self.wrong_node_index = 1
self.loss = self.loss_fun(self.target_index)
def test_value_loss_some_marble_left(self):
target_node = self.nodes[0]
# Penalty for the closest Marble that did not arrive at the target
expected = velocity_weighted_distance(target_node, self.marble_3,
pos_weight=self.pos_weight,
vel_weight=self.vel_weight)
# Penalties for wrong outputs
for marble, node in zip([self.marble_1, self.marble_2], self.nodes[1:]):
expected += -1/velocity_weighted_distance(node, marble,
pos_weight=self.pos_weight,
vel_weight=self.vel_weight)
torch.testing.assert_allclose(self.loss, expected)
class LossFunctionMultiplePredInclCorrectTestCase(unittest.TestCase):
"""
Testcases for NenwinLossFunction.__call__()
when multiple output-MarbleEaterNodes have eaten a Marble.
The losses should count up.
"""
def setUp(self):
"""
Sketch:
|
↑| M_1 M_3
y| ↓ ↓
|
0| N_1 ←M_2 N_0
|
|
+-------------------------------
-10 0 10 x→
Marbles M_1 and M_2 are moving towards N_1,
but should arrive at N_0.
M_3 arrives correctly at N_0, at the same moment.
"""
self.pos_weight = 0.5
self.vel_weight = 0.5
self.nodes = (gen_node_at(torch.tensor([10.0, 0])),
gen_node_at(torch.tensor([-10.0, 0])))
self.marble_1 = gen_marble_at(ZERO,
vel=torch.tensor([-3.0, 0]),
datum="M_1")
self.marble_2 = gen_marble_at(torch.tensor([-10.0, 10.0]),
vel=torch.tensor([0, -3.0]),
datum="M_2")
self.marble_3 = gen_marble_at(torch.tensor([10.0, 10.0]),
vel=torch.tensor([0, -3.0]),
datum="M_3")
marbles = [self.marble_1, self.marble_2, self.marble_3]
self.model = NenwinModel(self.nodes, marbles)
for _ in range(50): # Should be enough for the Marble to be eaten
self.model.make_timestep(0.1)
assert len(self.model.marbles) == 0, "Testcase badly designed."
self.loss_fun = NenwinLossFunction(self.nodes, self.model,
vel_weight=self.vel_weight,
pos_weight=self.pos_weight)
self.target_index = 0
self.wrong_node_index = 1
self.loss = self.loss_fun(self.target_index)
def test_value_loss_no_marble_left(self):
"""
For the two wrong Marbles (M_1 and M_2),
their loss should equal the velocity-weighted distance
of the target Node to the nearest Marble,
plus the *negative reciprocal* of the distance of the wrong node
(self.nodes[1]) to the Marble.
For the correcly arrived Marble, the loss should equal 0.
Case where no non-eaten Marble is available
at time of loss computation.
Not that this should matter, as a correct Marble was eaten!
"""
wrong_node = self.nodes[self.wrong_node_index]
expected = 0.0
for marble in [self.marble_1, self.marble_2]:
expected -= 1/velocity_weighted_distance(wrong_node, marble,
pos_weight=self.pos_weight,
vel_weight=self.vel_weight)
torch.testing.assert_allclose(self.loss, expected)
class LossFunctionCallWrongPredTestCase(unittest.TestCase):
"""
Testcases for NenwinLossFunction.__call__() for when another than the target
output-MarbleEaterNodes has eaten a Marble.
"""
def setUp(self):
"""
Sketch:
|
^|
y|
|
0| N_1 <M N_0
|
|
+-------------------------------
-10 0 10 x>
Marble M starts with moving towards N_1, but should arrive at N_0.
"""
self.pos_weight = 0.5
self.vel_weight = 0.5
self.nodes = (gen_node_at(torch.tensor([10.0, 0])),
gen_node_at(torch.tensor([-10.0, 0])))
self.marble = gen_marble_at(ZERO,
vel=torch.tensor([-3.0, 0]),
datum="original")
self.model = NenwinModel(self.nodes, [self.marble])
for _ in range(50): # Should be enough for the Marble to be eaten
self.model.make_timestep(0.1)
assert len(self.model.marbles) == 0, "Testcase badly designed."
self.loss_fun = NenwinLossFunction(self.nodes, self.model,
vel_weight=self.vel_weight,
pos_weight=self.pos_weight)
self.target_index = 0
self.wrong_node_index = 1
self.loss = self.loss_fun(self.target_index)
def test_value_loss_wrong_pred_no_marble_left(self):
"""
The loss should equal the velocity-weighted distance
of the target Node to the nearest Marble,
plus the *negative reciprocal* of the distance of the wrong node
(self.nodes[1]) to the Marble.
Case where no non-eaten Marble is available.
"""
target_node = self.nodes[self.target_index]
wrong_node = self.nodes[self.wrong_node_index]
expected = velocity_weighted_distance(target_node, self.marble,
pos_weight=self.pos_weight,
vel_weight=self.vel_weight)
expected -= 1/velocity_weighted_distance(wrong_node, self.marble,
pos_weight=self.pos_weight,
vel_weight=self.vel_weight)
torch.testing.assert_allclose(self.loss, expected)
def test_value_loss_wrong_pred_some_marble_left(self):
"""
The loss should equal the velocity-weighted distance
of the target Node to the nearest Marble,
plus the *negative reciprocal* of the distance of the wrong node
(self.nodes[1]) to the Marble.
Case where another non-eaten Marble is still available
at time of loss computation.
"""
second_marble = gen_marble_at(ZERO, datum="second")
self.model.add_marbles([second_marble])
self.loss = self.loss_fun(self.target_index)
target_node = self.nodes[self.target_index]
wrong_node = self.nodes[self.wrong_node_index]
expected = velocity_weighted_distance(target_node, second_marble,
pos_weight=self.pos_weight,
vel_weight=self.vel_weight)
expected -= 1/velocity_weighted_distance(wrong_node, self.marble,
pos_weight=self.pos_weight,
vel_weight=self.vel_weight)
torch.testing.assert_allclose(self.loss, expected)
def test_grads_no_error_wrong_pred(self):
"""
Case in which no prediction output is given.
All learnable parameters should have a gradient value
that does not cause errors with an optimizer.
Case where another non-eaten Marble is still available
at time of loss computation.
"""
second_marble = gen_marble_at(ZERO)
self.model.add_marbles([second_marble])
optim = torch.optim.Adam(self.model.parameters())
optim.zero_grad()
try:
self.loss.backward()
optim.step()
except RuntimeError as e:
self.fail(f"Error occurred during backprop/optim step: {e}")
def test_grads_value_wrong_pred(self):
"""
Case in which no prediction output is given.
The learnable parameters of the Marble should have a gradient value
that does affect the values of the weights.
The loss should be lower for the second run.
"""
second_marble = gen_marble_at(ZERO)
self.model.add_marbles([second_marble])
optim = torch.optim.Adam(self.model.parameters())
optim.zero_grad()
self.loss.backward()
self.assertIsNotNone(self.marble.init_pos.grad)
self.assertIsNotNone(self.marble.init_vel.grad)
self.assertIsNotNone(self.marble.mass.grad)
optim.step()
# Now verify the loss improved.
self.model.reset()
self.model.make_timestep(0.1)
self.model.add_marbles([second_marble])
new_loss = self.loss_fun(self.target_index)
self.assertLess(new_loss.item(), self.loss.item())
class LossFunctionCallNoPredTestCase(unittest.TestCase):
"""
Testcases for NenwinLossFunction.__call__() in case none of the
output-MarbleEaterNodes has eaten any Marble.
"""
def setUp(self):
self.pos_weight = 0.5
self.vel_weight = 0.5
self.node = gen_node_at(ZERO)
self.marble = gen_marble_at(torch.tensor([10., 10.]))
self.model = NenwinModel([self.node], [self.marble])
self.model.make_timestep(0.1) # too short to cross the distance
self.loss_fun = NenwinLossFunction([self.node], self.model,
vel_weight=self.vel_weight,
pos_weight=self.pos_weight)
assert self.node.num_marbles_eaten == 0, "Testcase badly desgined."
self.target_index = 0
self.loss = self.loss_fun(self.target_index)
def test_value_loss_no_output(self):
"""
If no Marble has been eaten, the loss
should equal the velocity-weighted distance
of the target Node to the nearest Marble.
"""
expected = velocity_weighted_distance(self.node, self.marble,
pos_weight=self.pos_weight,
vel_weight=self.vel_weight)
torch.testing.assert_allclose(self.loss, expected)
def test_grads_no_error_no_output(self):
"""
Case in which no prediction output is given.
All learnable parameters should have a gradient value
that does not cause errors with an optimizer.
Passes if no errors occurs.
"""
optim = torch.optim.Adam(self.model.parameters())
optim.zero_grad()
try:
self.loss.backward()
optim.step()
except RuntimeError as e:
self.fail(f"Error occurred during backprop/optim step: {e}")
def test_grads_value_no_output(self):
"""
Case in which no prediction output is given.
The learnable parameters of the Marble should have a gradient value
that does affect the values of the weights.
The loss should be lower for the second run.
"""
optim = torch.optim.Adam(self.model.parameters())
optim.zero_grad()
self.loss.backward()
self.assertIsNotNone(self.marble.init_pos.grad)
self.assertIsNotNone(self.marble.init_vel.grad)
self.assertIsNotNone(self.marble.mass.grad)
optim.step()
# Now verify the loss improved.
self.model.reset()
self.model.make_timestep(0.1)
new_loss = self.loss_fun(self.target_index)
self.assertLess(new_loss.item(), self.loss.item())
