def test_linear_algebra(self):
a_scalar = AD(1, [1., 0])
b_scalar = AD(2, [0, 1.])
A = np.array([[a_scalar, a_scalar]])
B = np.array([[b_scalar, b_scalar]]).T
C = np.dot(A, B)
npc.assert_equal(C, [[AD(4, [4., 2])]])
# `matmul` not supported for `dtype=object` (#11332). `np.dot` should
# be used instead.
with self.assertRaises(TypeError):
C2 = np.matmul(A, B)
# Type mixing
Bf = np.array([[2., 2]]).T
C2 = np.dot(A, Bf) # Leverages implicit casting.
npc.assert_equal(C2, [[AD(4, [4., 0])]])
# Other methods.
X = np.array([[a_scalar, b_scalar], [b_scalar, a_scalar]])
npc.assert_equal(np.trace(X), AD(2, [2., 0]))
# `inv` is a ufunc that we must implement, if possible. However, given
# that this is currently `dtype=object`, it would be extremely unwise
# to do so. See #8116 for alternative.
with self.assertRaises(TypeError):
Y = np.linalg.inv(X)
# Use workaround for inverse. For now, just check values.
X_float = npc.to_float(X)
Xinv_float = np.linalg.inv(X_float)
Xinv = drake_math.inv(X)
np.testing.assert_equal(npc.to_float(Xinv), Xinv_float)
def test_linear_algebra(self):
a_scalar = AD(1, [1., 0])
b_scalar = AD(2, [0, 1.])
A = np.array([[a_scalar, a_scalar]])
B = np.array([[b_scalar, b_scalar]]).T
C = np.dot(A, B)
npc.assert_equal(C, [[AD(4, [4., 2])]])
# `matmul` not supported for `dtype=object` (#11332). `np.dot` should
# be used instead.
with self.assertRaises(TypeError):
C2 = np.matmul(A, B)
# Type mixing
Bf = np.array([[2., 2]]).T
C2 = np.dot(A, Bf) # Leverages implicit casting.
npc.assert_equal(C2, [[AD(4, [4., 0])]])
# Other methods.
X = np.array([[a_scalar, b_scalar], [b_scalar, a_scalar]])
npc.assert_equal(np.trace(X), AD(2, [2., 0]))
# `inv` is a ufunc that we must implement, if possible. However, given
# that this is currently `dtype=object`, it would be extremely unwise
# to do so. See #8116 for alternative.
with self.assertRaises(TypeError):
Y = np.linalg.inv(X)
# Use workaround for inverse. For now, just check values.
X_float = npc.to_float(X)
Xinv_float = np.linalg.inv(X_float)
Xinv = drake_math.inv(X)
np.testing.assert_equal(npc.to_float(Xinv), Xinv_float)
def check_overload(self, overload):
# TODO(eric.cousineau): Consider comparing against `numpy` ufunc
# methods.
import pydrake.math as drake_math
unary = [
(drake_math.log, math.log),
(drake_math.abs, math.fabs),
(drake_math.exp, math.exp),
(drake_math.sqrt, math.sqrt),
(drake_math.sin, math.sin),
(drake_math.cos, math.cos),
(drake_math.tan, math.tan),
(drake_math.asin, math.asin),
(drake_math.acos, math.acos),
(drake_math.atan, math.atan),
(drake_math.sinh, math.sinh),
(drake_math.cosh, math.cosh),
(drake_math.tanh, math.tanh),
(drake_math.ceil, math.ceil),
(drake_math.floor, math.floor),
]
binary = [
(drake_math.min, min),
(drake_math.max, max),
(drake_math.pow, pow),
(drake_math.atan2, math.atan2),
]
# Arbitrary values to test overloads with.
args_float_all = [0.1, 0.2]
def check_eval(functions, nargs):
# Generate arguments.
args_float = args_float_all[:nargs]
args_T = list(map(overload.to_type, args_float))
# Check each supported function.
for f_drake, f_builtin in functions:
if not overload.supports(f_drake):
continue
debug_print("- Functions: ", qualname(f_drake),
qualname(f_builtin))
y_builtin = f_builtin(*args_float)
y_float = f_drake(*args_float)
debug_print(" - - Float Eval:", repr(y_builtin), repr(y_float))
self.assertEqual(y_float, y_builtin)
self.assertIsInstance(y_float, float)
# Test method current overload, and ensure value is accurate.
y_T = f_drake(*args_T)
y_T_float = overload.to_float(y_T)
debug_print(" - - Overload Eval:", repr(y_T), repr(y_T_float))
self.assertIsInstance(y_T, overload.T)
# - Ensure the translated value is accurate.
self.assertEqual(y_T_float, y_float)
debug_print("\n\nOverload: ", qualname(type(overload)))
float_overload = FloatOverloads()
# Check each number of arguments.
debug_print("Unary:")
check_eval(unary, 1)
debug_print("Binary:")
check_eval(binary, 2)
# Check specialized linear / array algebra.
if overload.supports(drake_math.inv):
f_drake, f_builtin = drake_math.inv, np.linalg.inv
X_float = np.eye(2)
Y_builtin = f_builtin(X_float)
Y_float = f_drake(X_float)
self.assertIsInstance(Y_float[0, 0].item(), float)
np.testing.assert_equal(Y_builtin, Y_float)
to_type_array = np.vectorize(overload.to_type)
to_float_array = np.vectorize(overload.to_float)
X_T = to_type_array(X_float)
Y_T = drake_math.inv(X_T)
self.assertIsInstance(Y_T[0, 0], overload.T)
np.testing.assert_equal(to_float_array(Y_T), Y_float)
def check_overload(self, overload):
# TODO(eric.cousineau): Consider comparing against `numpy` ufunc
# methods.
import pydrake.math as drake_math
unary = [
(drake_math.log, math.log),
(drake_math.abs, math.fabs),
(drake_math.exp, math.exp),
(drake_math.sqrt, math.sqrt),
(drake_math.sin, math.sin),
(drake_math.cos, math.cos),
(drake_math.tan, math.tan),
(drake_math.asin, math.asin),
(drake_math.acos, math.acos),
(drake_math.atan, math.atan),
(drake_math.sinh, math.sinh),
(drake_math.cosh, math.cosh),
(drake_math.tanh, math.tanh),
(drake_math.ceil, math.ceil),
(drake_math.floor, math.floor),
]
binary = [
(drake_math.min, min),
(drake_math.max, max),
(drake_math.pow, pow),
(drake_math.atan2, math.atan2),
]
# Arbitrary values to test overloads with.
args_float_all = [0.1, 0.2]
def check_eval(functions, nargs):
# Generate arguments.
args_float = args_float_all[:nargs]
args_T = list(map(overload.to_type, args_float))
# Check each supported function.
for f_drake, f_builtin in functions:
if not overload.supports(f_drake):
continue
debug_print(
"- Functions: ", qualname(f_drake), qualname(f_builtin))
y_builtin = f_builtin(*args_float)
y_float = f_drake(*args_float)
debug_print(" - - Float Eval:", repr(y_builtin), repr(y_float))
self.assertEqual(y_float, y_builtin)
self.assertIsInstance(y_float, float)
# Test method current overload, and ensure value is accurate.
y_T = f_drake(*args_T)
y_T_float = overload.to_float(y_T)
debug_print(" - - Overload Eval:", repr(y_T), repr(y_T_float))
self.assertIsInstance(y_T, overload.T)
# - Ensure the translated value is accurate.
self.assertEqual(y_T_float, y_float)
debug_print("\n\nOverload: ", qualname(type(overload)))
float_overload = FloatOverloads()
# Check each number of arguments.
debug_print("Unary:")
check_eval(unary, 1)
debug_print("Binary:")
check_eval(binary, 2)
# Check specialized linear / array algebra.
if overload.supports(drake_math.inv):
f_drake, f_builtin = drake_math.inv, np.linalg.inv
X_float = np.eye(2)
Y_builtin = f_builtin(X_float)
Y_float = f_drake(X_float)
self.assertIsInstance(Y_float[0, 0].item(), float)
np.testing.assert_equal(Y_builtin, Y_float)
to_type_array = np.vectorize(overload.to_type)
to_float_array = np.vectorize(overload.to_float)
X_T = to_type_array(X_float)
Y_T = drake_math.inv(X_T)
self.assertIsInstance(Y_T[0, 0], overload.T)
np.testing.assert_equal(to_float_array(Y_T), Y_float)
