def test_check_gradient_jacobian_vector_arg():
helpers.check_gradient(
lambda x: numpy.array([x[0]**2 * x[1], 5 * x[0] + math.sin(x[1])]),
lambda x: numpy.array([[2 * x[0] * x[1], x[0]**2],
[5.0, math.cos(x[1])]]),
f_arg_tensor=numpy.random.random(2),
f_shape='jac')
def test_dsoftmax_matrix():
tensor_shape = [random.randint(2, 10) for _ in range(2)]
helpers.check_gradient(lambda X: calculate.softmax(X),
lambda X: calculate.dsoftmax(calculate.softmax(X)),
f_arg_tensor=numpy.random.random(tensor_shape),
f_shape='jac-stack')
def test_check_gradient_scalar_3tensor_arg():
helpers.check_gradient(lambda x: numpy.sum(x**2),
lambda x: 2.0 * x,
f_arg_tensor=numpy.random.random(
(random.randint(1, 10), random.randint(1, 10),
random.randint(1, 10))),
f_shape='scalar')
def test_dlogit_matrix():
tensor_shape = [random.randint(1, 10) for _ in range(2)]
helpers.check_gradient(lambda X: calculate.logit(X),
lambda X: calculate.dlogit(X),
f_arg_tensor=numpy.random.random(tensor_shape),
f_shape='lin')
def test_check_gradient_jacobian():
helpers.check_gradient(
lambda x: numpy.array([x[0]**2 * x[1], 5 * x[0] + math.sin(x[1])]),
lambda x: numpy.array([[2 * x[0] * x[1], x[0]**2],
[5.0, math.cos(x[1])]]),
inputs=numpy.random.rand(2),
f_shape='jac')
def test_check_gradient_jacobian_3tensor_arg():
helpers.check_gradient(lambda x: numpy.array(
[numpy.sum(x**2), numpy.sum(x**3)]),
lambda x: numpy.array([2 * x, 3 * x**2]),
f_arg_tensor=numpy.random.random(
(random.randint(1, 10), random.randint(1, 10),
random.randint(1, 10))),
f_shape='jac')
def check_error_gradient(error_func, tensor_d=1):
tensor_shape = [random.randint(1, 10) for _ in range(tensor_d)]
tensor_b = numpy.random.random(tensor_shape)
helpers.check_gradient(lambda X: error_func(X, tensor_b),
lambda X: error_func.derivative(X, tensor_b)[1],
f_arg_tensor=numpy.random.random(tensor_shape),
f_shape='scalar')
def check_error_gradient(error_func):
vec_length = random.randint(1, 10)
vec_b = numpy.random.random(vec_length)
helpers.check_gradient(
lambda X: error_func(X, vec_b),
lambda X: error_func.derivative(X, vec_b)[1],
inputs=numpy.random.random(vec_length),
f_shape='scalar')
def test_check_gradient_jacobian_matrix_arg_matrix_out():
helpers.check_gradient(
lambda x: numpy.array([[numpy.sum(
x**2), numpy.sum(x**3)], [numpy.sum(x**4),
numpy.sum(x**5)]]),
lambda x: numpy.array([[2 * x, 3 * x**2], [4 * x**3, 5 * x**4]]),
f_arg_tensor=numpy.random.random(
(random.randint(1, 10), random.randint(1, 10))),
f_shape='jac')
def test_check_gradient_jac_stack():
helpers.check_gradient(lambda x: numpy.hstack([
numpy.sum(x**2, axis=1, keepdims=True),
numpy.sum(x**3, axis=1, keepdims=True)
]),
lambda x: numpy.array([[2 * x[i], 3 * x[i]**2]
for i in range(x.shape[0])]),
f_arg_tensor=numpy.random.random(
(random.randint(1, 10), random.randint(1, 10))),
f_shape='jac-stack')
def test_check_gradient_lin_matrix_arg():
helpers.check_gradient(lambda x: x**2,
lambda x: 2.0 * x,
f_arg_tensor=numpy.random.random(
(random.randint(1, 10), random.randint(1, 10))),
f_shape='lin')
helpers.check_gradient(lambda x: numpy.sqrt(x),
lambda x: 1.0 / (2 * numpy.sqrt(x)),
f_arg_tensor=numpy.random.random(
(random.randint(1, 10), random.randint(1, 10))),
f_shape='lin')
def test_output_correct_jacobian():
rows = random.randint(1, 10)
cols = random.randint(1, 10)
flat_model_target_vec = numpy.random.random(rows * cols)
similarities = numpy.random.random(rows)
target_vec = numpy.random.random(cols)
helpers.check_gradient(
lambda x: ill._output_correct_objective(x, similarities, target_vec),
lambda x: ill._output_correct_jacobian(x, similarities, target_vec),
f_arg_tensor=flat_model_target_vec,
f_shape='scalar')
def _check_jacobian(make_model_func):
attrs = random.randint(1, 10)
outs = random.randint(1, 10)
model = make_model_func(attrs, random.randint(1, 10), outs)
inp_matrix, tar_matrix = datasets.get_random_regression(10, attrs, outs)
# Test jacobian of error function
f = lambda xk: mlp._mlp_obj(model, inp_matrix, tar_matrix, xk)
df = lambda xk: mlp._mlp_obj_jac(model, inp_matrix, tar_matrix, xk)[1]
helpers.check_gradient(
f, df, inputs=mlp._flatten(model._weight_matrices), f_shape='scalar')
def test_check_gradient_jacobian_matrix_arg():
helpers.check_gradient(
lambda x: numpy.array(
[x[0][0]**2 + x[1][0]**2, x[0][1]**2 + x[1][1]**2]),
lambda x: numpy.array([[[2 * x[0][0], 0], [2 * x[1][0], 0]],
[[0, 2 * x[0][1]], [0, 2 * x[1][1]]]]),
f_arg_tensor=numpy.random.random((2, 2)),
f_shape='jac')
helpers.check_gradient(lambda x: numpy.array(
[numpy.sum(x**2), numpy.sum(x**3)]),
lambda x: numpy.array([2 * x, 3 * x**2]),
f_arg_tensor=numpy.random.random(
(random.randint(1, 10), random.randint(1, 10))),
f_shape='jac')
def _check_jacobian(make_model_func):
attrs = random.randint(1, 10)
outs = random.randint(1, 10)
model = make_model_func(attrs, outs)
inp_matrix, tar_matrix = datasets.get_random_regression(10, attrs, outs)
# Test jacobian of error function
f = lambda xk: model._get_obj(xk, inp_matrix, tar_matrix)
df = lambda xk: model._get_obj_jac(xk, inp_matrix, tar_matrix)[1]
helpers.check_gradient(f,
df,
inputs=model._weight_matrix.ravel(),
f_shape='scalar')
def _check_jacobian(make_model_func):
attrs = random.randint(1, 10)
outs = random.randint(1, 10)
model = make_model_func(attrs, random.randint(1, 10), outs)
inp_matrix, tar_matrix = datasets.get_random_regression(random.randint(1, 10), attrs, outs)
# Test jacobian of error function
f = lambda xk: model._get_obj(xk, inp_matrix, tar_matrix)
df = lambda xk: model._get_obj_jac(xk, inp_matrix, tar_matrix)[1]
helpers.check_gradient(
f,
df,
f_arg_tensor=mlp._flatten(model._bias_vec, model._weight_matrices),
f_shape='scalar')
def test_check_gradient_lin():
helpers.check_gradient(lambda x: x**2, lambda x: 2 * x, f_shape='lin')
helpers.check_gradient(lambda x: numpy.sqrt(x),
lambda x: 1.0 / (2 * numpy.sqrt(x)),
f_shape='lin')
def test_dgaussian():
helpers.check_gradient(
calculate.gaussian,
lambda x: calculate.dgaussian(x, calculate.gaussian(x)),
f_shape='lin')
def test_L2Penalty_jacobian():
penalty_func = error.L2Penalty(penalty_weight=random.uniform(0.0, 2.0))
helpers.check_gradient(penalty_func, penalty_func.derivative)
def test_check_gradient_scalar():
helpers.check_gradient(lambda x: numpy.sum(x**2),
lambda x: 2.0 * x,
f_shape='scalar')
def test_dsoftmax():
helpers.check_gradient(calculate.softmax,
lambda x: calculate.dsoftmax(calculate.softmax(x)),
f_shape='jac')
def test_drelu():
helpers.check_gradient(calculate.relu, calculate.drelu, f_shape='lin')
def test_big_drelu():
helpers.check_gradient(calculate.relu,
calculate.drelu,
inputs=numpy.array([0., 1000.]),
f_shape='lin')
def test_dtanh():
helpers.check_gradient(calculate.tanh,
lambda x: calculate.dtanh(calculate.tanh(x)),
f_shape='lin')
def test_dlogit():
helpers.check_gradient(calculate.logit, calculate.dlogit, f_shape='lin')
