def test_initialize(self):
input_1 = 3
with self.assertRaises(TypeError):
Parameter(input_1)
input_2 = None
with self.assertRaises(TypeError):
Parameter(input_2)
def test_step(self):
# test for value
param = OrderedDict()
param['test_1'] = Parameter(np.ones(4))
param['test_1'].grad = np.ones(4)
optimizer = SGD(param)
optimizer.step()
self.assertTrue(
(param['test_1'].val == np.ones(4) *
(1. - 0.0001)).all()) # checking with hand calculation
def test_step(self):
# test for raising error
param = OrderedDict()
param['test_1'] = Parameter(np.ones(4))
optimizer = Optimizer(param)
with self.assertRaises(
NotImplementedError
): # not imeplement the step method of Optimizer
optimizer.step()
def test_initialize(self):
# test for raising error
# invalid alpha
param = OrderedDict()
param['test_1'] = Parameter(np.ones(4))
with self.assertRaises(TypeError):
SGD(param, alpha=[0.1]) # wrong value of alpha
with self.assertRaises(ValueError):
SGD(param, alpha=-0.1) # wrong type of alpha
def test_step(self):
# test for value
param = OrderedDict()
param['test_1'] = Parameter(np.ones(4))
param['test_1'].grad = np.ones(4)
optimizer = MomentumSGD(param)
optimizer.step() # first time
self.assertTrue(
(param['test_1'].val == np.ones(4) *
(1. - 0.0001)).all()) # checking with hand calculation
pre_grad = -np.ones(4) * 0.0001 # omega
grad = -np.ones(4) * 0.0001 + pre_grad * 0.9 # omega
val = param['test_1'].val + grad
optimizer.step() # second time
self.assertTrue((param['test_1'].val == val
).all()) # checking with hand calculation
def test_register_parameters(self):
input_1 = 0.5 # float
with self.assertRaises(TypeError):
self.module.register_parameters(input_1)
input_2 = np.array([0.1]) # numpy.ndarray
with self.assertRaises(TypeError):
self.module.register_parameters(input_2)
input_3 = Parameter(np.zeros(5))
self.module.register_parameters(input_3)
self.assertTrue(id(input_3) == id(self.module.parameters['param_0']))
def test_initialize_numerical_gradient(self):
# test for type(x)
input_1_param = 3.
input_1_fn = lambda weight: None
with self.assertRaises(TypeError):
numerical_gradient(input_1_param, input_1_fn)
# test for type(f)
input_2_param = OrderedDict()
input_2_param['test_1'] = Parameter(np.ones(4))
input_2_fn = 4.
with self.assertRaises(TypeError):
numerical_gradient(input_2_param, input_2_fn)