def test_both_regularized(self):
guided_vars_dict = {
'ex_index': {
'regularization_type': 'l1_norm',
'regularization_norm': 1,
'regularization_alpha': 1,
},
'errant_tag': {
'regularization_type': 'l1_norm',
'regularization_norm': 1,
'regularization_alpha': 1,
},
}
act_fn_dict = {
'dp-ex_index': lambda x: x,
'dp-errant_tag': lambda x: x
}
raw_ex_vars = np.array([0, 1, 4])
ex_index = np.array([1, 1, 1, 0, 2, 2, 2])
raw_errant_vars = np.array([0, 1, 4])
reg_loss = utils.get_total_regularization_loss(guided_vars_dict,
act_fn_dict,
raw_ex_vars, ex_index,
raw_errant_vars)
self.assertEqual(reg_loss, 20)
def test_errant_tag_l1norm(self):
guided_vars_dict = {
'errant_tag': {
'regularization_type': 'l1_norm',
'regularization_norm': 1,
'regularization_alpha': 1,
},
}
# Just pass values, same as guided_parameters_utils.activation_self_fn.
act_fn_dict = {'dp-errant_tag': lambda x: x}
raw_errant_vars = np.array([0, 1, 4])
reg_loss = utils.get_total_regularization_loss(
guided_vars_dict, act_fn_dict, raw_errant_vars=raw_errant_vars)
self.assertEqual(reg_loss, 5)
def test_ex_index_l1norm_alpha_half(self):
guided_vars_dict = {
'ex_index': {
'regularization_type': 'l1_norm',
'regularization_norm': 1,
'regularization_alpha': 0.5,
},
}
act_fn_dict = {'dp-ex_index': lambda x: x}
raw_ex_vars = np.array([0, 1, 4])
ex_index = np.array([1, 1, 1, 0, 2, 2, 2])
reg_loss = utils.get_total_regularization_loss(guided_vars_dict,
act_fn_dict,
raw_ex_vars, ex_index)
self.assertEqual(reg_loss, 7.5)
def test_ex_index_lpnorm(self):
guided_vars_dict = {
'ex_index': {
'regularization_type': 'lp_norm',
'regularization_norm': 1.5,
'regularization_alpha': 1,
},
}
# Just pass values, same as guided_parameters_utils.activation_self_fn.
act_fn_dict = {'dp-ex_index': lambda x: x}
raw_ex_vars = np.array([0, 1, 4])
ex_index = np.array([1, 1, 1, 0, 2, 2, 2])
reg_loss = utils.get_total_regularization_loss(guided_vars_dict,
act_fn_dict,
raw_ex_vars, ex_index)
self.assertAlmostEqual(float(reg_loss), 9, delta=1e-5)