def test_oracle(self):
cache_refs = [{
'dw': torch.tensor([[0.7357], [1.4836], [-0.3669]]),
'obj': torch.tensor(1.3400)
}, {
'dw': torch.tensor([[0.7414], [1.4836], [-0.3669]]),
'obj': torch.tensor(1.3397)
}, {
'dw': torch.tensor([[0.7414], [1.4836], [-0.3669]]),
'obj': torch.tensor(1.3397)
}, {
'dw': torch.tensor([[0.7414], [1.4836], [-0.3669]]),
'obj': torch.tensor(1.3397)
}, {
'dw': torch.tensor([[0.7414], [1.4836], [-0.3669]]),
'obj': torch.tensor(1.3397)
}]
temps = [1e-1, 1e-2, 1e-3, 1e-4, 1e-5]
for temp, cache_ref in zip(temps, cache_refs):
self.hparams.temp = temp
self.obj = SmoothedLasso_Gradient(self.hparams)
cache_test = self.obj.oracle(self.w, self.x, self.y)
assert_all_close_dict(
cache_ref, cache_test,
"oracle_info with parameter temp={}".format(temp))
if self.w.grad is not None:
self.w.grad.zero_()
class TestObj_SmoothedLasso_Gradient(unittest.TestCase):
def setUp(self):
_init_lasso(self)
self.obj = SmoothedLasso_Gradient(self.hparams)
def test_error(self):
error_test = self.obj.task_error(self.w, self.x, self.y)
error_ref = torch.tensor(1.3251)
assert_all_close(error_test, error_ref, "task_error returned value")
def test_oracle(self):
cache_refs = [{
'dw': torch.tensor([[0.7357], [1.4836], [-0.3669]]),
'obj': torch.tensor(1.3400)
}, {
'dw': torch.tensor([[0.7319], [1.4774], [-0.3645]]),
'obj': torch.tensor(1.3511)
}, {
'dw': torch.tensor([[0.7315], [1.4740], [-0.3579]]),
'obj': torch.tensor(1.5336)
}]
temps = [0.1, 1, 10]
for temp, cache_ref in zip(temps, cache_refs):
self.hparams.temp = temp
self.obj = SmoothedLasso_Gradient(self.hparams)
cache_test = self.obj.oracle(self.w, self.x, self.y)
assert_all_close_dict(
cache_ref, cache_test,
"oracle_info with parameter temp={}".format(temp))
if self.w.grad is not None:
self.w.grad.zero_()
def test_oracle(self):
cache_refs = [{
'dw': torch.tensor([[0.7357], [1.4836], [-0.3669]]),
'obj': torch.tensor(1.3400)},{
'dw': torch.tensor([[0.7319], [1.4774], [-0.3645]]),
'obj': torch.tensor(1.3511)},{
'dw': torch.tensor([[0.7315], [1.4740], [-0.3579]]),
'obj': torch.tensor(1.5336)}
]
temps = [0.1, 1, 10]
for temp, cache_ref in zip(temps, cache_refs):
self.hparams.temp = temp
self.obj = SmoothedLasso_Gradient(self.hparams)
cache_test = self.obj.oracle(self.w, self.x, self.y)
assert_all_close_dict(cache_ref, cache_test, "oracle_info with parameter temp={}".format(temp))
self.w.grad.zero_()
def get_objective(args, hparams):
print('Reg Coefficient mu: \t {}'.format(args.mu))
if args.obj == 'svm':
if args.opt in ('sgd', 'gd'):
obj = SVM_SubGradient(hparams)
elif args.opt == 'bcfw':
obj = SVM_ConditionalGradient(hparams)
else:
raise ValueError
elif args.obj == 'logistic':
if args.opt in ('sgd', 'gd'):
obj = Logistic_Gradient(hparams)
else:
raise ValueError
elif args.obj == 'ridge':
if args.opt in ('sgd', 'gd'):
obj = Ridge_Gradient(hparams)
elif args.opt == 'closed-form':
obj = Ridge_ClosedForm(hparams)
else:
raise ValueError
elif args.obj == 'lasso':
if args.opt in ('sgd', 'gd'):
obj = Lasso_subGradient(hparams)
else:
raise ValueError
elif args.obj == 'smooth-lasso':
if args.opt in ('sgd', 'gd'):
obj = SmoothedLasso_Gradient(hparams)
else:
raise ValueError
else:
raise ValueError(
'Did not recognize objective {obj}'.format(obj=args.obj))
return obj
def setUp(self):
_init_lasso(self)
self.obj = SmoothedLasso_Gradient(self.hparams)