def test_fit(self):
torch.manual_seed(42)
m = OCE(lambda x: 1 - torch.exp(-x))
self.train_oce(m)
x = torch.randn(10000)
result = m(x)
expect = torch.log(EntropicLoss()(x))
assert torch.isclose(result, expect, rtol=1e-02)
def test_repr(self):
loss = EntropicLoss()
assert repr(loss) == "EntropicLoss()"
loss = EntropicLoss(a=10.0)
assert repr(loss) == "EntropicLoss(a=10.0)"
def test_shape(self):
loss = EntropicLoss()
self.assert_shape(loss)
def test_error_a(self):
with pytest.raises(ValueError):
EntropicLoss(0)
with pytest.raises(ValueError):
EntropicLoss(-1)
def test_value(self, n_paths, a):
value = 1.0
loss = EntropicLoss(a)
result = loss(torch.full((n_paths, ), value))
expect = torch.exp(-a * torch.tensor(value))
assert torch.isclose(result, expect)
def test_cash(self, n_paths, a):
loss = EntropicLoss(a)
self.assert_cash(loss, torch.randn(n_paths))
def test_convex(self, n_paths, risk, a):
loss = EntropicLoss(risk)
x1 = torch.randn(n_paths)
x2 = torch.randn(n_paths)
self.assert_convex(loss, x1, x2, a)
def test_nonincreasing(self, n_paths, risk, a):
loss = EntropicLoss(risk)
self.assert_nonincreasing(loss, torch.randn(n_paths), a)