def test_LossDict_aggregate_max_depth_gt_0(self):
def loss():
return torch.tensor(1.0)
loss_dict = pystiche.LossDict(
(("0.0.0", loss()), ("0.0.1", loss()), ("0.1", loss()), ("1", loss()))
)
actual = loss_dict.aggregate(1)
desired = pystiche.LossDict((("0", 3 * loss()), ("1", loss())))
self.assertDictEqual(actual, desired)
actual = loss_dict.aggregate(2)
desired = pystiche.LossDict(
(("0.0", 2 * loss()), ("0.1", loss()), ("1", loss()))
)
self.assertDictEqual(actual, desired)
actual = loss_dict.aggregate(3)
desired = loss_dict
self.assertDictEqual(actual, desired)
actual = loss_dict.aggregate(4)
desired = loss_dict
self.assertDictEqual(actual, desired)
def test_setitem_LossDict(self):
name = "loss"
loss_dict = pystiche.LossDict()
num_sub_losses = 3
loss = pystiche.LossDict([(str(idx), torch.tensor(idx))
for idx in range(num_sub_losses)])
loss_dict[name] = loss
for idx in range(num_sub_losses):
actual = loss_dict[f"{name}.{idx}"]
desired = loss[str(idx)]
ptu.assert_allclose(actual, desired)
def test_default_image_optim_log_fn_loss_dict_smoke(self):
class MockOptimLogger:
def __init__(self):
self.msg = None
@contextlib.contextmanager
def environment(self, header):
yield
def message(self, msg):
self.msg = msg
loss_dict = pystiche.LossDict(
(("a", torch.tensor(0.0)), ("b.c", torch.tensor(1.0))))
log_freq = 1
max_depth = 1
optim_logger = MockOptimLogger()
log_fn = optim.default_image_optim_log_fn(optim_logger,
log_freq=log_freq,
max_depth=max_depth)
step = log_freq
log_fn(step, loss_dict)
actual = optim_logger.msg
desired = loss_dict.format(max_depth=max_depth)
self.assertEqual(actual, desired)
def test_backward(self):
losses = [
torch.tensor(val, dtype=torch.float, requires_grad=True)
for val in range(3)
]
def zero_grad():
for loss in losses:
loss.grad = None
def extract_grads():
return [loss.grad.clone() for loss in losses]
zero_grad()
loss_dict = pystiche.LossDict([(str(idx), loss)
for idx, loss in enumerate(losses)])
loss_dict.backward()
actuals = extract_grads()
zero_grad()
total = sum(losses)
total.backward()
desireds = extract_grads()
for actual, desired in zip(actuals, desireds):
ptu.assert_allclose(actual, desired)
def test_item(self):
losses = (1.0, 2.0)
loss_dict = pystiche.LossDict([(f"loss{idx}", torch.tensor(val))
for idx, val in enumerate(losses)])
actual = loss_dict.item()
desired = sum(losses)
assert actual == pytest.approx(desired)
def test_LossDict_item(self):
losses = (1.0, 2.0)
loss_dict = pystiche.LossDict([(f"loss{idx}", torch.tensor(val))
for idx, val in enumerate(losses)])
actual = loss_dict.item()
desired = sum(losses)
self.assertAlmostEqual(actual, desired)
def test_call(self):
input = torch.tensor(0.0)
named_ops = [(str(idx), self.Operator(idx + 1.0)) for idx in range(3)]
op_container = ops.OperatorContainer(named_ops)
actual = op_container(input)
desired = pystiche.LossDict([(name, input + op.bias) for name, op in named_ops])
ptu.assert_allclose(actual, desired)
def test_LossDict_total(self):
loss1 = torch.tensor(1.0)
loss2 = torch.tensor(2.0)
loss_dict = pystiche.LossDict((("loss1", loss1), ("loss2", loss2)))
actual = loss_dict.total()
desired = loss1 + loss2
self.assertTensorAlmostEqual(actual, desired)
def test_total(self):
loss1 = torch.tensor(1.0)
loss2 = torch.tensor(2.0)
loss_dict = pystiche.LossDict((("loss1", loss1), ("loss2", loss2)))
actual = loss_dict.total()
desired = loss1 + loss2
ptu.assert_allclose(actual, desired)
def test_setitem_Tensor(self):
name = "loss"
loss_dict = pystiche.LossDict()
loss = torch.tensor(1.0)
loss_dict[name] = loss
actual = loss_dict[name]
desired = loss
ptu.assert_allclose(actual, desired)
def test_LossDict_setitem_Tensor(self):
name = "loss"
loss_dict = pystiche.LossDict()
loss = torch.tensor(1.0)
loss_dict[name] = loss
actual = loss_dict[name]
desired = loss
self.assertTensorAlmostEqual(actual, desired)
def forward(self, input_image: torch.Tensor) -> pystiche.LossDict:
for encoder in self._multi_layer_encoders:
encoder.encode(input_image)
loss = pystiche.LossDict([(name, op(input_image))
for name, op in self.named_children()])
for encoder in self._multi_layer_encoders:
encoder.empty_storage()
return loss
def test_mul(self):
losses = (2.0, 3.0)
factor = 4.0
loss_dict = pystiche.LossDict([(f"loss{idx}", torch.tensor(val))
for idx, val in enumerate(losses)])
loss_dict = loss_dict * factor
for idx, loss in enumerate(losses):
actual = float(loss_dict[f"loss{idx}"])
desired = loss * factor
assert actual == ptu.approx(desired)
def test_OperatorContainer_call(self):
class TestOperator(ops.Operator):
def __init__(self, bias):
super().__init__()
self.bias = bias
def process_input_image(self, image):
return image + self.bias
input = torch.tensor(0.0)
named_ops = [(str(idx), TestOperator(idx + 1.0)) for idx in range(3)]
op_container = ops.OperatorContainer(named_ops)
actual = op_container(input)
desired = pystiche.LossDict([(name, input + op.bias) for name, op in named_ops])
self.assertTensorDictAlmostEqual(actual, desired)
def test_MultiOperatorLoss_call():
class TestOperator(ops.Operator):
def __init__(self, bias):
super().__init__()
self.bias = bias
def process_input_image(self, image):
return image + self.bias
input = torch.tensor(0.0)
named_ops = [(str(idx), TestOperator(idx + 1.0)) for idx in range(3)]
multi_op_loss = loss.MultiOperatorLoss(named_ops)
actual = multi_op_loss(input)
desired = pystiche.LossDict([(name, input + op.bias)
for name, op in named_ops])
ptu.assert_allclose(actual, desired)
def test_LossMeter_update_LossDict(self):
actual_meter = optim.LossMeter("actual_meter")
desired_meter = optim.LossMeter("desired_meter")
losses = torch.arange(3, dtype=torch.float)
loss_dict = pystiche.LossDict([(str(idx), loss)
for idx, loss in enumerate(losses)])
actual_meter.update(loss_dict)
desired_meter.update(torch.sum(losses).item())
for attr in (
"count",
"last_val",
"global_sum",
"global_min",
"global_max",
"global_avg",
"local_avg",
):
actual = getattr(actual_meter, attr)
desired = getattr(desired_meter, attr)
self.assertAlmostEqual(actual, desired)
def test_setitem_other(self):
name = "loss"
loss_dict = pystiche.LossDict()
with pytest.raises(TypeError):
loss_dict[name] = 1.0
def test_setitem_non_scalar_Tensor(self):
name = "loss"
loss_dict = pystiche.LossDict()
with pytest.raises(TypeError):
loss_dict[name] = torch.ones(1)
def test_LossDict_float(self):
loss_dict = pystiche.LossDict(
(("a", torch.tensor(0.0)), ("b", torch.tensor(1.0))))
self.assertAlmostEqual(float(loss_dict), loss_dict.item())
def process_input_image(self, input_image: torch.Tensor) -> pystiche.LossDict:
return pystiche.LossDict(
[(name, op(input_image)) for name, op in self.named_children()]
)
def forward(self, input_image: torch.Tensor) -> pystiche.LossDict:
with self._mle_handler(input_image):
return pystiche.LossDict(
[(name, op(input_image)) for name, op in self.named_children()]
)
def test_repr_smoke(self):
loss_dict = pystiche.LossDict(
(("a", torch.tensor(0.0)), ("b", torch.tensor(1.0))))
assert isinstance(repr(loss_dict), str)
def test_float(self):
loss_dict = pystiche.LossDict(
(("a", torch.tensor(0.0)), ("b", torch.tensor(1.0))))
assert float(loss_dict) == pytest.approx(loss_dict.item())
def test_LossDict_setitem_other(self):
name = "loss"
loss_dict = pystiche.LossDict()
with self.assertRaises(TypeError):
loss_dict[name] = 1.0