def test_wrong_input_shapes():
m = FractionalBias()
with pytest.raises(
ValueError,
match=r"Input data shapes should be the same, but given"):
m.update((torch.rand(4, 1, 2), torch.rand(4, 1)))
with pytest.raises(
ValueError,
match=r"Input data shapes should be the same, but given"):
m.update((torch.rand(4, 1), torch.rand(4, 1, 2)))
with pytest.raises(
ValueError,
match=r"Input data shapes should be the same, but given"):
m.update((
torch.rand(4, 1, 2),
torch.rand(4, ),
))
with pytest.raises(
ValueError,
match=r"Input data shapes should be the same, but given"):
m.update((
torch.rand(4, ),
torch.rand(4, 1, 2),
))
def test_zero_sample():
m = FractionalBias()
with pytest.raises(
NotComputableError,
match=
r"FractionalBias must have at least one example before it can be computed"
):
m.compute()
def test_wrong_input_shapes():
m = FractionalBias()
with pytest.raises(ValueError):
m.update((torch.rand(4, 1, 2), torch.rand(4, 1)))
with pytest.raises(ValueError):
m.update((torch.rand(4, 1), torch.rand(4, 1, 2)))
with pytest.raises(ValueError):
m.update((torch.rand(4, 1, 2), torch.rand(4,)))
with pytest.raises(ValueError):
m.update((torch.rand(4,), torch.rand(4, 1, 2)))
def _test(n_epochs, metric_device):
metric_device = torch.device(metric_device)
n_iters = 80
s = 16
n_classes = 2
offset = n_iters * s
y_true = torch.rand(size=(offset * idist.get_world_size(), ),
dtype=torch.double).to(device)
y_preds = torch.rand(size=(offset * idist.get_world_size(), ),
dtype=torch.double).to(device)
def update(engine, i):
return (
y_preds[i * s + rank * offset:(i + 1) * s + rank * offset],
y_true[i * s + rank * offset:(i + 1) * s + rank * offset],
)
engine = Engine(update)
m = FractionalBias(device=metric_device)
m.attach(engine, "fb")
data = list(range(n_iters))
engine.run(data=data, max_epochs=n_epochs)
assert "fb" in engine.state.metrics
res = engine.state.metrics["fb"]
if isinstance(res, torch.Tensor):
res = res.cpu().numpy()
np_y_true = y_true.cpu().numpy()
np_y_preds = y_preds.cpu().numpy()
np_sum = (2 * (np_y_true - np_y_preds) /
(np_y_preds + np_y_true + 1e-30)).sum()
np_len = len(y_preds)
np_ans = np_sum / np_len
assert pytest.approx(res, rel=tol) == np_ans
def _test(y_pred, y, batch_size):
def update_fn(engine, batch):
idx = (engine.state.iteration - 1) * batch_size
y_true_batch = np_y[idx : idx + batch_size]
y_pred_batch = np_y_pred[idx : idx + batch_size]
return torch.from_numpy(y_pred_batch), torch.from_numpy(y_true_batch)
engine = Engine(update_fn)
m = FractionalBias()
m.attach(engine, "fb")
np_y = y.double().numpy().ravel()
np_y_pred = y_pred.double().numpy().ravel()
data = list(range(y_pred.shape[0] // batch_size))
fb = engine.run(data, max_epochs=1).metrics["fb"]
np_sum = (2 * (np_y - np_y_pred) / (np_y_pred + np_y)).sum()
np_len = len(y_pred)
np_ans = np_sum / np_len
assert np_ans == pytest.approx(fb)
def _test(metric_device):
metric_device = torch.device(metric_device)
m = FractionalBias(device=metric_device)
torch.manual_seed(10 + rank)
y_pred = torch.randint(0, 10, size=(10,), device=device).float()
y = torch.randint(0, 10, size=(10,), device=device).float()
m.update((y_pred, y))
# gather y_pred, y
y_pred = idist.all_gather(y_pred)
y = idist.all_gather(y)
np_y_pred = y_pred.cpu().numpy()
np_y = y.cpu().numpy()
res = m.compute()
np_sum = (2 * (np_y - np_y_pred) / (np_y_pred + np_y + 1e-30)).sum()
np_len = len(y_pred)
np_ans = np_sum / np_len
assert np_ans == pytest.approx(res, rel=tol)
def test_fractional_bias():
a = np.random.randn(4)
b = np.random.randn(4)
c = np.random.randn(4)
d = np.random.randn(4)
ground_truth = np.random.randn(4)
m = FractionalBias()
m.update((torch.from_numpy(a), torch.from_numpy(ground_truth)))
np_sum = (2 * (ground_truth - a) / (a + ground_truth)).sum()
np_len = len(a)
np_ans = np_sum / np_len
assert m.compute() == pytest.approx(np_ans)
m.update((torch.from_numpy(b), torch.from_numpy(ground_truth)))
np_sum += (2 * (ground_truth - b) / (b + ground_truth)).sum()
np_len += len(b)
np_ans = np_sum / np_len
assert m.compute() == pytest.approx(np_ans)
m.update((torch.from_numpy(c), torch.from_numpy(ground_truth)))
np_sum += (2 * (ground_truth - c) / (c + ground_truth)).sum()
np_len += len(c)
np_ans = np_sum / np_len
assert m.compute() == pytest.approx(np_ans)
m.update((torch.from_numpy(d), torch.from_numpy(ground_truth)))
np_sum += (2 * (ground_truth - d) / (d + ground_truth)).sum()
np_len += len(d)
np_ans = np_sum / np_len
assert m.compute() == pytest.approx(np_ans)
def test_error_is_not_nan():
m = FractionalBias()
m.update((torch.zeros(4), torch.zeros(4)))
assert not (torch.isnan(m._sum_of_errors).any() or torch.isinf(m._sum_of_errors).any()), m._sum_of_errors
def test_zero_div():
m = FractionalBias()
with pytest.raises(NotComputableError):
m.compute()
