def test_loss_boundary(loss):
"""Test interval ranges of y_true and y_pred in losses."""
# make sure low and high are always within the interval, used for linspace
if loss.is_multiclass:
y_true = np.linspace(0, 9, num=10)
else:
low, high = _inclusive_low_high(loss.interval_y_true)
y_true = np.linspace(low, high, num=10)
# add boundaries if they are included
if loss.interval_y_true.low_inclusive:
y_true = np.r_[y_true, loss.interval_y_true.low]
if loss.interval_y_true.high_inclusive:
y_true = np.r_[y_true, loss.interval_y_true.high]
assert loss.in_y_true_range(y_true)
n = y_true.shape[0]
low, high = _inclusive_low_high(loss.interval_y_pred)
if loss.is_multiclass:
y_pred = np.empty((n, 3))
y_pred[:, 0] = np.linspace(low, high, num=n)
y_pred[:, 1] = 0.5 * (1 - y_pred[:, 0])
y_pred[:, 2] = 0.5 * (1 - y_pred[:, 0])
else:
y_pred = np.linspace(low, high, num=n)
assert loss.in_y_pred_range(y_pred)
# calculating losses should not fail
raw_prediction = loss.link.link(y_pred)
loss.loss(y_true=y_true, raw_prediction=raw_prediction)
def random_y_true_raw_prediction(loss,
n_samples,
y_bound=(-100, 100),
raw_bound=(-5, 5),
seed=42):
"""Random generate y_true and raw_prediction in valid range."""
rng = np.random.RandomState(seed)
if loss.is_multiclass:
raw_prediction = np.empty((n_samples, loss.n_classes))
raw_prediction.flat[:] = rng.uniform(
low=raw_bound[0],
high=raw_bound[1],
size=n_samples * loss.n_classes,
)
y_true = np.arange(n_samples).astype(float) % loss.n_classes
else:
raw_prediction = rng.uniform(low=raw_bound[0],
high=raw_bound[0],
size=n_samples)
# generate a y_true in valid range
low, high = _inclusive_low_high(loss.interval_y_true)
low = max(low, y_bound[0])
high = min(high, y_bound[1])
y_true = rng.uniform(low, high, size=n_samples)
# set some values at special boundaries
if loss.interval_y_true.low == 0 and loss.interval_y_true.low_inclusive:
y_true[::(n_samples // 3)] = 0
if loss.interval_y_true.high == 1 and loss.interval_y_true.high_inclusive:
y_true[1::(n_samples // 3)] = 1
return y_true, raw_prediction
def test_is_in_range(interval):
# make sure low and high are always within the interval, used for linspace
low, high = _inclusive_low_high(interval)
x = np.linspace(low, high, num=10)
assert interval.includes(x)
# x contains lower bound
assert interval.includes(np.r_[x, interval.low]) == interval.low_inclusive
# x contains upper bound
assert interval.includes(np.r_[x, interval.high]) == interval.high_inclusive
# x contains upper and lower bound
assert interval.includes(np.r_[x, interval.low, interval.high]) == (
interval.low_inclusive and interval.high_inclusive
)
