test.py

@pytest.mark.parametrize("loss", ['huber', 'ls', 'lad', 'quantile'])
@pytest.mark.parametrize("use_sample_weight", [False, True])
def test_regressor_train_loss_convergence(loss, use_sample_weight):
    rng = np.random.RandomState(42)
    n_samples, n_features = 30, 5
    n_estimators = 300

    # Make random data (without duplicated samples) to make sure
    # it's possible to build an invertible (overfitting) mapping
    # from X to y that therefore should lead to a regression loss
    # of zero if n_estimators is large enough.
    X = rng.randn(n_samples, n_features)
    y = rng.randn(n_samples)

    if use_sample_weight:
        sample_weight = rng.uniform(0, 10, size=n_samples)
        sample_weight[sample_weight < 2] = 0
    else:
        sample_weight = None

    gbr = GradientBoostingRegressor(
        learning_rate=0.1,
        max_depth=3,
        loss=loss,
        n_estimators=n_estimators,
        n_iter_no_change=None,  # make sure early stopping is disabled
    )
    gbr.fit(X, y, sample_weight=sample_weight)

    train_loss = gbr.loss_(y, gbr._raw_predict(X),
                           sample_weight=sample_weight)

    assert len(gbr.train_score_) == n_estimators
    # assert gbr.train_score_[-1] == pytest.approx(train_loss)
    assert train_loss < 1e-6, gbr.train_score_

train_loss is probably wrong but I do not know why. Maybe in the mean time we should just check:

assert gbr.train_score_[-1] < 1e-6, gbr.train_score_