Fixed issue 3605. (#3628)

* Fixed issue 3605.

- https://github.com/dmlc/xgboost/issues/3605

* Fixed the bug in a better way.

* Added a test to catch the bug.

* Fixed linter errors.

src/common/hist_util.cu

@@ -257,13 +257,13 @@ struct GPUSketcher {
       n_cuts_cur_[icol] = std::min(n_cuts_, n_unique);
       // if less elements than cuts: copy all elements with their weights
       if (n_cuts_ > n_unique) {
-        auto weights2_iter = weights2_.begin();
-        auto fvalues_iter = fvalues_cur_.begin();
-        auto cuts_iter = cuts_d_.begin() + icol * n_cuts_;
+        float* weights2_ptr = weights2_.data().get();
+        float* fvalues_ptr = fvalues_cur_.data().get();
+        WXQSketch::Entry* cuts_ptr = cuts_d_.data().get() + icol * n_cuts_;
         dh::LaunchN(device_, n_unique, [=]__device__(size_t i) {
-            bst_float rmax = weights2_iter[i];
-            bst_float rmin = i > 0 ? weights2_iter[i - 1] : 0;
-            cuts_iter[i] = WXQSketch::Entry(rmin, rmax, rmax - rmin, fvalues_iter[i]);
+            bst_float rmax = weights2_ptr[i];
+            bst_float rmin = i > 0 ? weights2_ptr[i - 1] : 0;
+            cuts_ptr[i] = WXQSketch::Entry(rmin, rmax, rmax - rmin, fvalues_ptr[i]);
           });
       } else if (n_cuts_cur_[icol] > 0) {
         // if more elements than cuts: use binary search on cumulative weights

tests/python-gpu/test_gpu_updaters.py

@@ -14,7 +14,7 @@ def assert_gpu_results(cpu_results, gpu_results):
         assert np.allclose(cpu_res["eval"][-1], gpu_res["eval"][-1], 1e-2, 1e-2)
 
 datasets = ["Boston", "Cancer", "Digits", "Sparse regression",
-            "Sparse regression with weights"]
+            "Sparse regression with weights", "Small weights regression"]
 
 class TestGPU(unittest.TestCase):
     def test_gpu_exact(self):

tests/python/regression_test_utilities.py

@@ -55,12 +55,20 @@ def get_sparse():
 
 
 def get_sparse_weights():
+    return get_weights_regression(1, 10)
+
+
+def get_small_weights():
+    return get_weights_regression(1e-6, 1e-5)
+
+
+def get_weights_regression(min_weight, max_weight):
     rng = np.random.RandomState(199)
     n = 10000
     sparsity = 0.25
     X, y = datasets.make_regression(n, random_state=rng)
     X = np.array([[np.nan if rng.uniform(0, 1) < sparsity else x for x in x_row] for x_row in X])
-    w = np.array([rng.uniform(1, 10) for i in range(n)])
+    w = np.array([rng.uniform(min_weight, max_weight) for i in range(n)])
     return X, y, w
 
 
@@ -130,6 +138,8 @@ def run_suite(param, num_rounds=10, select_datasets=None, scale_features=False):
         Dataset("Sparse regression", get_sparse, "reg:linear", "rmse"),
         Dataset("Sparse regression with weights", get_sparse_weights,
                 "reg:linear", "rmse", has_weights=True),
+        Dataset("Small weights regression", get_small_weights,
+                "reg:linear", "rmse", has_weights=True),
         Dataset("Boston External Memory", get_boston, "reg:linear", "rmse",
                 use_external_memory=True)
     ]