Fix segfault and other errors in ForestInference.load_from_sklearn (#…

…5973)

Closes #5551

* Replace `np.float32` with `"float32"` so that we don't reference the `np` module. By the time `__dealloc__` method is called, modules may have already been unloaded.
* Improve the user experience by raising a helpful error when the user attempts to predict with an empty forest.

Authors:
  - Philip Hyunsu Cho (https://github.com/hcho3)
  - Dante Gama Dessavre (https://github.com/dantegd)

Approvers:
  - Dante Gama Dessavre (https://github.com/dantegd)

URL: #5973

python/cuml/cuml/fil/fil.pyx

@@ -356,7 +356,7 @@ cdef class ForestInference_impl():
         return None
 
     def get_dtype(self):
-        dtype_array = [np.float32, np.float64]
+        dtype_array = ["float32", "float64"]
         return dtype_array[self.forest_data.index()]
 
     def get_algo(self, algo_str):
@@ -477,14 +477,24 @@ cdef class ForestInference_impl():
         cdef uintptr_t preds_ptr
         preds_ptr = preds.ptr
 
-        if fil_dtype == np.float32:
+        if fil_dtype == "float32":
+            if self.get_forest32() == NULL:
+                raise RuntimeError(
+                    "Cannot call predict() with empty forest. "
+                    "Please load the forest first with load() or "
+                    "load_from_sklearn()")
             predict(handle_[0],
                     self.get_forest32(),
                     <float*> preds_ptr,
                     <float*> X_ptr,
                     <size_t> n_rows,
                     <bool> predict_proba)
-        elif fil_dtype == np.float64:
+        elif fil_dtype == "float64":
+            if self.get_forest64() == NULL:
+                raise RuntimeError(
+                    "Cannot call predict() with empty forest. "
+                    "Please load the forest first with load() or "
+                    "load_from_sklearn()")
             predict(handle_[0],
                     self.get_forest64(),
                     <double*> preds_ptr,
@@ -493,7 +503,7 @@ cdef class ForestInference_impl():
                     <bool> predict_proba)
         else:
             # should not reach here
-            assert False, 'invalid fil_dtype, must be np.float32 or np.float64'
+            assert False, 'invalid fil_dtype, must be float32 or float64'
 
         self.handle.sync()
 
@@ -557,15 +567,15 @@ cdef class ForestInference_impl():
     def __dealloc__(self):
         cdef handle_t* handle_ = <handle_t*><size_t>self.handle.getHandle()
         fil_dtype = self.get_dtype()
-        if fil_dtype == np.float32:
+        if fil_dtype == "float32":
             if self.get_forest32() != NULL:
                 free[float](handle_[0], self.get_forest32())
-        elif fil_dtype == np.float64:
+        elif fil_dtype == "float64":
             if self.get_forest64() != NULL:
                 free[double](handle_[0], self.get_forest64())
         else:
             # should not reach here
-            assert False, 'invalid fil_dtype, must be np.float32 or np.float64'
+            assert False, 'invalid fil_dtype, must be float32 or float64'
 
 
 class ForestInference(Base,
@@ -747,7 +757,7 @@ class ForestInference(Base,
            Optional 'out' location to store inference results
 
         safe_dtype_conversion : bool (default = False)
-            FIL converts data to np.float32 when needed. Set this parameter to
+            FIL converts data to float32 when needed. Set this parameter to
             True to enable checking for information loss during that
             conversion, but note that this check can have a significant
             performance penalty. Parameter will be dropped in a future
@@ -776,7 +786,7 @@ class ForestInference(Base,
            Optional 'out' location to store inference results
 
         safe_dtype_conversion : bool (default = False)
-            FIL converts data to np.float32 when needed. Set this parameter to
+            FIL converts data to float32 when needed. Set this parameter to
             True to enable checking for information loss during that
             conversion, but note that this check can have a significant
             performance penalty. Parameter will be dropped in a future

python/cuml/cuml/tests/dask/test_dask_random_forest.py

@@ -653,7 +653,7 @@ def test_rf_broadcast(model_type, fit_broadcast, transform_broadcast, client):
 
     if model_type == "classification":
         X, y = make_classification(
-            n_samples=n_workers * 1000,
+            n_samples=n_workers * 10000,
             n_features=20,
             n_informative=15,
             n_classes=4,
@@ -663,7 +663,7 @@ def test_rf_broadcast(model_type, fit_broadcast, transform_broadcast, client):
         y = y.astype(np.int32)
     else:
         X, y = make_regression(
-            n_samples=n_workers * 1000,
+            n_samples=n_workers * 10000,
             n_features=20,
             n_informative=5,
             random_state=123,