Improved FIL documentation, put FIL on the front page.

README.md

@@ -83,6 +83,7 @@ repo](https://github.com/rapidsai/notebooks-contrib).
 | | Stochastic Gradient Descent (SGD), Coordinate Descent (CD), and Quasi-Newton (QN) (including L-BFGS and OWL-QN) solvers for linear models  | |
 | **Nonlinear Models for Regression or Classification** | Random Forest (RF) Classification | Experimental multi-node multi-GPU via Dask |
 | | Random Forest (RF) Regression | Experimental multi-node multi-GPU via Dask |
+| | Inference for decision tree-based models | Forest Inference Library (FIL) |
 |  | K-Nearest Neighbors (KNN) | Multi-node multi-GPU via Dask, uses [Faiss](https://github.com/facebookresearch/faiss) for Nearest Neighbors Query. |
 |  | K-Nearest Neighbors (KNN) Classification | |
 |  | K-Nearest Neighbors (KNN) Regression | |

python/cuml/fil/fil.pyx

@@ -316,29 +316,30 @@ cdef class ForestInference_impl():
 
 
 class ForestInference(Base):
-    """
-    ForestInference provides GPU-accelerated inference (prediction)
+    """ForestInference provides GPU-accelerated inference (prediction)
     for random forest and boosted decision tree models.
 
     This module does not support training models. Rather, users should
     train a model in another package and save it in a
     treelite-compatible format. (See https://github.com/dmlc/treelite)
-    Currently, LightGBM and XGBoost GBDT and random forest models are
-    supported.
+    Currently, LightGBM, XGBoost and SKLearn GBDT and random forest models
+    are supported.
 
-    Users typically create a ForestInference object by loading a
-    saved model file with ForestInference.load. The resulting object
+    Users typically create a ForestInference object by loading a saved model
+    file with ForestInference.load. It is also possible to create it from an
+    SKLearn model using ForestInference.load_from_sklearn. The resulting object
     provides a `predict` method for carrying out inference.
 
     **Known limitations**:
-     * Trees are represented as complete binary trees, so a tree of depth k
-       will be stored in (2**k) - 1 nodes. This will be less space-efficient
-       for sparse trees.
-     * While treelite supports additional formats, only XGBoost and LightGBM
-       are tested in FIL currently.
-     * LightGBM categorical features are not supported
+     * A single row of data should fit into the shared memory of a thread block, 
+       which means that more than 12288 features are not supported.
+     * From sklearn.ensemble, only
+       {RandomForest,GradientBoosting}{Classifier,Regressor} models are
+       supported; other sklearn.ensemble models are currently not supported.
+     * Importing large SKLearn models can be slow, as it is done in Python.
+     * LightGBM categorical features are not supported.
      * Inference uses a dense matrix format, which is efficient for many
-       problems but will be suboptimal for sparse datasets.
+       problems but can be suboptimal for sparse datasets.
      * Only binary classification and regression are supported.
 
     Parameters