From afd671d76408686fb6cda80e7b684486d99fb768 Mon Sep 17 00:00:00 2001
From: James Lamb <jaylamb20@gmail.com>
Date: Mon, 14 Dec 2020 15:07:06 -0600
Subject: [PATCH 1/2] [doc] [dask] Add example on early stopping with Dask

---
 doc/tutorials/dask.rst | 77 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 77 insertions(+)

diff --git a/doc/tutorials/dask.rst b/doc/tutorials/dask.rst
index 1d00f7754696..6706a69ff8cb 100644
--- a/doc/tutorials/dask.rst
+++ b/doc/tutorials/dask.rst
@@ -273,6 +273,83 @@ actual computation will return a coroutine and hence require awaiting:
         # Use `client.compute` instead of the `compute` method from dask collection
         print(await client.compute(prediction))
 
+*****************************
+Evaluation and Early Stopping
+*****************************
+
+.. versionadded:: 1.3.0
+
+The Dask interface allows the use of validation sets that are stored in distributed collections (Dask DataFrame or Dask Array). These can be used for evaluation and early stopping.
+
+To enable early stopping, pass one or more validation sets containing ``DaskDMatrix`` objects.
+
+.. code-block:: python
+
+    import dask.array as da
+    import xgboost as xgb
+
+    num_rows = 1e6
+    num_features = 100
+    num_partitions = 10
+    rows_per_chunk = num_rows / num_partitions
+
+    data = da.random.random(
+        size=(num_rows, num_features),
+        chunks=(rows_per_chunk, num_features)
+    )
+
+    labels = da.random.random(
+        size=(num_rows, 1),
+        chunks=(rows_per_chunk, 1)
+    )
+
+    X_eval = da.random.random(
+        size=(num_rows, num_features),
+        chunks=(rows_per_chunk, num_features)
+    )
+
+    y_eval = da.random.random(
+        size=(num_rows, 1),
+        chunks=(rows_per_chunk, 1)
+    )
+
+    dtrain = xgb.dask.DaskDMatrix(
+        client=client,
+        data=data,
+        label=labels
+    )
+
+    dvalid = xgb.dask.DaskDMatrix(
+        client=client,
+        data=X_eval,
+        label=y_eval
+    )
+
+    result = xgb.dask.train(
+        client=client,
+        params={
+            "objective": "reg:squarederror",
+        },
+        dtrain=dtrain,
+        num_boost_round=10,
+        evals=[(dvalid, "valid1")],
+        early_stopping_rounds=3
+    )
+
+When validation sets are provided to ``xgb.dask.train()`` in this way, the model object returned by ``xgb.dask.train()`` contains a history of evaluation metrics for each validation set, across all boosting rounds.
+
+.. code-block:: python
+
+    print(result["history"])
+    # {'valid1': OrderedDict([('rmse', [0.28857, 0.28858, 0.288592, 0.288598])])}
+
+If early stopping is enabled by also passing ``early_stopping_rounds``, you can check the best iteration in the returned booster.
+
+.. code-block:: python
+
+    bst = result["booster"]
+    print(booster.best_iteration)
+
 *****************************************************************************
 Why is the initialization of ``DaskDMatrix``  so slow and throws weird errors
 *****************************************************************************

From bf58056cc827ddbe383e9390235e7026217d2492 Mon Sep 17 00:00:00 2001
From: fis <jm.yuan@outlook.com>
Date: Tue, 15 Dec 2020 16:42:07 +0800
Subject: [PATCH 2/2] Typo, model slice.

---
 doc/tutorials/dask.rst | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/doc/tutorials/dask.rst b/doc/tutorials/dask.rst
index 6706a69ff8cb..80754baaca53 100644
--- a/doc/tutorials/dask.rst
+++ b/doc/tutorials/dask.rst
@@ -347,8 +347,9 @@ If early stopping is enabled by also passing ``early_stopping_rounds``, you can
 
 .. code-block:: python
 
-    bst = result["booster"]
+    booster = result["booster"]
     print(booster.best_iteration)
+    best_model = booster[: booster.best_iteration]
 
 *****************************************************************************
 Why is the initialization of ``DaskDMatrix``  so slow and throws weird errors