Save Scikit-Learn attributes into learner attributes. (#5245)

* Remove the recommendation for pickle.

* Save skl attributes in booster.attr

* Test loading scikit-learn model with native booster.

src/common/io.h

@@ -52,7 +52,7 @@ class PeekableInStream : public dmlc::Stream {
 class FixedSizeStream : public PeekableInStream {
  public:
   explicit FixedSizeStream(PeekableInStream* stream);
-  ~FixedSizeStream() = default;
+  ~FixedSizeStream() override = default;
 
   size_t Read(void* dptr, size_t size) override;
   size_t PeekRead(void* dptr, size_t size) override;

src/common/json.cc

@@ -1,6 +1,7 @@
 /*!
  * Copyright (c) by Contributors 2019
  */
+#include <cctype>
 #include <sstream>
 #include <limits>
 #include <cmath>
@@ -351,7 +352,9 @@ Json JsonReader::Parse() {
       return ParseObject();
     } else if ( c == '[' ) {
       return ParseArray();
-    } else if ( c == '-' || std::isdigit(c) ) {
+    } else if ( c == '-' || std::isdigit(c) ||
+                c == 'N' ) {
+      // For now we only accept `NaN`, not `nan` as the later violiates LR(1) with `null`.
       return ParseNumber();
     } else if ( c == '\"' ) {
       return ParseString();
@@ -547,6 +550,13 @@ Json JsonReader::ParseNumber() {
 
   // TODO(trivialfis): Add back all the checks for number
   bool negative = false;
+  if (XGBOOST_EXPECT(*p == 'N', false)) {
+    GetChar('N');
+    GetChar('a');
+    GetChar('N');
+    return Json(static_cast<Number::Float>(std::numeric_limits<float>::quiet_NaN()));
+  }
+
   if ('-' == *p) {
     ++p;
     negative = true;

src/learner.cc

@@ -661,13 +661,13 @@ class LearnerImpl : public Learner {
       CHECK(header == serialisation_header_)  // NOLINT
           << R"doc(
 
-If you are loading a serialized model (like pickle in Python) generated by older XGBoost,
-please export the model by calling `Booster.save_model` from that version first, then load
-it back in current version.  See:
+  If you are loading a serialized model (like pickle in Python) generated by older
+  XGBoost, please export the model by calling `Booster.save_model` from that version
+  first, then load it back in current version.  See:
 
-  https://xgboost.readthedocs.io/en/latest/tutorials/saving_model.html
+    https://xgboost.readthedocs.io/en/latest/tutorials/saving_model.html
 
-for more details about differences between saving model and serializing.
+  for more details about differences between saving model and serializing.
 
 )doc";
       int64_t json_offset {-1};