Top k multi error (#2178)

* Implement top-k multiclass error metric. Add new parameter top_k_threshold. * Add test for multiclass metrics * Make test less sensitive to avoid floating-point issues. * Change tabs to spaces. * Fix problem with test in Python 2. Refactor to use np.testing. Decrease number of training rounds so loss is larger and easier to compare. * Move multiclass tests into test_engine.py * Change parameter name from top_k_threshold to multi_error_top_k. * Fix top-k error metric to handle case where scores are equal. Update tests and docs. * Change name of top-k metric to multi_error@k. * Change tabs to spaces. * Fix formatting. * Fix minor issues in docs.
microsoft · May 26, 2019 · b3db9e9 · b3db9e9
1 parent 19de2be
commit b3db9e9
Show file tree

Hide file tree

Showing 5 changed files with 100 additions and 15 deletions.
diff --git a/docs/Parameters.rst b/docs/Parameters.rst
@@ -843,6 +843,18 @@ Metric Parameters
 
    -  `NDCG <https://en.wikipedia.org/wiki/Discounted_cumulative_gain#Normalized_DCG>`__ and `MAP <https://makarandtapaswi.wordpress.com/2012/07/02/intuition-behind-average-precision-and-map/>`__ evaluation positions, separated by ``,``
 
+-  ``multi_error_top_k`` :raw-html:`<a id="multi_error_top_k" title="Permalink to this parameter" href="#multi_error_top_k">&#x1F517;&#xFE0E;</a>`, default = ``1``, type = int, constraints: ``multi_error_top_k > 0``
+
+   -  used only with ``multi_error`` metric
+
+   -  threshold for top-k multi-error metric
+
+   -  the error on each sample is ``0`` if the true class is among the top ``multi_error_top_k`` predictions, and ``1`` otherwise
+
+      -  more precisely, the error on a sample is ``0`` if there are at least ``num_classes - multi_error_top_k`` predictions strictly less than the prediction on the true class
+
+   -  when ``multi_error_top_k=1`` this is equivalent to the usual multi-error metric
+
 Network Parameters
 ------------------
 

diff --git a/include/LightGBM/config.h b/include/LightGBM/config.h
@@ -747,6 +747,14 @@ struct Config {
   // desc = `NDCG <https://en.wikipedia.org/wiki/Discounted_cumulative_gain#Normalized_DCG>`__ and `MAP <https://makarandtapaswi.wordpress.com/2012/07/02/intuition-behind-average-precision-and-map/>`__ evaluation positions, separated by ``,``
   std::vector<int> eval_at;
 
+  // check = >0
+  // desc = used only with ``multi_error`` metric
+  // desc = threshold for top-k multi-error metric
+  // desc = the error on each sample is ``0`` if the true class is among the top ``multi_error_top_k`` predictions, and ``1`` otherwise
+  // descl2 = more precisely, the error on a sample is ``0`` if there are at least ``num_classes - multi_error_top_k`` predictions strictly less than the prediction on the true class
+  // desc = when ``multi_error_top_k=1`` this is equivalent to the usual multi-error metric
+  int multi_error_top_k = 1;
+
   #pragma endregion
 
   #pragma region Network Parameters

diff --git a/src/io/config_auto.cpp b/src/io/config_auto.cpp
@@ -260,6 +260,7 @@ std::unordered_set<std::string> Config::parameter_set({
   "metric_freq",
   "is_provide_training_metric",
   "eval_at",
+  "multi_error_top_k",
   "num_machines",
   "local_listen_port",
   "time_out",
@@ -521,6 +522,9 @@ void Config::GetMembersFromString(const std::unordered_map<std::string, std::str
     eval_at = Common::StringToArray<int>(tmp_str, ',');
   }
 
+  GetInt(params, "multi_error_top_k", &multi_error_top_k);
+  CHECK(multi_error_top_k >0);
+
   GetInt(params, "num_machines", &num_machines);
   CHECK(num_machines >0);
 
@@ -637,6 +641,7 @@ std::string Config::SaveMembersToString() const {
   str_buf << "[metric_freq: " << metric_freq << "]\n";
   str_buf << "[is_provide_training_metric: " << is_provide_training_metric << "]\n";
   str_buf << "[eval_at: " << Common::Join(eval_at, ",") << "]\n";
+  str_buf << "[multi_error_top_k: " << multi_error_top_k << "]\n";
   str_buf << "[num_machines: " << num_machines << "]\n";
   str_buf << "[local_listen_port: " << local_listen_port << "]\n";
   str_buf << "[time_out: " << time_out << "]\n";

diff --git a/src/metric/multiclass_metric.hpp b/src/metric/multiclass_metric.hpp
@@ -20,15 +20,15 @@ namespace LightGBM {
 template<typename PointWiseLossCalculator>
 class MulticlassMetric: public Metric {
  public:
-  explicit MulticlassMetric(const Config& config) {
+  explicit MulticlassMetric(const Config& config) :config_(config){
     num_class_ = config.num_class;
   }
 
   virtual ~MulticlassMetric() {
   }
 
   void Init(const Metadata& metadata, data_size_t num_data) override {
-    name_.emplace_back(PointWiseLossCalculator::Name());
+    name_.emplace_back(PointWiseLossCalculator::Name(config_));
     num_data_ = num_data;
     // get label
     label_ = metadata.label();
@@ -72,7 +72,7 @@ class MulticlassMetric: public Metric {
           std::vector<double> rec(num_pred_per_row);
           objective->ConvertOutput(raw_score.data(), rec.data());
           // add loss
-          sum_loss += PointWiseLossCalculator::LossOnPoint(label_[i], rec);
+          sum_loss += PointWiseLossCalculator::LossOnPoint(label_[i], rec, config_);
         }
       } else {
         #pragma omp parallel for schedule(static) reduction(+:sum_loss)
@@ -85,7 +85,7 @@ class MulticlassMetric: public Metric {
           std::vector<double> rec(num_pred_per_row);
           objective->ConvertOutput(raw_score.data(), rec.data());
           // add loss
-          sum_loss += PointWiseLossCalculator::LossOnPoint(label_[i], rec) * weights_[i];
+          sum_loss += PointWiseLossCalculator::LossOnPoint(label_[i], rec, config_) * weights_[i];
         }
       }
     } else {
@@ -98,7 +98,7 @@ class MulticlassMetric: public Metric {
             rec[k] = static_cast<double>(score[idx]);
           }
           // add loss
-          sum_loss += PointWiseLossCalculator::LossOnPoint(label_[i], rec);
+          sum_loss += PointWiseLossCalculator::LossOnPoint(label_[i], rec, config_);
         }
       } else {
         #pragma omp parallel for schedule(static) reduction(+:sum_loss)
@@ -109,7 +109,7 @@ class MulticlassMetric: public Metric {
             rec[k] = static_cast<double>(score[idx]);
           }
           // add loss
-          sum_loss += PointWiseLossCalculator::LossOnPoint(label_[i], rec) * weights_[i];
+          sum_loss += PointWiseLossCalculator::LossOnPoint(label_[i], rec, config_) * weights_[i];
         }
       }
     }
@@ -129,25 +129,28 @@ class MulticlassMetric: public Metric {
   /*! \brief Name of this test set */
   std::vector<std::string> name_;
   int num_class_;
+  /*! \brief config parameters*/
+  Config config_;
 };
 
-/*! \brief L2 loss for multiclass task */
+/*! \brief top-k error for multiclass task; if k=1 (default) this is the usual multi-error */
 class MultiErrorMetric: public MulticlassMetric<MultiErrorMetric> {
  public:
   explicit MultiErrorMetric(const Config& config) :MulticlassMetric<MultiErrorMetric>(config) {}
 
-  inline static double LossOnPoint(label_t label, std::vector<double>& score) {
+  inline static double LossOnPoint(label_t label, std::vector<double>& score, const Config& config) {
     size_t k = static_cast<size_t>(label);
+    int num_larger = 0;
     for (size_t i = 0; i < score.size(); ++i) {
-      if (i != k && score[i] >= score[k]) {
-        return 1.0f;
-      }
+      if (score[i] >= score[k]) ++num_larger;
+      if (num_larger > config.multi_error_top_k) return 1.0f;
     }
     return 0.0f;
   }
 
-  inline static const char* Name() {
-    return "multi_error";
+  inline static const std::string Name(const Config& config) {
+    if (config.multi_error_top_k == 1) return "multi_error";
+    else return "multi_error@" + std::to_string(config.multi_error_top_k);
   }
 };
 
@@ -156,7 +159,7 @@ class MultiSoftmaxLoglossMetric: public MulticlassMetric<MultiSoftmaxLoglossMetr
  public:
   explicit MultiSoftmaxLoglossMetric(const Config& config) :MulticlassMetric<MultiSoftmaxLoglossMetric>(config) {}
 
-  inline static double LossOnPoint(label_t label, std::vector<double>& score) {
+  inline static double LossOnPoint(label_t label, std::vector<double>& score, const Config&) {
     size_t k = static_cast<size_t>(label);
     if (score[k] > kEpsilon) {
       return static_cast<double>(-std::log(score[k]));
@@ -165,7 +168,7 @@ class MultiSoftmaxLoglossMetric: public MulticlassMetric<MultiSoftmaxLoglossMetr
     }
   }
 
-  inline static const char* Name() {
+  inline static const std::string Name(const Config&) {
     return "multi_logloss";
   }
 };

diff --git a/tests/python_package_test/test_engine.py b/tests/python_package_test/test_engine.py
@@ -26,6 +26,13 @@ def multi_logloss(y_true, y_pred):
     return np.mean([-math.log(y_pred[i][y]) for i, y in enumerate(y_true)])
 
 
+def top_k_error(y_true, y_pred, k):
+    if k == y_pred.shape[1]:
+        return 0
+    max_rest = np.max(-np.partition(-y_pred, k)[:, k:], axis=1)
+    return 1 - np.mean((y_pred[np.arange(len(y_true)), y_true] > max_rest))
+
+
 class TestEngine(unittest.TestCase):
     def test_binary(self):
         X, y = load_breast_cancer(True)
@@ -363,6 +370,56 @@ def test_multiclass_prediction_early_stopping(self):
         ret = multi_logloss(y_test, gbm.predict(X_test, **pred_parameter))
         self.assertLess(ret, 0.2)
 
+    def test_multi_class_error(self):
+        X, y = load_digits(return_X_y=True)
+        params = {'objective': 'multiclass', 'num_classes': 10, 'metric': 'multi_error', 'num_leaves': 4, 'seed': 0,
+                  'num_rounds': 30, 'verbose': -1}
+        lgb_data = lgb.Dataset(X, label=y)
+        results = {}
+        est = lgb.train(params, lgb_data, valid_sets=[lgb_data], valid_names=['train'], evals_result=results)
+        predict_default = est.predict(X)
+        params = {'objective': 'multiclass', 'num_classes': 10, 'metric': 'multi_error', 'multi_error_top_k': 1,
+                  'num_leaves': 4, 'seed': 0, 'num_rounds': 30, 'verbose': -1, 'metric_freq': 10}
+        results = {}
+        est = lgb.train(params, lgb_data, valid_sets=[lgb_data], valid_names=['train'], evals_result=results)
+        predict_1 = est.predict(X)
+        # check that default gives same result as k = 1
+        np.testing.assert_array_almost_equal(predict_1, predict_default, 5)
+        # check against independent calculation for k = 1
+        err = top_k_error(y, predict_1, 1)
+        np.testing.assert_almost_equal(results['train']['multi_error'][-1], err, 5)
+        # check against independent calculation for k = 2
+        params = {'objective': 'multiclass', 'num_classes': 10, 'metric': 'multi_error', 'multi_error_top_k': 2,
+                  'num_leaves': 4, 'seed': 0, 'num_rounds': 30, 'verbose': -1, 'metric_freq': 10}
+        results = {}
+        est = lgb.train(params, lgb_data, valid_sets=[lgb_data], valid_names=['train'], evals_result=results)
+        predict_2 = est.predict(X)
+        err = top_k_error(y, predict_2, 2)
+        np.testing.assert_almost_equal(results['train']['multi_error@2'][-1], err, 5)
+        # check against independent calculation for k = 10
+        params = {'objective': 'multiclass', 'num_classes': 10, 'metric': 'multi_error', 'multi_error_top_k': 10,
+                  'num_leaves': 4, 'seed': 0, 'num_rounds': 30, 'verbose': -1, 'metric_freq': 10}
+        results = {}
+        est = lgb.train(params, lgb_data, valid_sets=[lgb_data], valid_names=['train'], evals_result=results)
+        predict_2 = est.predict(X)
+        err = top_k_error(y, predict_2, 10)
+        np.testing.assert_almost_equal(results['train']['multi_error@10'][-1], err, 5)
+        # check case where predictions are equal
+        X = np.array([[0, 0], [0, 0]])
+        y = np.array([0, 1])
+        lgb_data = lgb.Dataset(X, label=y)
+        params = {'objective': 'multiclass', 'num_classes': 2, 'metric': 'multi_error', 'multi_error_top_k': 1,
+                  'num_leaves': 4, 'seed': 0, 'num_rounds': 1, 'verbose': -1, 'metric_freq': 10}
+        results = {}
+        lgb.train(params, lgb_data, valid_sets=[lgb_data], valid_names=['train'], evals_result=results)
+        np.testing.assert_almost_equal(results['train']['multi_error'][-1], 1, 5)
+        lgb_data = lgb.Dataset(X, label=y)
+        params = {'objective': 'multiclass', 'num_classes': 2, 'metric': 'multi_error', 'multi_error_top_k': 2,
+                  'num_leaves': 4, 'seed': 0, 'num_rounds': 1, 'verbose': -1, 'metric_freq': 10}
+        results = {}
+        lgb.train(params, lgb_data, valid_sets=[lgb_data], valid_names=['train'], evals_result=results)
+        np.testing.assert_almost_equal(results['train']['multi_error@2'][-1], 0, 5)
+
     def test_early_stopping(self):
         X, y = load_breast_cancer(True)
         params = {