Merge pull request #1180 from Wanglongzhi2001/master

fix: fix EarlyStopping
SciSharp · Sep 21, 2023 · 3811e4e · 3811e4e
2 parents 725ec1e + f809f6e
commit 3811e4e
Show file tree

Hide file tree

Showing 2 changed files with 48 additions and 22 deletions.
diff --git a/src/TensorFlowNET.Core/NumPy/Numpy.Math.cs b/src/TensorFlowNET.Core/NumPy/Numpy.Math.cs
@@ -85,5 +85,11 @@ public static NDArray dot(NDArray x1, NDArray x2, NDArray? axes = null, string?
 
         [AutoNumPy]
         public static NDArray add(NDArray x, NDArray y) => new NDArray(math_ops.add(x, y));
+
+        [AutoNumPy]
+        public static NDArray greater(NDArray x, NDArray y) => new NDArray(tf.greater(x, y));
+
+        [AutoNumPy]
+        public static NDArray less(NDArray x, NDArray y) => new NDArray(tf.less(x, y));
     }
 }
diff --git a/src/TensorFlowNET.Keras/Callbacks/Earlystopping.cs b/src/TensorFlowNET.Keras/Callbacks/Earlystopping.cs
@@ -19,8 +19,10 @@ public class EarlyStopping: ICallback
     string _monitor;
     string _mode;
     bool _restore_best_weights;
-    List<IVariableV1>? _best_weights;
+    List<NDArray>? _best_weights;
     CallbackParams _parameters;
+    Func<NDArray, NDArray, NDArray> _monitor_op;
+
     public Dictionary<string, List<float>>? history { get; set; }
     // user need to pass a CallbackParams to EarlyStopping, CallbackParams at least need the model
     public EarlyStopping(CallbackParams parameters,string monitor = "val_loss", float min_delta = 0f, int patience = 0,
@@ -38,17 +40,49 @@ public EarlyStopping(CallbackParams parameters,string monitor = "val_loss", floa
         _min_delta = Math.Abs(min_delta);
         _restore_best_weights = restore_best_weights;
         _mode = mode;
-        if (mode != "auto" && mode != "min" && mode != "max")
+
+        if (_mode != "auto" && _mode != "min" && _mode != "max")
+        {
+            Console.WriteLine($"EarlyStopping mode {_mode} is unknown, fallback to auto mode.");
+            _mode = "auto";
+        }
+
+        if (_mode == "min")
+        {
+            _monitor_op = np.less;
+        }
+        else if (_mode == "max")
+        {
+            _monitor_op = np.greater;
+        }
+        else
+        {
+            if (_monitor.EndsWith("acc") || _monitor.EndsWith("accuracy") || _monitor.EndsWith("auc"))
+            {
+                _monitor_op = np.greater;
+            }
+            else
+            {
+                _monitor_op = np.less;
+            }   
+        }
+
+        if (_monitor_op == np.greater)
         {
-            Console.WriteLine("EarlyStopping mode %s is unknown, fallback to auto mode.", mode);
+            _min_delta *= 1;
+        }
+        else
+        {
+            _min_delta *= -1;
         }
     }
     public void on_train_begin()
     {
         _wait = 0;
         _stopped_epoch = 0;
+        _best = _monitor_op == np.less ? (float)np.Inf : (float)-np.Inf;
+        _best_weights = null;
         _best_epoch = 0;
-        _best = (float)np.Inf;
     }
 
     public void on_epoch_begin(int epoch)
@@ -74,7 +108,7 @@ public void on_epoch_end(int epoch, Dictionary<string, float> epoch_logs)
         // Restore the weights after first epoch if no progress is ever made.
         if (_restore_best_weights && _best_weights == null)
         {
-            _best_weights = _parameters.Model.Weights;
+            _best_weights = _parameters.Model.get_weights();
         }
         _wait += 1;
 
@@ -83,7 +117,7 @@ public void on_epoch_end(int epoch, Dictionary<string, float> epoch_logs)
             _best = current;
             _best_epoch = epoch;
             if (_restore_best_weights)
-                _best_weights = _parameters.Model.TrainableWeights;
+                _best_weights = _parameters.Model.get_weights();
             // Only restart wait if we beat both the baseline and our previous best.
             if (_baseline == 0f || _is_improvement(current, _baseline))
                 _wait = 0;
@@ -99,7 +133,7 @@ public void on_epoch_end(int epoch, Dictionary<string, float> epoch_logs)
                 {
                     Console.WriteLine($"Restoring model weights from the end of the best epoch: {_best_epoch + 1}");
                 }
-                _parameters.Model.Weights = _best_weights;
+                _parameters.Model.set_weights(_best_weights);
             }
         }
     }
@@ -131,21 +165,7 @@ float get_monitor_value(Dictionary<string, float> logs)
     }
     public bool _is_improvement(float monitor_value, float reference_value)
     {
-        bool less_op = (monitor_value - _min_delta) < reference_value;
-        bool greater_op = (monitor_value - _min_delta) >= reference_value;
-        if (_mode == "min")
-            return less_op;
-        else if (_mode == "max")
-            return greater_op;
-        else
-        {
-            if (_monitor.EndsWith("acc") || _monitor.EndsWith("accuracy") || _monitor.EndsWith("auc"))
-            {
-                return greater_op;
-            }
-            else
-                return less_op;
-        }
+        return _monitor_op(monitor_value - _min_delta, reference_value);
     }
 
     public void on_test_end(Dictionary<string, float> logs)