Merge pull request #792 from birdnet-team/focal-loss-autotune

Fix focal loss in autotune and apply sigmoid for test data evaluation

Author: max-mauermann

birdnet_analyzer/model.py

@@ -478,6 +478,9 @@ def save_model_params(path):
             "Upsamling ratio",
             "use mixup",
             "use label smoothing",
+            "use focal loss",
+            "focal loss alpha",
+            "focal loss gamma",
             "BirdNET Model version",
         ),
         (
@@ -492,6 +495,9 @@ def save_model_params(path):
             cfg.UPSAMPLING_RATIO,
             cfg.TRAIN_WITH_MIXUP,
             cfg.TRAIN_WITH_LABEL_SMOOTHING,
+            cfg.TRAIN_WITH_FOCAL_LOSS,
+            cfg.FOCAL_LOSS_ALPHA,
+            cfg.FOCAL_LOSS_GAMMA,
             cfg.MODEL_VERSION,
         ),
     )

birdnet_analyzer/train/utils.py

@@ -496,6 +496,10 @@ def run_trial(self, trial, *args, **kwargs):
         cfg.UPSAMPLING_RATIO = best_params["upsampling_ratio"]
         cfg.TRAIN_WITH_MIXUP = best_params["mixup"]
         cfg.TRAIN_WITH_LABEL_SMOOTHING = best_params["label_smoothing"]
+        cfg.TRAIN_WITH_FOCAL_LOSS = best_params["focal_loss"]
+        if cfg.TRAIN_WITH_FOCAL_LOSS:
+            cfg.FOCAL_LOSS_ALPHA = best_params["focal_loss_alpha"]
+            cfg.FOCAL_LOSS_GAMMA = best_params["focal_loss_gamma"]
 
         print("Best params: ")
         print("hidden_units: ", cfg.TRAIN_HIDDEN_UNITS)
@@ -507,6 +511,10 @@ def run_trial(self, trial, *args, **kwargs):
             print("upsampling_mode: ", cfg.UPSAMPLING_MODE)
         print("mixup: ", cfg.TRAIN_WITH_MIXUP)
         print("label_smoothing: ", cfg.TRAIN_WITH_LABEL_SMOOTHING)
+        print("focal_loss: ", cfg.TRAIN_WITH_FOCAL_LOSS)
+        if cfg.TRAIN_WITH_FOCAL_LOSS:
+            print("focal_loss_alpha: ", cfg.FOCAL_LOSS_ALPHA)
+            print("focal_loss_gamma: ", cfg.FOCAL_LOSS_GAMMA)
 
     # Build model
     print("Building model...", flush=True)
@@ -724,6 +732,8 @@ def evaluate_model(classifier, x_test, y_test, labels, threshold=None):
     # Make predictions
     y_pred_prob = classifier.predict(x_test)
 
+    y_pred_prob = model.flat_sigmoid(y_pred_prob, sensitivity=-1, bias=1.0)
+
     # Calculate metrics for each class
     metrics = {}