Metric aggregation testing

CHANGELOG.md

@@ -71,6 +71,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Fixed determinism in `DDPSpawnBackend` when using `seed_everything` in main process ([#3335](https://github.com/PyTorchLightning/pytorch-lightning/pull/3335))
 
+- Fixed aggregation of metrics ([#3517](https://github.com/PyTorchLightning/pytorch-lightning/pull/3517))
+
 ## [0.9.0] - YYYY-MM-DD
 
 ### Added

pytorch_lightning/metrics/classification.py

@@ -21,17 +21,18 @@
     auroc,
     average_precision,
     confusion_matrix,
+    _confmat_normalize,
     dice_score,
     f1_score,
     fbeta_score,
     iou,
     multiclass_precision_recall_curve,
     multiclass_roc,
-    precision,
     precision_recall_curve,
-    recall,
     roc,
+    precision_recall
 )
+from pytorch_lightning.metrics.functional.reduction import class_reduce
 from pytorch_lightning.metrics.metric import TensorMetric
 
 
@@ -44,8 +45,8 @@ class Accuracy(TensorMetric):
         >>> pred = torch.tensor([0, 1, 2, 3])
         >>> target = torch.tensor([0, 1, 2, 2])
         >>> metric = Accuracy()
-        >>> metric(pred, target).item()
-        0.75
+        >>> metric(pred, target)
+        tensor(0.7500)
 
     """
 
@@ -84,7 +85,14 @@ def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
             A Tensor with the classification score.
         """
         return accuracy(pred=pred, target=target,
-                        num_classes=self.num_classes, class_reduction=self.class_reduction)
+                        num_classes=self.num_classes,
+                        class_reduction='none',
+                        return_state=True)
+
+    @staticmethod
+    def compute(self, data: Any, output: Any):
+        tps, sups = output['tps'], output['sups']
+        return class_reduce(tps, sups, sups, class_reduction=self.class_reduction)
 
 
 class ConfusionMatrix(TensorMetric):
@@ -135,16 +143,16 @@ def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
             A Tensor with the confusion matrix.
         """
         return confusion_matrix(pred=pred, target=target,
-                                normalize=self.normalize,
+                                normalize=False,  # we normalize after ddp sync
                                 num_classes=self.num_classes)
 
-    def aggregate(self, *tensors: torch.Tensor) -> torch.Tensor:
-        """Aggregates results by stacking them instead of concatenating before averaging.
-
-        Returns:
-            the aggregated results
-        """
-        return torch.stack(tensors).mean(0)
+    @staticmethod
+    def compute(self, data: Any, output: Any):
+        """ Confusion matrix normalization needs to happen after ddp sync """
+        confmat = output
+        if self.normalize:
+            confmat = _confmat_normalize(confmat)
+        return confmat
 
 
 class PrecisionRecallCurve(TensorMetric):
@@ -202,7 +210,8 @@ def forward(
             - recall values
             - threshold values
         """
-        return precision_recall_curve(pred=pred, target=target, sample_weight=sample_weight, pos_label=self.pos_label)
+        return precision_recall_curve(pred=pred, target=target,
+                                      sample_weight=sample_weight, pos_label=self.pos_label)
 
 
 class Precision(TensorMetric):
@@ -256,9 +265,15 @@ def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
         Return:
             A Tensor with the classification score.
         """
-        return precision(pred=pred, target=target,
-                         num_classes=self.num_classes,
-                         class_reduction=self.class_reduction)
+        return precision_recall(pred=pred, target=target,
+                                num_classes=self.num_classes,
+                                class_reduction='none',
+                                return_state=True)
+
+    @staticmethod
+    def compute(self, data: Any, output: Any):
+        tps, fps, sups = output['tps'], output['fps'], output['sups']
+        return class_reduce(tps, tps + fps, sups, class_reduction=self.class_reduction)
 
 
 class Recall(TensorMetric):
@@ -313,10 +328,15 @@ def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
         Return:
             A Tensor with the classification score.
         """
-        return recall(pred=pred,
-                      target=target,
-                      num_classes=self.num_classes,
-                      class_reduction=self.class_reduction)
+        return precision_recall(pred=pred, target=target,
+                                num_classes=self.num_classes,
+                                class_reduction='none',
+                                return_state=True)
+
+    @staticmethod
+    def compute(self, data: Any, output: Any):
+        tps, fns, sups = output['tps'], output['fns'], output['sups']
+        return class_reduce(tps, tps + fns, sups, class_reduction=self.class_reduction)
 
 
 class AveragePrecision(TensorMetric):
@@ -470,12 +490,28 @@ def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
         Return:
             torch.Tensor: classification score
         """
-        return fbeta_score(pred=pred, target=target,
-                           beta=self.beta, num_classes=self.num_classes,
-                           class_reduction=self.class_reduction)
+        return precision_recall(pred=pred, target=target,
+                                num_classes=self.num_classes,
+                                class_reduction='none',
+                                return_state=True)
+
+    @staticmethod
+    def compute(self, data: Any, output: Any):
+        """ tps, fps, fns, sups needs to be synced before we do any calculations """
+        tps, fps, fns, sups = output['tps'], output['fps'], output['fns'], output['sups']
+
+        intermidiate_reduction = 'none' if self.class_reduction != "micro" else 'micro'
+        precision = class_reduce(tps, tps + fps, sups, class_reduction=intermidiate_reduction)
+        recall = class_reduce(tps, tps + fns, sups, class_reduction=intermidiate_reduction)
+
+        num = (1 + self.beta ** 2) * precision * recall
+        denom = ((self.beta ** 2) * precision + recall)
+        if intermidiate_reduction == 'micro':
+            return torch.sum(num) / torch.sum(denom)
+        return class_reduce(num, denom, sups, class_reduction=self.class_reduction)
 
 
-class F1(TensorMetric):
+class F1(FBeta):
     """
     Computes the F1 score, which is the harmonic mean of the precision and recall.
     It ranges between 1 and 0, where 1 is perfect and the worst value is 0.
@@ -507,29 +543,15 @@ def __init__(
 
             reduce_group: the process group to reduce metric results from DDP
         """
-        super().__init__(
+        super(TensorMetric, self).__init__(
             name="f1",
             reduce_group=reduce_group,
         )
 
         self.num_classes = num_classes
         assert class_reduction in ('micro', 'macro', 'weighted', 'none')
         self.class_reduction = class_reduction
-
-    def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
-        """
-        Actual metric computation
-
-        Args:
-            pred: predicted labels
-            target: groundtruth labels
-
-        Return:
-            torch.Tensor: classification score
-        """
-        return f1_score(pred=pred, target=target,
-                        num_classes=self.num_classes,
-                        class_reduction=self.class_reduction)
+        self.beta = 1.0
 
 
 class ROC(TensorMetric):

pytorch_lightning/metrics/functional/classification.py

@@ -239,7 +239,8 @@ def accuracy(
         pred: torch.Tensor,
         target: torch.Tensor,
         num_classes: Optional[int] = None,
-        class_reduction: str = 'micro'
+        class_reduction: str = 'micro',
+        return_state=False
 ) -> torch.Tensor:
     """
     Computes the accuracy classification score
@@ -254,7 +255,8 @@ def accuracy(
             - ``'macro'``: calculate metrics for each label, and find their unweighted mean.
             - ``'weighted'``: calculate metrics for each label, and find their weighted mean.
             - ``'none'``: returns calculated metric per class
-
+        return_state: returns a internal state that can be ddp reduced
+            before doing the final calculation
     Return:
          A Tensor with the accuracy score.
 
@@ -271,10 +273,21 @@ def accuracy(
 
     tps, fps, tns, fns, sups = stat_scores_multiple_classes(
         pred=pred, target=target, num_classes=num_classes)
-
+    if return_state:
+        return {'tps': tps, 'sups': sups}
     return class_reduce(tps, sups, sups, class_reduction=class_reduction)
 
 
+def _confmat_normalize(cm):
+    """ Normalization function for confusion matrix """
+    cm = cm / cm.sum(-1, keepdim=True)
+    nan_elements = cm[torch.isnan(cm)].nelement()
+    if nan_elements != 0:
+        cm[torch.isnan(cm)] = 0
+        rank_zero_warn(f'{nan_elements} nan values found in confusion matrix have been replaced with zeros.')
+    return cm
+
+
 def confusion_matrix(
         pred: torch.Tensor,
         target: torch.Tensor,
@@ -312,11 +325,7 @@ def confusion_matrix(
     cm = bins.reshape(num_classes, num_classes).squeeze().float()
 
     if normalize:
-        cm = cm / cm.sum(-1, keepdim=True)
-        nan_elements = cm[torch.isnan(cm)].nelement()
-        if nan_elements != 0:
-            cm[torch.isnan(cm)] = 0
-            rank_zero_warn(f'{nan_elements} nan values found in confusion matrix have been replaced with zeros.')
+        cm = _confmat_normalize(cm)
 
     return cm
 
@@ -326,7 +335,8 @@ def precision_recall(
         target: torch.Tensor,
         num_classes: Optional[int] = None,
         class_reduction: str = 'micro',
-        return_support: bool = False
+        return_support: bool = False,
+        return_state: bool = False
 ) -> Tuple[torch.Tensor, torch.Tensor]:
     """
     Computes precision and recall for different thresholds
@@ -343,6 +353,8 @@ def precision_recall(
             - ``'none'``: returns calculated metric per class
 
         return_support: returns the support for each class, need for fbeta/f1 calculations
+        return_state: returns a internal state that can be ddp reduced
+            before doing the final calculation
 
     Return:
         Tensor with precision and recall
@@ -359,6 +371,8 @@ def precision_recall(
 
     precision = class_reduce(tps, tps + fps, sups, class_reduction=class_reduction)
     recall = class_reduce(tps, tps + fns, sups, class_reduction=class_reduction)
+    if return_state:
+        return {'tps': tps, 'fps': fps, 'fns': fns, 'sups': sups}
     if return_support:
         return precision, recall, sups
     return precision, recall