use EMA normalizer in RetinaNet

Summary: fix facebookresearch#868

Reviewed By: rbgirshick, alexander-kirillov

Differential Revision: D20062149

fbshipit-source-id: 5fcf0537730f3b3f7217fde42f8b97e506eab316

detectron2/modeling/meta_arch/retinanet.py

@@ -8,6 +8,7 @@
 
 from detectron2.layers import ShapeSpec, batched_nms, cat
 from detectron2.structures import Boxes, ImageList, Instances, pairwise_iou
+from detectron2.utils.events import get_event_storage
 from detectron2.utils.logger import log_first_n
 
 from ..anchor_generator import build_anchor_generator
@@ -98,6 +99,15 @@ def __init__(self, cfg):
         self.normalizer = lambda x: (x - pixel_mean) / pixel_std
         self.to(self.device)
 
+        """
+        In Detectron1, loss is normalized by number of foreground samples in the batch.
+        When batch size is 1 per GPU, #foreground has a large variance and
+        using it lead to lower performance. Here we maintain an EMA of #foreground to
+        stabilize the normalizer.
+        """
+        self.loss_normalizer = 100  # initialize with any reasonable #fg that's not too small
+        self.loss_normalizer_momentum = 0.9
+
     def forward(self, batched_inputs):
         """
         Args:
@@ -172,7 +182,12 @@ def losses(self, gt_classes, gt_anchors_deltas, pred_class_logits, pred_anchor_d
 
         valid_idxs = gt_classes >= 0
         foreground_idxs = (gt_classes >= 0) & (gt_classes != self.num_classes)
-        num_foreground = foreground_idxs.sum()
+        num_foreground = foreground_idxs.sum().item()
+        get_event_storage().put_scalar("num_foreground", num_foreground)
+        self.loss_normalizer = (
+            self.loss_normalizer_momentum * self.loss_normalizer
+            + (1 - self.loss_normalizer_momentum) * num_foreground
+        )
 
         gt_classes_target = torch.zeros_like(pred_class_logits)
         gt_classes_target[foreground_idxs, gt_classes[foreground_idxs]] = 1
@@ -184,15 +199,15 @@ def losses(self, gt_classes, gt_anchors_deltas, pred_class_logits, pred_anchor_d
             alpha=self.focal_loss_alpha,
             gamma=self.focal_loss_gamma,
             reduction="sum",
-        ) / max(1, num_foreground)
+        ) / max(1, self.loss_normalizer)
 
         # regression loss
         loss_box_reg = smooth_l1_loss(
             pred_anchor_deltas[foreground_idxs],
             gt_anchors_deltas[foreground_idxs],
             beta=self.smooth_l1_loss_beta,
             reduction="sum",
-        ) / max(1, num_foreground)
+        ) / max(1, self.loss_normalizer)
 
         return {"loss_cls": loss_cls, "loss_box_reg": loss_box_reg}