diff --git a/scenic/projects/baselines/centernet/modeling/centernet.py b/scenic/projects/baselines/centernet/modeling/centernet.py
index caa4710ad..f506bbc7a 100644
--- a/scenic/projects/baselines/centernet/modeling/centernet.py
+++ b/scenic/projects/baselines/centernet/modeling/centernet.py
@@ -214,15 +214,17 @@ def heatmap_focal_loss(self, heatmaps, gt_heatmaps):
     pos_w = (gt_heatmaps == 1.).astype(jnp.float32)  # B x m x C
     pos_loss = jnp.log(pred) * jnp.power(1 - pred, self.focal_gamma) * pos_w
     neg_loss = jnp.log(1. - pred) * jnp.power(pred, self.focal_gamma) * neg_w
-    norm = jnp.maximum(pos_w.sum(), 1.)  # scalar
+    bs = pos_w.shape[0]
+    norm = jnp.maximum(pos_w.reshape((bs, -1)).sum(1), 1.0)
+    norm = jnp.mean(norm)  # scalar
     if self.sync_device_norm:  # sync across GPUs. Helpful for small batch size.
-      norm = jax.lax.pmean(norm, axis_name='batch')
-    pos_loss = pos_loss.sum() / norm  # scalar
-    neg_loss = neg_loss.sum() / norm  # scalar
+      norm = jax.lax.pmean(norm, axis_name='batch')  # scalar
+    pos_loss = jnp.mean(pos_loss.reshape((bs, -1)).sum(1)) / norm  # scalar
+    neg_loss = jnp.mean(neg_loss.reshape((bs, -1)).sum(1)) / norm  # scalar
     if self.focal_alpha >= 0:
       pos_loss = self.focal_alpha * pos_loss
       neg_loss = (1. - self.focal_alpha) * neg_loss
-    return - pos_loss, - neg_loss, norm / heatmaps.shape[0]
+    return - pos_loss, - neg_loss, norm
 
   def reg_loss(self, box_regs, gt_regs):
     """Compute regression loss.
@@ -242,11 +244,12 @@ def reg_loss(self, box_regs, gt_regs):
     """
     reg_inds = gt_regs.max(axis=2) >= 0  # B x m: find valid pixels.
     gious = centernet_utils.giou_loss(box_regs, gt_regs)  # B x m
-    norm = jnp.maximum(reg_inds.sum(), 1.)  # scalar
+    norm = jnp.maximum(reg_inds.sum(1), 1.)
+    norm = jnp.mean(norm)  # scalar
     if self.sync_device_norm:
       norm = jax.lax.pmean(norm, axis_name='batch')
-    reg_loss = (gious * reg_inds).sum() / norm  # scalar
-    return reg_loss, gious, norm / reg_inds.shape[0]
+    reg_loss = jnp.mean((gious * reg_inds).sum(1)) / norm  # scalar
+    return reg_loss, gious, norm
 
   def _get_bbox_ltrb(self, grids, boxes, m, n):
     """generate FCOS style regression targets.