diff --git a/easy_rec/python/model/esmm.py b/easy_rec/python/model/esmm.py
index 50567ae63..a83e304a8 100644
--- a/easy_rec/python/model/esmm.py
+++ b/easy_rec/python/model/esmm.py
@@ -71,7 +71,7 @@ def build_loss_graph(self):
           tf.float32)
       cvr_losses = tf.keras.backend.binary_crossentropy(
           ctcvr_label, self._prediction_dict['probs_ctcvr'])
-      cvr_loss = tf.reduce_sum(cvr_losses, name='ctcvr_loss')
+      cvr_loss = tf.reduce_mean(cvr_losses, name='ctcvr_loss')
       # The weight defaults to 1.
       self._loss_dict['weighted_cross_entropy_loss_%s' %
                       cvr_tower_name] = self._cvr_tower_cfg.weight * cvr_loss
@@ -84,14 +84,15 @@ def build_loss_graph(self):
       cvr_loss = tf.losses.mean_squared_error(
           labels=ctcvr_label,
           predictions=self._prediction_dict['y_ctcvr'],
-          weights=self._sample_weight)
+          weights=self._sample_weight,
+          reduction=tf.losses.Reduction.MEAN)
       self._loss_dict['weighted_l2_loss_%s' %
                       cvr_tower_name] = self._cvr_tower_cfg.weight * cvr_loss
     _labels = tf.cast(self._labels[ctr_label_name], tf.float32)
     _logits = self._prediction_dict['logits_%s' % ctr_tower_name]
     cross = tf.nn.sigmoid_cross_entropy_with_logits(
         labels=_labels, logits=_logits, name='ctr_loss')
-    ctr_loss = tf.reduce_sum(cross)
+    ctr_loss = tf.reduce_mean(cross)
     self._loss_dict['weighted_cross_entropy_loss_%s' %
                     ctr_tower_name] = self._ctr_tower_cfg.weight * ctr_loss
     return self._loss_dict