Add a tf.summary.scalar call for L2 loss in resnet

agmitre · Mar 22, 2018 · d821ab8 · d821ab8
1 parent 4b85dab
commit d821ab8
Showing 1 changed file with 3 additions and 1 deletion.
diff --git a/official/resnet/resnet_run_loop.py b/official/resnet/resnet_run_loop.py
@@ -235,9 +235,11 @@ def exclude_batch_norm(name):
   loss_filter_fn = loss_filter_fn or exclude_batch_norm
 
   # Add weight decay to the loss.
-  loss = cross_entropy + weight_decay * tf.add_n(
+  l2_loss = weight_decay * tf.add_n(
       [tf.nn.l2_loss(v) for v in tf.trainable_variables()
        if loss_filter_fn(v.name)])
+  tf.summary.scalar('l2_loss', l2_loss)
+  loss = cross_entropy + l2_loss
 
   if mode == tf.estimator.ModeKeys.TRAIN:
     global_step = tf.train.get_or_create_global_step()