Fixes semantic loss top_k computation

official/vision/beta/projects/panoptic_maskrcnn/losses/panoptic_deeplab_losses.py

@@ -76,18 +76,38 @@ def __call__(self, logits, labels, sample_weight=None):
     if self._top_k_percent_pixels >= 1.0:
       loss = tf.reduce_sum(cross_entropy_loss) / normalizer
     else:
-      cross_entropy_loss = tf.reshape(cross_entropy_loss, shape=[-1])
-      top_k_pixels = tf.cast(
-          self._top_k_percent_pixels *
-          tf.cast(tf.size(cross_entropy_loss), tf.float32), tf.int32)
-      top_k_losses, _ = tf.math.top_k(
-          cross_entropy_loss, k=top_k_pixels, sorted=True)
-      normalizer = tf.reduce_sum(
-          tf.cast(tf.not_equal(top_k_losses, 0.0), tf.float32)) + EPSILON
-      loss = tf.reduce_sum(top_k_losses) / normalizer
-
+      loss = self._compute_top_k_loss(cross_entropy_loss)
     return loss
 
+  def _compute_top_k_loss(self, loss):
+    batch_size = tf.shape(loss)[0]
+    loss = tf.reshape(loss, shape=[batch_size, -1])
+
+    top_k_pixels = tf.cast(
+        self._top_k_percent_pixels *
+        tf.cast(tf.shape(loss)[-1], dtype=tf.float32),
+        dtype=tf.int32)
+
+    # shape: [batch_size, top_k_pixels]
+    per_sample_top_k_loss = tf.map_fn(
+        fn=lambda x: tf.nn.top_k(x, k=top_k_pixels, sorted=False)[0],
+        elems=loss,
+        parallel_iterations=32,
+        fn_output_signature=tf.float32)
+
+    # shape: [batch_size]
+    per_sample_normalizer = tf.reduce_sum(
+        tf.cast(
+            tf.not_equal(per_sample_top_k_loss, 0.0),
+            dtype=tf.float32),
+        axis=-1) + EPSILON
+    per_sample_normalized_loss = tf.reduce_sum(
+        per_sample_top_k_loss, axis=-1) / per_sample_normalizer
+
+    normalized_loss = tf_utils.safe_mean(per_sample_normalized_loss)
+    return normalized_loss
+
+
 class CenterHeatmapLoss:
   def __init__(self):
     self._loss_fn = tf.losses.mean_squared_error