Internal change

PiperOrigin-RevId: 302921283
brkygokcen · Mar 25, 2020 · 7c83a9d · 7c83a9d
1 parent 9fc4fd0
commit 7c83a9d
Show file tree

Hide file tree

Showing 2 changed files with 22 additions and 1 deletion.
diff --git a/official/nlp/transformer/misc.py b/official/nlp/transformer/misc.py
@@ -239,7 +239,10 @@ def get_callbacks(steps_per_epoch):
   """Returns common callbacks."""
   callbacks = []
   if FLAGS.enable_time_history:
-    time_callback = keras_utils.TimeHistory(FLAGS.batch_size, FLAGS.log_steps)
+    time_callback = keras_utils.TimeHistory(
+        FLAGS.batch_size,
+        FLAGS.log_steps,
+        FLAGS.model_dir if FLAGS.enable_tensorboard else None)
     callbacks.append(time_callback)
 
   if FLAGS.enable_tensorboard:

diff --git a/official/nlp/transformer/transformer_main.py b/official/nlp/transformer/transformer_main.py
@@ -246,6 +246,11 @@ def train(self):
 
     callbacks = self._create_callbacks(flags_obj.model_dir, 0, params)
 
+    # Only TimeHistory callback is supported for CTL
+    if params["use_ctl"]:
+      callbacks = [cb for cb in callbacks
+                   if isinstance(cb, keras_utils.TimeHistory)]
+
     # TODO(b/139418525): Refactor the custom training loop logic.
     @tf.function
     def train_steps(iterator, steps):
@@ -299,8 +304,13 @@ def _step_fn(inputs):
         if not self.use_tpu:
           raise NotImplementedError(
               "Custom training loop on GPUs is not implemented.")
+
         # Runs training steps.
         with summary_writer.as_default():
+          for cb in callbacks:
+            cb.on_epoch_begin(current_iteration)
+            cb.on_batch_begin(0)
+
           train_steps(
               train_ds_iterator,
               tf.convert_to_tensor(train_steps_per_eval, dtype=tf.int32))
@@ -309,10 +319,18 @@ def _step_fn(inputs):
           logging.info("Train Step: %d/%d / loss = %s", current_step,
                        flags_obj.train_steps, train_loss)
 
+          for cb in callbacks:
+            cb.on_batch_end(train_steps_per_eval - 1)
+            cb.on_epoch_end(current_iteration)
+
           if params["enable_tensorboard"]:
             for metric_obj in train_metrics:
               tf.compat.v2.summary.scalar(metric_obj.name, metric_obj.result(),
                                           current_step)
+              summary_writer.flush()
+
+        for cb in callbacks:
+          cb.on_train_end()
 
         if flags_obj.enable_checkpointing:
           # avoid check-pointing when running for benchmarking.