Single timestamp list and return average_exp_per_second. (#6032)

* Single timestamp list and return average_exp_per_second.

* Add paren to subtract before division.

official/resnet/keras/keras_common.py

@@ -51,14 +51,12 @@ class TimeHistory(tf.keras.callbacks.Callback):
       batch_size: Total batch size.
 
     """
-    self._batch_size = batch_size
+    self.batch_size = batch_size
     super(TimeHistory, self).__init__()
     self.log_steps = log_steps
 
-    # has stats for all batches
-    self.batch_start_timestamps = []
-    # only has stats for batch_index % log_steps == 0 (excluding 0)
-    self.batch_end_timestamps = []
+    # Logs start of step 0 then end of each step based on log_steps interval.
+    self.timestamp_log = []
 
   def on_train_begin(self, logs=None):
     self.record_batch = True
@@ -71,16 +69,17 @@ class TimeHistory(tf.keras.callbacks.Callback):
       timestamp = time.time()
       self.start_time = timestamp
       self.record_batch = False
-      self.batch_start_timestamps.append(BatchTimestamp(batch, timestamp))
+      if batch == 0:
+        self.timestamp_log.append(BatchTimestamp(batch, timestamp))
 
   def on_batch_end(self, batch, logs=None):
     if batch % self.log_steps == 0:
       timestamp = time.time()
       elapsed_time = timestamp - self.start_time
-      examples_per_second = (self._batch_size * self.log_steps) / elapsed_time
-      self.record_batch = True
+      examples_per_second = (self.batch_size * self.log_steps) / elapsed_time
       if batch != 0:
-        self.batch_end_timestamps.append(BatchTimestamp(batch, timestamp))
+        self.record_batch = True
+        self.timestamp_log.append(BatchTimestamp(batch, timestamp))
         tf.logging.info("BenchmarkMetric: {'num_batches':%d, 'time_taken': %f,"
                         "'images_per_second': %f}" %
                         (batch, elapsed_time, examples_per_second))
@@ -154,6 +153,7 @@ def build_stats(history, eval_output, time_callback):
       and sparse_categorical_accuracy.
     eval_output: Output of the eval step. Assumes first value is eval_loss and
       second value is accuracy_top_1.
+    time_callback: Time tracking callback likely used during keras.fit.
 
   Returns:
     Dictionary of normalized results.
@@ -174,9 +174,14 @@ def build_stats(history, eval_output, time_callback):
       stats[TRAIN_TOP_1] = train_hist['sparse_categorical_accuracy'][-1].item()
 
   if time_callback:
-    stats['batch_start_timestamps'] = time_callback.batch_start_timestamps
-    stats['batch_end_timestamps'] = time_callback.batch_end_timestamps
+    timestamp_log = time_callback.timestamp_log
+    stats['step_timestamp_log'] = timestamp_log
     stats['train_finish_time'] = time_callback.train_finish_time
+    if len(timestamp_log) > 1:
+      stats['avg_exp_per_second'] = (
+          time_callback.batch_size * time_callback.log_steps *
+          (len(time_callback.timestamp_log)-1) /
+          (timestamp_log[-1].timestamp - timestamp_log[0].timestamp))
 
   return stats