Internal change

PiperOrigin-RevId: 309586836

official/benchmark/bert_benchmark.py

@@ -31,6 +31,7 @@ import tensorflow as tf
 # pylint: enable=g-bad-import-order
 
 from official.benchmark import bert_benchmark_utils as benchmark_utils
+from official.benchmark import owner_utils
 from official.nlp.bert import configs
 from official.nlp.bert import run_classifier
 from official.utils.misc import distribution_utils
@@ -55,7 +56,6 @@ class BertClassifyBenchmarkBase(benchmark_utils.BertBenchmarkBase):
     super(BertClassifyBenchmarkBase, self).__init__(output_dir)
     self.num_epochs = None
     self.num_steps_per_epoch = None
-    self.tpu = tpu
     FLAGS.steps_per_loop = 50
 
   @flagsaver.flagsaver
@@ -74,9 +74,9 @@ class BertClassifyBenchmarkBase(benchmark_utils.BertBenchmarkBase):
     warmup_steps = int(epochs * steps_per_epoch * 0.1)
     eval_steps = int(
         math.ceil(input_meta_data['eval_data_size'] / FLAGS.eval_batch_size))
-    if self.tpu:
+    if self.default_flags['tpu']:
       strategy = distribution_utils.get_distribution_strategy(
-          distribution_strategy='tpu', tpu_address=self.tpu)
+          distribution_strategy='tpu', tpu_address=self.default_flags['tpu'])
     else:
       strategy = distribution_utils.get_distribution_strategy(
           distribution_strategy='mirrored' if use_ds else 'off',
@@ -211,6 +211,7 @@ class BertClassifyBenchmarkReal(BertClassifyBenchmarkBase):
                                 'summaries/training_summary.txt')
     self._run_and_report_benchmark(summary_path, use_ds=False)
 
+  @owner_utils.Owner('tf-model-garden')
   def benchmark_8_gpu_mrpc(self):
     """Test BERT model performance with 8 GPUs."""
 
@@ -264,6 +265,7 @@ class BertClassifyBenchmarkReal(BertClassifyBenchmarkBase):
                                 'summaries/training_summary.txt')
     self._run_and_report_benchmark(summary_path, use_ds=False)
 
+  @owner_utils.Owner('tf-model-garden')
   def benchmark_2x2_tpu_mrpc(self):
     """Test BERT model performance with 2x2 TPU."""
 
@@ -289,14 +291,14 @@ class BertClassifyAccuracy(BertClassifyBenchmarkBase):
   `benchmark_(number of gpus)_gpu_(dataset type)` format.
   """
 
-  def __init__(self, output_dir=TMP_DIR, **kwargs):
+  def __init__(self, output_dir=TMP_DIR, tpu=None, **kwargs):
     self.train_data_path = CLASSIFIER_TRAIN_DATA_PATH
     self.eval_data_path = CLASSIFIER_EVAL_DATA_PATH
     self.bert_config_file = MODEL_CONFIG_FILE_PATH
     self.input_meta_data_path = CLASSIFIER_INPUT_META_DATA_PATH
     self.pretrained_checkpoint_path = PRETRAINED_CHECKPOINT_PATH
 
-    super(BertClassifyAccuracy, self).__init__(output_dir=output_dir)
+    super(BertClassifyAccuracy, self).__init__(output_dir=output_dir, tpu=tpu)
 
   @benchmark_wrappers.enable_runtime_flags
   def _run_and_report_benchmark(self,
@@ -326,6 +328,7 @@ class BertClassifyAccuracy(BertClassifyBenchmarkBase):
     FLAGS.bert_config_file = self.bert_config_file
     FLAGS.init_checkpoint = self.pretrained_checkpoint_path
 
+  @owner_utils.Owner('tf-model-garden')
   def benchmark_8_gpu_mrpc(self):
     """Run BERT model accuracy test with 8 GPUs.
 
@@ -349,6 +352,16 @@ class BertClassifyAccuracy(BertClassifyBenchmarkBase):
                                 'summaries/training_summary.txt')
     self._run_and_report_benchmark(summary_path)
 
+  @owner_utils.Owner('tf-model-garden')
+  def benchmark_2x2_tpu_mrpc(self):
+    """Run BERT model accuracy test on 2x2 TPU."""
+    self._setup()
+    FLAGS.model_dir = self._get_model_dir('benchmark_2x2_tpu_mrpc')
+
+    summary_path = os.path.join(FLAGS.model_dir,
+                                'summaries/training_summary.txt')
+    self._run_and_report_benchmark(summary_path)
+
 
 if __name__ == '__main__':
   tf.test.main()

official/benchmark/bert_pretrain_benchmark.py

@@ -28,6 +28,7 @@ import tensorflow as tf  # pylint: disable=g-bad-import-order
 
 from official.benchmark import benchmark_wrappers
 from official.benchmark import bert_benchmark_utils
+from official.benchmark import owner_utils
 from official.nlp.bert import run_pretraining
 from official.utils.flags import core as flags_core
 from official.utils.misc import distribution_utils
@@ -64,10 +65,10 @@ class BertPretrainAccuracyBenchmark(bert_benchmark_utils.BertBenchmarkBase):
         output_dir=output_dir, tpu=tpu, **kwargs)
 
   @benchmark_wrappers.enable_runtime_flags
-  def _run_and_report_benchmark(self, summary_path: str):
+  def _run_and_report_benchmark(self, summary_path: str, report_accuracy: bool):
     """Runs and reports the benchmark given the provided configuration."""
     distribution = distribution_utils.get_distribution_strategy(
-        distribution_strategy='tpu', tpu_address=FLAGS.tpu)
+        distribution_strategy='tpu', tpu_address=self.default_flags['tpu'])
     logging.info('Flags: %s', flags_core.get_nondefault_flags_as_str())
     start_time_sec = time.time()
     run_pretraining.run_bert_pretrain(
@@ -76,22 +77,26 @@ class BertPretrainAccuracyBenchmark(bert_benchmark_utils.BertBenchmarkBase):
 
     with tf.io.gfile.GFile(summary_path, 'rb') as reader:
       summary = json.loads(reader.read().decode('utf-8'))
-    self._report_benchmark(summary, start_time_sec, wall_time_sec)
+    self._report_benchmark(summary, start_time_sec, wall_time_sec,
+                           report_accuracy)
 
-  def _report_benchmark(self, summary, start_time_sec, wall_time_sec):
+  def _report_benchmark(self, summary, start_time_sec, wall_time_sec,
+                        report_accuracy):
     metrics = [{
         'name': 'train_loss',
         'value': summary['train_loss'],
     }, {
         'name':
-            'example_per_second',
+            'exp_per_second',
         'value':
             self.timer_callback.get_examples_per_sec(FLAGS.train_batch_size *
                                                      FLAGS.steps_per_loop)
     }, {
         'name': 'startup_time',
         'value': self.timer_callback.get_startup_time(start_time_sec)
-    }, {
+    }]
+    if report_accuracy:
+      metrics.extend([{
           'name': 'masked_lm_accuracy',
           'value': summary['masked_lm_accuracy'],
           'min_value': MIN_MLM_ACCURACY,
@@ -101,7 +106,7 @@ class BertPretrainAccuracyBenchmark(bert_benchmark_utils.BertBenchmarkBase):
           'value': summary['next_sentence_accuracy'],
           'min_value': MIN_NSP_ACCURACY,
           'max_value': MAX_NSP_ACCURACY,
-    }]
+      }])
     self.report_benchmark(
         iters=summary['total_training_steps'],
         wall_time=wall_time_sec,
@@ -120,7 +125,8 @@ class BertPretrainAccuracyBenchmark(bert_benchmark_utils.BertBenchmarkBase):
     FLAGS.max_predictions_per_seq = 20
     FLAGS.dtype = 'bf16'
 
-  def benchmark_8x8_tpu_bf16_seq128_1m_steps(self):
+  @owner_utils.Owner('tf-model-garden')
+  def benchmark_accuracy_8x8_tpu_bf16_seq128_1m_steps(self):
     """Test bert pretraining with 8x8 TPU for 1 million steps."""
     # This is used for accuracy test.
     self._setup()
@@ -128,23 +134,26 @@ class BertPretrainAccuracyBenchmark(bert_benchmark_utils.BertBenchmarkBase):
     FLAGS.num_steps_per_epoch = 250000
     FLAGS.num_train_epochs = 4
     FLAGS.model_dir = self._get_model_dir(
-        'benchmark_8x8_tpu_bf16_seq128_1m_steps')
+        'benchmark_accuracy_8x8_tpu_bf16_seq128_1m_steps')
     summary_path = os.path.join(FLAGS.model_dir,
                                 'summaries/training_summary.txt')
-    self._run_and_report_benchmark(summary_path=summary_path)
+    self._run_and_report_benchmark(summary_path=summary_path,
+                                   report_accuracy=True)
 
-  def benchmark_4x4_tpu_bf16_seq128_1k_steps(self):
-    """Test bert pretraining with 4x4 TPU for 1000 steps."""
-    # This is used for througput test.
+  @owner_utils.Owner('tf-model-garden')
+  def benchmark_perf_8x8_tpu_bf16_seq128_10k_steps(self):
+    """Test bert pretraining with 8x8 TPU for 10000 steps."""
     self._setup()
     self._specify_common_flags()
-    FLAGS.num_steps_per_epoch = 1000
-    FLAGS.num_train_epochs = 1
+    FLAGS.num_steps_per_epoch = 5000
+    FLAGS.num_train_epochs = 2
     FLAGS.model_dir = self._get_model_dir(
-        'benchmark_4x4_tpu_bf16_seq128_1k_steps')
+        'benchmark_perf_8x8_tpu_bf16_seq128_10k_steps')
     summary_path = os.path.join(FLAGS.model_dir,
                                 'summaries/training_summary.txt')
-    self._run_and_report_benchmark(summary_path=summary_path)
+    # Disable accuracy check.
+    self._run_and_report_benchmark(summary_path=summary_path,
+                                   report_accuracy=False)
 
 
 if __name__ == '__main__':

official/benchmark/bert_squad_benchmark.py

@@ -30,6 +30,7 @@ import tensorflow as tf
 # pylint: enable=g-bad-import-order
 
 from official.benchmark import bert_benchmark_utils as benchmark_utils
+from official.benchmark import owner_utils
 from official.nlp.bert import run_squad
 from official.utils.misc import distribution_utils
 from official.utils.misc import keras_utils
@@ -79,9 +80,9 @@ class BertSquadBenchmarkBase(benchmark_utils.BertBenchmarkBase):
     Returns:
       A `tf.distribute.DistibutionStrategy` object.
     """
-    if FLAGS.tpu or ds_type == 'tpu':
+    if self.default_flags['tpu'] or ds_type == 'tpu':
       return distribution_utils.get_distribution_strategy(
-          distribution_strategy='tpu', tpu_address=FLAGS.tpu)
+          distribution_strategy='tpu', tpu_address=self.default_flags['tpu'])
     elif ds_type == 'multi_worker_mirrored':
       # Configures cluster spec for multi-worker distribution strategy.
       _ = distribution_utils.configure_cluster(FLAGS.worker_hosts,
@@ -225,26 +226,7 @@ class BertSquadBenchmarkReal(BertSquadBenchmarkBase):
 
     self._run_and_report_benchmark(ds_type='off', run_eagerly=True)
 
-  def benchmark_2_gpu(self):
-    """Tests BERT SQuAD model performance with 2 GPUs."""
-
-    self._setup()
-    self.num_gpus = 2
-    FLAGS.model_dir = self._get_model_dir('benchmark_2_gpu_squad')
-    FLAGS.train_batch_size = 8
-
-    self._run_and_report_benchmark()
-
-  def benchmark_4_gpu(self):
-    """Tests BERT SQuAD model performance with 4 GPUs."""
-
-    self._setup()
-    self.num_gpus = 4
-    FLAGS.model_dir = self._get_model_dir('benchmark_4_gpu_squad')
-    FLAGS.train_batch_size = 16
-
-    self._run_and_report_benchmark()
-
+  @owner_utils.Owner('tf-model-garden')
   def benchmark_8_gpu(self):
     """Tests BERT SQuAD model performance with 8 GPUs."""
 
@@ -293,30 +275,6 @@ class BertSquadBenchmarkReal(BertSquadBenchmarkBase):
 
     self._run_and_report_benchmark()
 
-  def benchmark_2_gpu_fp16(self):
-    """Tests BERT SQuAD model performance with 2 GPUs and FP16."""
-
-    self._setup()
-    self.num_gpus = 2
-    FLAGS.model_dir = self._get_model_dir('benchmark_2_gpu_squad_fp16')
-    FLAGS.train_batch_size = 8
-    FLAGS.dtype = 'fp16'
-    FLAGS.loss_scale = 'dynamic'
-
-    self._run_and_report_benchmark()
-
-  def benchmark_4_gpu_fp16(self):
-    """Tests BERT SQuAD model performance with 4 GPUs and FP16."""
-
-    self._setup()
-    self.num_gpus = 4
-    FLAGS.model_dir = self._get_model_dir('benchmark_4_gpu_squad_fp16')
-    FLAGS.train_batch_size = 16
-    FLAGS.dtype = 'fp16'
-    FLAGS.loss_scale = 'dynamic'
-
-    self._run_and_report_benchmark()
-
   def benchmark_8_gpu_fp16(self):
     """Tests BERT SQuAD model performance with 8 GPUs."""
 
@@ -355,18 +313,6 @@ class BertSquadBenchmarkReal(BertSquadBenchmarkBase):
 
     self._run_and_report_benchmark()
 
-  def benchmark_4_gpu_amp(self):
-    """Tests BERT SQuAD model performance with 1 GPU with automatic mixed precision."""
-
-    self._setup()
-    self.num_gpus = 4
-    FLAGS.model_dir = self._get_model_dir('benchmark_4_gpu_amp_squad')
-    FLAGS.train_batch_size = 16
-    FLAGS.dtype = 'fp16'
-    FLAGS.fp16_implementation = 'graph_rewrite'
-
-    self._run_and_report_benchmark()
-
   def benchmark_8_gpu_amp(self):
     """Tests BERT SQuAD model performance with 1 GPU with automatic mixed precision."""
 
@@ -380,6 +326,7 @@ class BertSquadBenchmarkReal(BertSquadBenchmarkBase):
 
     self._run_and_report_benchmark()
 
+  @owner_utils.Owner('tf-model-garden')
   def benchmark_2x2_tpu(self):
     """Tests BERT SQuAD model performance with 2x2 TPU."""
 
@@ -449,6 +396,7 @@ class BertSquadAccuracy(BertSquadBenchmarkBase):
 
     self._run_and_report_benchmark(ds_type='off', run_eagerly=True)
 
+  @owner_utils.Owner('tf-model-garden')
   def benchmark_8_gpu(self):
     """Tests BERT SQuAD model accuracy with 8 GPUs."""
 
@@ -485,6 +433,7 @@ class BertSquadAccuracy(BertSquadBenchmarkBase):
 
     self._run_and_report_benchmark()
 
+  @owner_utils.Owner('tf-model-garden')
   def benchmark_2x2_tpu(self):
     """Tests BERT SQuAD model accuracy with 2x2 TPU."""
 

official/benchmark/unet3d_benchmark.py

@@ -26,6 +26,7 @@ import tensorflow as tf  # pylint: disable=g-bad-import-order
 
 from official.benchmark import benchmark_wrappers
 from official.benchmark import keras_benchmark
+from official.benchmark import owner_utils
 from official.vision.segmentation import unet_main as unet_training_lib
 from official.vision.segmentation import unet_model as unet_model_lib
 
@@ -119,6 +120,7 @@ class Unet3DAccuracyBenchmark(keras_benchmark.KerasBenchmark):
   def _get_model_dir(self, folder_name):
     return os.path.join(self.output_dir, folder_name)
 
+  @owner_utils.Owner('tf-model-garden')
   def benchmark_4x4_tpu_bf16(self):
     """Test Keras model with 4x4 TPU, fp16."""
     experiment_name = 'benchmark_4x4_tpu_fp16'

official/benchmark/xlnet_benchmark.py

@@ -29,6 +29,7 @@ import tensorflow as tf
 # pylint: enable=g-bad-import-order
 
 from official.benchmark import bert_benchmark_utils as benchmark_utils
+from official.benchmark import owner_utils
 from official.nlp.xlnet import run_classifier
 from official.nlp.xlnet import run_squad
 from official.benchmark import benchmark_wrappers
@@ -47,8 +48,8 @@ FLAGS = flags.FLAGS
 class XLNetBenchmarkBase(benchmark_utils.BertBenchmarkBase):
   """Base class to hold methods common to test classes in the module."""
 
-  def __init__(self, output_dir=None):
-    super(XLNetBenchmarkBase, self).__init__(output_dir)
+  def __init__(self, output_dir=None, tpu=None):
+    super(XLNetBenchmarkBase, self).__init__(output_dir=output_dir, tpu=tpu)
     self.num_epochs = None
     self.num_steps_per_epoch = None
 
@@ -71,12 +72,12 @@ class XLNetClassifyAccuracy(XLNetBenchmarkBase):
   `benchmark_(number of gpus)_gpu_(dataset type)` format.
   """
 
-  def __init__(self, output_dir=None, **kwargs):
+  def __init__(self, output_dir=None, tpu=None, **kwargs):
     self.train_data_path = CLASSIFIER_TRAIN_DATA_PATH
     self.eval_data_path = CLASSIFIER_EVAL_DATA_PATH
     self.pretrained_checkpoint_path = PRETRAINED_CHECKPOINT_PATH
 
-    super(XLNetClassifyAccuracy, self).__init__(output_dir=output_dir)
+    super(XLNetClassifyAccuracy, self).__init__(output_dir=output_dir, tpu=tpu)
 
   @benchmark_wrappers.enable_runtime_flags
   def _run_and_report_benchmark(self,
@@ -123,6 +124,7 @@ class XLNetClassifyAccuracy(XLNetBenchmarkBase):
     FLAGS.train_tfrecord_path = self.train_data_path
     FLAGS.test_tfrecord_path = self.eval_data_path
 
+  @owner_utils.Owner('tf-model-garden')
   def benchmark_8_gpu_imdb(self):
     """Run XLNet model accuracy test with 8 GPUs."""
     self._setup()
@@ -134,6 +136,19 @@ class XLNetClassifyAccuracy(XLNetBenchmarkBase):
                                 'summaries/training_summary.txt')
     self._run_and_report_benchmark(summary_path)
 
+  @owner_utils.Owner('tf-model-garden')
+  def benchmark_2x2_tpu_imdb(self):
+    """Run XLNet model accuracy test on 2x2 tpu."""
+    self._setup()
+    FLAGS.strategy_type = 'tpu'
+    FLAGS.model_dir = self._get_model_dir('benchmark_2x2_tpu_imdb')
+    # Sets timer_callback to None as we do not use it now.
+    self.timer_callback = None
+
+    summary_path = os.path.join(FLAGS.model_dir,
+                                'summaries/training_summary.txt')
+    self._run_and_report_benchmark(summary_path)
+
 
 class XLNetSquadAccuracy(XLNetBenchmarkBase):
   """Short accuracy test for XLNet squad model.
@@ -143,14 +158,14 @@ class XLNetSquadAccuracy(XLNetBenchmarkBase):
   `benchmark_(number of gpus)_gpu_(dataset type)` format.
   """
 
-  def __init__(self, output_dir=None, **kwargs):
+  def __init__(self, output_dir=None, tpu=None, **kwargs):
     self.train_data_path = SQUAD_DATA_PATH
     self.predict_file = os.path.join(SQUAD_DATA_PATH, "dev-v2.0.json")
     self.test_data_path = os.path.join(SQUAD_DATA_PATH, "12048.eval.tf_record")
     self.spiece_model_file = os.path.join(SQUAD_DATA_PATH, "spiece.cased.model")
     self.pretrained_checkpoint_path = PRETRAINED_CHECKPOINT_PATH
 
-    super(XLNetSquadAccuracy, self).__init__(output_dir=output_dir)
+    super(XLNetSquadAccuracy, self).__init__(output_dir=output_dir, tpu=tpu)
 
   @benchmark_wrappers.enable_runtime_flags
   def _run_and_report_benchmark(self,
@@ -196,9 +211,10 @@ class XLNetSquadAccuracy(XLNetBenchmarkBase):
     FLAGS.test_tfrecord_path = self.test_data_path
     FLAGS.spiece_model_file = self.spiece_model_file
     FLAGS.predict_file = self.predict_file
-    FLAGS.adam_epsilon=1e-6
-    FLAGS.lr_layer_decay_rate=0.75
+    FLAGS.adam_epsilon = 1e-6
+    FLAGS.lr_layer_decay_rate = 0.75
 
+  @owner_utils.Owner('tf-model-garden')
   def benchmark_8_gpu_squadv2(self):
     """Run XLNet model squad v2 accuracy test with 8 GPUs."""
     self._setup()
@@ -211,6 +227,20 @@ class XLNetSquadAccuracy(XLNetBenchmarkBase):
                                 'summaries/training_summary.txt')
     self._run_and_report_benchmark(summary_path)
 
+  @owner_utils.Owner('tf-model-garden')
+  def benchmark_2x2_tpu_squadv2(self):
+    """Run XLNet model squad v2 accuracy test on 2x2 tpu."""
+    self._setup()
+    FLAGS.strategy_type = 'tpu'
+    FLAGS.model_dir = self._get_model_dir('benchmark_2x2_tpu_squadv2')
+    FLAGS.predict_dir = FLAGS.model_dir
+    # Sets timer_callback to None as we do not use it now.
+    self.timer_callback = None
+
+    summary_path = os.path.join(FLAGS.model_dir,
+                                'summaries/training_summary.txt')
+    self._run_and_report_benchmark(summary_path)
+
 
 if __name__ == '__main__':
   tf.test.main()