Remove 'default' in get_distribution_strategy which is complex and error-prone

PiperOrigin-RevId: 282669615

official/benchmark/keras_cifar_benchmark.py

@@ -241,7 +241,7 @@ class Resnet56KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     self._setup()
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu')
     FLAGS.batch_size = 128
     self._run_and_report_benchmark()
@@ -253,7 +253,7 @@ class Resnet56KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.enable_eager = True
     FLAGS.run_eagerly = False
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_xla')
     FLAGS.batch_size = 128
     self._run_and_report_benchmark()
@@ -263,7 +263,7 @@ class Resnet56KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     self._setup()
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_force_v1_path')
     FLAGS.batch_size = 128
     FLAGS.force_v2_in_keras_compile = False
@@ -275,7 +275,7 @@ class Resnet56KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = False
     FLAGS.run_eagerly = False
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu')
     FLAGS.batch_size = 128
     self._run_and_report_benchmark()
@@ -346,7 +346,7 @@ class Resnet56KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 2
     FLAGS.enable_eager = True
     FLAGS.run_eagerly = False
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_2_gpu')
     FLAGS.batch_size = 128 * 2  # 2 GPUs
     self._run_and_report_benchmark()
@@ -357,7 +357,7 @@ class Resnet56KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 2
     FLAGS.enable_eager = False
     FLAGS.run_eagerly = False
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_graph_2_gpu')
     FLAGS.batch_size = 128 * 2  # 2 GPUs
     self._run_and_report_benchmark()

official/benchmark/keras_imagenet_benchmark.py

@@ -319,7 +319,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
 
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu')
     FLAGS.batch_size = 128
     self._run_and_report_benchmark()
@@ -332,7 +332,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.enable_eager = True
     FLAGS.dtype = 'fp16'
     FLAGS.fp16_implementation = 'graph_rewrite'
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_amp')
     FLAGS.batch_size = 256
     self._run_and_report_benchmark()
@@ -344,7 +344,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = True
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu')
     FLAGS.batch_size = 128
     self._run_and_report_benchmark()
@@ -358,7 +358,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.dtype = 'fp16'
     FLAGS.fp16_implementation = 'graph_rewrite'
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_amp')
     FLAGS.batch_size = 256
     self._run_and_report_benchmark()
@@ -369,7 +369,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
 
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_fp16')
     FLAGS.dtype = 'fp16'
     FLAGS.batch_size = 256
@@ -381,7 +381,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
 
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_fp16_dynamic')
     FLAGS.dtype = 'fp16'
     FLAGS.batch_size = 256
@@ -395,7 +395,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = True
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16')
     FLAGS.dtype = 'fp16'
     FLAGS.batch_size = 256
@@ -408,7 +408,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = True
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16_tweaked')
     FLAGS.dtype = 'fp16'
     FLAGS.batch_size = 256
@@ -423,7 +423,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = True
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16_dynamic')
     FLAGS.dtype = 'fp16'
     FLAGS.batch_size = 256
@@ -436,7 +436,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
 
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = False
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu')
     FLAGS.batch_size = 128
     self._run_and_report_benchmark()
@@ -448,7 +448,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = False
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_1_gpu')
     FLAGS.batch_size = 128
     self._run_and_report_benchmark()
@@ -460,7 +460,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 1
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = False
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu_fp16')
     FLAGS.batch_size = 256
     self._run_and_report_benchmark()
@@ -473,7 +473,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = False
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_1_gpu_fp16')
     FLAGS.batch_size = 256
     self._run_and_report_benchmark()
@@ -485,7 +485,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 1
     FLAGS.enable_eager = False
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir(
         'benchmark_graph_xla_1_gpu_fp16_tweaked')
     FLAGS.dtype = 'fp16'
@@ -500,7 +500,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
 
     FLAGS.num_gpus = 8
     FLAGS.enable_eager = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
     FLAGS.batch_size = 128 * 8  # 8 GPUs
     self._run_and_report_benchmark()
@@ -513,7 +513,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.enable_eager = True
     FLAGS.dtype = 'fp16'
     FLAGS.fp16_implementation = 'graph_rewrite'
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_amp')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
     self._run_and_report_benchmark()
@@ -524,7 +524,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
 
     FLAGS.num_gpus = 8
     FLAGS.enable_eager = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_tweaked')
     FLAGS.batch_size = 128 * 8  # 8 GPUs
     FLAGS.use_tensor_lr = True
@@ -538,7 +538,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 8
     FLAGS.enable_eager = True
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu')
     FLAGS.batch_size = 128 * 8  # 8 GPUs
     self._run_and_report_benchmark()
@@ -552,7 +552,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.dtype = 'fp16'
     FLAGS.fp16_implementation = 'graph_rewrite'
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_amp')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
     self._run_and_report_benchmark()
@@ -564,7 +564,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 8
     FLAGS.enable_eager = True
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_tweaked')
     FLAGS.batch_size = 128 * 8
     FLAGS.use_tensor_lr = True
@@ -579,7 +579,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 8
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
     self._run_and_report_benchmark()
@@ -591,7 +591,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 8
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16_tweaked')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
     FLAGS.use_tensor_lr = True
@@ -605,7 +605,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 8
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir(
         'benchmark_8_gpu_fp16_dynamic_tweaked')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
@@ -622,7 +622,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = True
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_fp16')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
     self._run_and_report_benchmark()
@@ -635,7 +635,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = True
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_fp16_tweaked')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
     FLAGS.use_tensor_lr = True
@@ -654,7 +654,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = True
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir(
         'benchmark_xla_8_gpu_fp16_tweaked_delay_measure')
     FLAGS.batch_size = 256 * 8
@@ -671,7 +671,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = True
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir(
         'benchmark_xla_8_gpu_fp16_dynamic_tweaked')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
@@ -687,7 +687,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
 
     FLAGS.num_gpus = 8
     FLAGS.enable_eager = False
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu')
     FLAGS.batch_size = 128 * 8  # 8 GPUs
     self._run_and_report_benchmark()
@@ -699,7 +699,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 8
     FLAGS.enable_eager = False
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_8_gpu')
     FLAGS.batch_size = 128 * 8  # 8 GPUs
     self._run_and_report_benchmark()
@@ -711,7 +711,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 8
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = False
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu_fp16')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
     self._run_and_report_benchmark()
@@ -724,7 +724,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = False
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_8_gpu_fp16')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
     self._run_and_report_benchmark()
@@ -736,7 +736,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 8
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = False
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu_fp16_tweaked')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
     FLAGS.use_tensor_lr = True
@@ -751,7 +751,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = False
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir(
         'benchmark_graph_xla_8_gpu_fp16_tweaked')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
@@ -770,7 +770,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = False
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir(
         'benchmark_graph_xla_8_gpu_fp16_tweaked_delay_measure')
     FLAGS.batch_size = 256 * 8
@@ -786,7 +786,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.num_gpus = 8
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = False
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir(
         'benchmark_graph_8_gpu_fp16_dynamic_tweaked')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
@@ -803,7 +803,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.dtype = 'fp16'
     FLAGS.enable_eager = False
     FLAGS.enable_xla = True
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir(
         'benchmark_graph_xla_8_gpu_fp16_dynamic_tweaked')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
@@ -888,7 +888,7 @@ class TrivialKerasBenchmarkReal(keras_benchmark.KerasBenchmark):
     def_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
     def_flags['train_steps'] = 600
     def_flags['log_steps'] = 100
-    def_flags['distribution_strategy'] = 'default'
+    def_flags['distribution_strategy'] = 'mirrored'
 
     super(TrivialKerasBenchmarkReal, self).__init__(
         output_dir=output_dir,

official/benchmark/resnet_ctl_imagenet_benchmark.py

@@ -231,7 +231,7 @@ class Resnet50CtlBenchmarkBase(CtlBenchmark):
     self._setup()
 
     FLAGS.num_gpus = 1
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu')
     FLAGS.batch_size = 128
     self._run_and_report_benchmark()
@@ -241,7 +241,7 @@ class Resnet50CtlBenchmarkBase(CtlBenchmark):
     self._setup()
 
     FLAGS.num_gpus = 1
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_amp')
     FLAGS.batch_size = 256
     FLAGS.dtype = 'fp16'
@@ -253,7 +253,7 @@ class Resnet50CtlBenchmarkBase(CtlBenchmark):
     self._setup()
 
     FLAGS.num_gpus = 1
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_amp')
     FLAGS.batch_size = 256
     FLAGS.dtype = 'fp16'
@@ -266,7 +266,7 @@ class Resnet50CtlBenchmarkBase(CtlBenchmark):
     self._setup()
 
     FLAGS.num_gpus = 1
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'one_device'
     FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_eager')
     FLAGS.batch_size = 64
     FLAGS.use_tf_function = False
@@ -278,7 +278,7 @@ class Resnet50CtlBenchmarkBase(CtlBenchmark):
     self._setup()
 
     FLAGS.num_gpus = 8
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
     FLAGS.batch_size = 128 * 8  # 8 GPUs
     self._run_and_report_benchmark()
@@ -288,7 +288,7 @@ class Resnet50CtlBenchmarkBase(CtlBenchmark):
     self._setup()
 
     FLAGS.num_gpus = 8
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_amp')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
     FLAGS.dtype = 'fp16'
@@ -300,7 +300,7 @@ class Resnet50CtlBenchmarkBase(CtlBenchmark):
     self._setup()
 
     FLAGS.num_gpus = 8
-    FLAGS.distribution_strategy = 'default'
+    FLAGS.distribution_strategy = 'mirrored'
     FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_amp')
     FLAGS.batch_size = 256 * 8  # 8 GPUs
     FLAGS.dtype = 'fp16'

official/utils/flags/_base.py

@@ -139,9 +139,9 @@ def define_base(data_dir=True, model_dir=True, clean=False, train_epochs=False,
 
   if distribution_strategy:
     flags.DEFINE_string(
-        name="distribution_strategy", short_name="ds", default="default",
+        name="distribution_strategy", short_name="ds", default="mirrored",
         help=help_wrap("The Distribution Strategy to use for training. "
-                       "Accepted values are 'off', 'default', 'one_device', "
+                       "Accepted values are 'off', 'one_device', "
                        "'mirrored', 'parameter_server', 'collective', "
                        "case insensitive. 'off' means not to use "
                        "Distribution Strategy; 'default' means to choose "

official/utils/misc/distribution_utils.py

@@ -81,7 +81,7 @@ def _mirrored_cross_device_ops(all_reduce_alg, num_packs):
   return cross_device_ops_class(num_packs=num_packs)
 
 
-def get_distribution_strategy(distribution_strategy="default",
+def get_distribution_strategy(distribution_strategy="mirrored",
                               num_gpus=0,
                               num_workers=1,
                               all_reduce_alg=None,
@@ -91,11 +91,9 @@ def get_distribution_strategy(distribution_strategy="default",
 
   Args:
     distribution_strategy: a string specifying which distribution strategy to
-      use. Accepted values are 'off', 'default', 'one_device', 'mirrored',
+      use. Accepted values are 'off', 'one_device', 'mirrored',
       'parameter_server', 'multi_worker_mirrored', and 'tpu' -- case insensitive.
-      'off' means not to use Distribution Strategy; 'default' means to choose from
-      `MirroredStrategy`, `MultiWorkerMirroredStrategy`, or `OneDeviceStrategy`
-      according to the number of GPUs and number of workers. 'tpu' means to use
+      'off' means not to use Distribution Strategy; 'tpu' means to use
       TPUStrategy using `tpu_address`.
     num_gpus: Number of GPUs to run this model.
     num_workers: Number of workers to run this model.
@@ -135,19 +133,16 @@ def get_distribution_strategy(distribution_strategy="default",
     return tf.distribute.experimental.MultiWorkerMirroredStrategy(
         communication=_collective_communication(all_reduce_alg))
 
-  if (distribution_strategy == "one_device" or
-      (distribution_strategy == "default" and num_gpus <= 1)):
+  if distribution_strategy == "one_device":
     if num_gpus == 0:
       return tf.distribute.OneDeviceStrategy("device:CPU:0")
-    else:
     if num_gpus > 1:
       raise ValueError("`OneDeviceStrategy` can not be used for more than "
                        "one device.")
     return tf.distribute.OneDeviceStrategy("device:GPU:0")
 
-  if distribution_strategy in ("mirrored", "default"):
+  if distribution_strategy == "mirrored":
     if num_gpus == 0:
-      assert distribution_strategy == "mirrored"
       devices = ["device:CPU:0"]
     else:
       devices = ["device:GPU:%d" % i for i in range(num_gpus)]

official/vision/image_classification/resnet_cifar_test.py

@@ -104,7 +104,7 @@ class KerasCifarTest(googletest.TestCase):
 
     extra_flags = [
         "-num_gpus", "1",
-        "-distribution_strategy", "default",
+        "-distribution_strategy", "mirrored",
         "-model_dir", "keras_cifar_1_gpu",
         "-data_format", "channels_last",
     ]
@@ -126,7 +126,7 @@ class KerasCifarTest(googletest.TestCase):
     extra_flags = [
         "-num_gpus", "1",
         "-noenable_eager",
-        "-distribution_strategy", "default",
+        "-distribution_strategy", "mirrored",
         "-model_dir", "keras_cifar_graph_1_gpu",
         "-data_format", "channels_last",
     ]
@@ -150,7 +150,7 @@ class KerasCifarTest(googletest.TestCase):
 
     extra_flags = [
         "-num_gpus", "2",
-        "-distribution_strategy", "default",
+        "-distribution_strategy", "mirrored",
         "-model_dir", "keras_cifar_2_gpu",
     ]
     extra_flags = extra_flags + self._extra_flags
@@ -171,7 +171,7 @@ class KerasCifarTest(googletest.TestCase):
     extra_flags = [
         "-num_gpus", "2",
         "-enable_eager", "false",
-        "-distribution_strategy", "default",
+        "-distribution_strategy", "mirrored",
         "-model_dir", "keras_cifar_graph_2_gpu",
     ]
     extra_flags = extra_flags + self._extra_flags

official/vision/image_classification/resnet_ctl_imagenet_test.py

@@ -82,7 +82,7 @@ class CtlImagenetTest(tf.test.TestCase):
 
     extra_flags = [
         '-num_gpus', num_gpus,
-        '-distribution_strategy', 'default',
+        '-distribution_strategy', 'mirrored',
         '-model_dir', 'ctl_imagenet_2_gpu',
         '-data_format', 'channels_last',
     ]

official/vision/image_classification/resnet_imagenet_test.py

@@ -94,7 +94,7 @@ class KerasImagenetTest(tf.test.TestCase):
 
     extra_flags = [
         "-num_gpus", "1",
-        "-distribution_strategy", "default",
+        "-distribution_strategy", "mirrored",
         "-data_format", "channels_last",
         "-enable_checkpoint_and_export", "1",
     ]
@@ -119,7 +119,7 @@ class KerasImagenetTest(tf.test.TestCase):
     extra_flags = [
         "-num_gpus", "1",
         "-dtype", "fp16",
-        "-distribution_strategy", "default",
+        "-distribution_strategy", "mirrored",
         "-data_format", "channels_last",
     ]
     extra_flags = extra_flags + self._extra_flags
@@ -143,7 +143,7 @@ class KerasImagenetTest(tf.test.TestCase):
 
     extra_flags = [
         "-num_gpus", "2",
-        "-distribution_strategy", "default",
+        "-distribution_strategy", "mirrored",
     ]
     extra_flags = extra_flags + self._extra_flags
 
@@ -166,7 +166,7 @@ class KerasImagenetTest(tf.test.TestCase):
     extra_flags = [
         "-num_gpus", "2",
         "-enable_xla", "true",
-        "-distribution_strategy", "default",
+        "-distribution_strategy", "mirrored",
     ]
     extra_flags = extra_flags + self._extra_flags
 
@@ -189,7 +189,7 @@ class KerasImagenetTest(tf.test.TestCase):
     extra_flags = [
         "-num_gpus", "2",
         "-dtype", "fp16",
-        "-distribution_strategy", "default",
+        "-distribution_strategy", "mirrored",
     ]
     extra_flags = extra_flags + self._extra_flags
 
@@ -213,7 +213,7 @@ class KerasImagenetTest(tf.test.TestCase):
         "-num_gpus", "2",
         "-dtype", "fp16",
         "-enable_xla", "true",
-        "-distribution_strategy", "default",
+        "-distribution_strategy", "mirrored",
     ]
     extra_flags = extra_flags + self._extra_flags