Add tweaked cloning tests (#6916)

official/resnet/keras/keras_imagenet_benchmark.py

@@ -510,6 +510,23 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.data_delay_prefetch = True
     self._run_and_report_benchmark()
 
+  def benchmark_8_gpu_fp16_cloning_tweaked(self):
+    """Test Keras model with 8 GPUs, fp16, cloning, and manual config tuning."""
+    self._setup()
+
+    FLAGS.num_gpus = 8
+    FLAGS.dtype = 'fp16'
+    FLAGS.enable_eager = True
+    FLAGS.distribution_strategy = 'default'
+    FLAGS.clone_model_in_keras_dist_strat = True
+    FLAGS.model_dir = self._get_model_dir(
+        'benchmark_8_gpu_fp16_cloning_tweaked')
+    FLAGS.batch_size = 256 * 8
+    FLAGS.use_tensor_lr = True
+    FLAGS.tf_gpu_thread_mode = 'gpu_private'
+    FLAGS.data_delay_prefetch = True
+    self._run_and_report_benchmark()
+
   def benchmark_8_gpu_fp16_dynamic_tweaked(self):
     """Test Keras model with 8 GPUs, fp16, dynamic loss scaling, and manual
        config tuning.
@@ -572,6 +589,26 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.data_delay_prefetch = True
     self._run_and_report_benchmark()
 
+  def benchmark_xla_8_gpu_fp16_cloning_tweaked(self):
+    """Test Keras model with manual config tuning, XLA, 8 GPUs, fp16, and
+       cloning.
+    """
+    self._setup()
+
+    FLAGS.num_gpus = 8
+    FLAGS.dtype = 'fp16'
+    FLAGS.enable_eager = True
+    FLAGS.enable_xla = True
+    FLAGS.distribution_strategy = 'default'
+    FLAGS.clone_model_in_keras_dist_strat = True
+    FLAGS.model_dir = self._get_model_dir(
+        'benchmark_xla_8_gpu_fp16_cloning_tweaked')
+    FLAGS.batch_size = 256 * 8
+    FLAGS.use_tensor_lr = True
+    # FLAGS.tf_gpu_thread_mode = 'gpu_private'
+    FLAGS.data_delay_prefetch = True
+    self._run_and_report_benchmark()
+
   def benchmark_xla_8_gpu_fp16_tweaked_delay_measure(self):
     """Test Keras model with manual config tuning, XLA, 8 GPUs and fp16. Delay
        performance measurement for stable performance on 96 vCPU platforms.
@@ -592,6 +629,28 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
     FLAGS.train_steps = 310
     self._run_and_report_benchmark()
 
+  def benchmark_xla_8_gpu_fp16_cloning_tweaked_delay_measure(self):
+    """Test Keras model with manual config tuning, XLA, 8 GPUs, fp16, and
+       cloning. Delay performance measurement for stable performance on 96 vCPU
+       platforms.
+    """
+    self._setup()
+
+    FLAGS.num_gpus = 8
+    FLAGS.dtype = 'fp16'
+    FLAGS.enable_eager = True
+    FLAGS.enable_xla = True
+    FLAGS.distribution_strategy = 'default'
+    FLAGS.clone_model_in_keras_dist_strat = True
+    FLAGS.model_dir = self._get_model_dir(
+        'benchmark_xla_8_gpu_fp16_cloning_tweaked_delay_measure')
+    FLAGS.batch_size = 256 * 8
+    FLAGS.use_tensor_lr = True
+    FLAGS.tf_gpu_thread_mode = 'gpu_private'
+    FLAGS.data_delay_prefetch = True
+    FLAGS.train_steps = 310
+    self._run_and_report_benchmark()
+
   def benchmark_xla_8_gpu_fp16_tweaked_optional_next(self):
     """Test Keras model with manual config tuning, XLA, 8 GPUs, fp16.