Skip to content

GitLab

Commit 0672bdcf authored by Zongwei Zhou's avatar Zongwei Zhou Committed by A. Unique TensorFlower
Browse files

Set dataset private threads and GPU private threads together for better Resnet... 

Set dataset private threads and GPU private threads together for better Resnet performance; remove outdated TF1 benchmarks; remove redundant classifier param runtime.gpu_threads_enabled

PiperOrigin-RevId: 306341022
parent 3db5347d
Hide whitespace changes
Inline Side-by-side
Showing with 13 additions and 342 deletions
official/benchmark/keras_imagenet_benchmark.py
View file @ 0672bdcf
...@@ -187,8 +187,7 @@ class Resnet50KerasAccuracy(keras_benchmark.KerasBenchmark): ...@@ -187,8 +187,7 @@ class Resnet50KerasAccuracy(keras_benchmark.KerasBenchmark):
per_replica_batch_size=128, per_replica_batch_size=128,
epochs=90, epochs=90,
epochs_between_evals=10, epochs_between_evals=10,
dtype='float32', dtype='float32')
dataset_num_private_threads=14)
def benchmark_8_gpu_fp16(self): def benchmark_8_gpu_fp16(self):
"""Tests Keras model with eager, dist_strat, 8 GPUs, and fp16.""" """Tests Keras model with eager, dist_strat, 8 GPUs, and fp16."""
...@@ -199,8 +198,7 @@ class Resnet50KerasAccuracy(keras_benchmark.KerasBenchmark): ...@@ -199,8 +198,7 @@ class Resnet50KerasAccuracy(keras_benchmark.KerasBenchmark):
per_replica_batch_size=256, per_replica_batch_size=256,
epochs=90, epochs=90,
epochs_between_evals=10, epochs_between_evals=10,
dtype='float16', dtype='float16')
gpu_thread_mode='gpu_private')
def benchmark_xla_8_gpu_fp16(self): def benchmark_xla_8_gpu_fp16(self):
"""Tests Keras model with XLA, eager, dist_strat, 8 GPUs and fp16.""" """Tests Keras model with XLA, eager, dist_strat, 8 GPUs and fp16."""
...@@ -212,8 +210,7 @@ class Resnet50KerasAccuracy(keras_benchmark.KerasBenchmark): ...@@ -212,8 +210,7 @@ class Resnet50KerasAccuracy(keras_benchmark.KerasBenchmark):
epochs=90, epochs=90,
epochs_between_evals=10, epochs_between_evals=10,
dtype='float16', dtype='float16',
enable_xla=True, enable_xla=True)
gpu_thread_mode='gpu_private')
def benchmark_xla_8_gpu_fp16_dynamic(self): def benchmark_xla_8_gpu_fp16_dynamic(self):
"""Tests Keras model with XLA, eager, dist_strat, 8 GPUs, dynamic fp16.""" """Tests Keras model with XLA, eager, dist_strat, 8 GPUs, dynamic fp16."""
...@@ -226,8 +223,7 @@ class Resnet50KerasAccuracy(keras_benchmark.KerasBenchmark): ...@@ -226,8 +223,7 @@ class Resnet50KerasAccuracy(keras_benchmark.KerasBenchmark):
epochs=90, epochs=90,
epochs_between_evals=10, epochs_between_evals=10,
dtype='float16', dtype='float16',
loss_scale='dynamic', loss_scale='dynamic')
gpu_thread_mode='gpu_private')
def _get_model_dir(self, folder_name): def _get_model_dir(self, folder_name):
return os.path.join(self.output_dir, folder_name) return os.path.join(self.output_dir, folder_name)
...@@ -270,8 +266,6 @@ class MobilenetV1KerasAccuracy(keras_benchmark.KerasBenchmark): ...@@ -270,8 +266,6 @@ class MobilenetV1KerasAccuracy(keras_benchmark.KerasBenchmark):
FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu') FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
FLAGS.dtype = 'fp32' FLAGS.dtype = 'fp32'
FLAGS.enable_eager = True FLAGS.enable_eager = True
# Add some thread tunings to improve performance.
FLAGS.datasets_num_private_threads = 14
self._run_and_report_benchmark() self._run_and_report_benchmark()
@benchmark_wrappers.enable_runtime_flags @benchmark_wrappers.enable_runtime_flags
...@@ -474,25 +468,6 @@ class Resnet50KerasClassifierBenchmarkBase(keras_benchmark.KerasBenchmark): ...@@ -474,25 +468,6 @@ class Resnet50KerasClassifierBenchmarkBase(keras_benchmark.KerasBenchmark):
per_replica_batch_size=256, per_replica_batch_size=256,
loss_scale='dynamic') loss_scale='dynamic')
def benchmark_graph_1_gpu(self):
"""Tests Keras model in legacy graph mode with 1 GPU."""
self._setup()
self._run_and_report_benchmark(
experiment_name='benchmark_graph_1_gpu',
num_gpus=1,
distribution_strategy='one_device',
per_replica_batch_size=128)
def benchmark_graph_xla_1_gpu(self):
"""Tests Keras model in legacy graph mode with XLA and 1 GPU."""
self._setup()
self._run_and_report_benchmark(
experiment_name='benchmark_graph_xla_1_gpu',
num_gpus=1,
enable_xla=True,
distribution_strategy='one_device',
per_replica_batch_size=128)
def benchmark_8_gpu(self): def benchmark_8_gpu(self):
"""Tests Keras model with 8 GPUs.""" """Tests Keras model with 8 GPUs."""
self._setup() self._setup()
...@@ -553,7 +528,8 @@ class Resnet50KerasClassifierBenchmarkBase(keras_benchmark.KerasBenchmark): ...@@ -553,7 +528,8 @@ class Resnet50KerasClassifierBenchmarkBase(keras_benchmark.KerasBenchmark):
dtype='float16', dtype='float16',
distribution_strategy='mirrored', distribution_strategy='mirrored',
per_replica_batch_size=256, per_replica_batch_size=256,
gpu_thread_mode='gpu_private') gpu_thread_mode='gpu_private',
dataset_num_private_threads=40)
def benchmark_8_gpu_fp16_dynamic_tweaked(self): def benchmark_8_gpu_fp16_dynamic_tweaked(self):
"""Tests Keras model with 8 GPUs, fp16, dynamic loss scaling, and tuned.""" """Tests Keras model with 8 GPUs, fp16, dynamic loss scaling, and tuned."""
...@@ -565,7 +541,8 @@ class Resnet50KerasClassifierBenchmarkBase(keras_benchmark.KerasBenchmark): ...@@ -565,7 +541,8 @@ class Resnet50KerasClassifierBenchmarkBase(keras_benchmark.KerasBenchmark):
distribution_strategy='mirrored', distribution_strategy='mirrored',
per_replica_batch_size=256, per_replica_batch_size=256,
loss_scale='dynamic', loss_scale='dynamic',
gpu_thread_mode='gpu_private') gpu_thread_mode='gpu_private',
dataset_num_private_threads=40)
def benchmark_xla_8_gpu_fp16(self): def benchmark_xla_8_gpu_fp16(self):
"""Tests Keras model with XLA, 8 GPUs and fp16.""" """Tests Keras model with XLA, 8 GPUs and fp16."""
...@@ -605,6 +582,7 @@ class Resnet50KerasClassifierBenchmarkBase(keras_benchmark.KerasBenchmark): ...@@ -605,6 +582,7 @@ class Resnet50KerasClassifierBenchmarkBase(keras_benchmark.KerasBenchmark):
distribution_strategy='mirrored', distribution_strategy='mirrored',
per_replica_batch_size=256, per_replica_batch_size=256,
gpu_thread_mode='gpu_private', gpu_thread_mode='gpu_private',
dataset_num_private_threads=48,
steps=310) steps=310)
def benchmark_xla_8_gpu_fp16_dynamic_tweaked(self): def benchmark_xla_8_gpu_fp16_dynamic_tweaked(self):
...@@ -621,25 +599,6 @@ class Resnet50KerasClassifierBenchmarkBase(keras_benchmark.KerasBenchmark): ...@@ -621,25 +599,6 @@ class Resnet50KerasClassifierBenchmarkBase(keras_benchmark.KerasBenchmark):
loss_scale='dynamic', loss_scale='dynamic',
dataset_num_private_threads=48) dataset_num_private_threads=48)
def benchmark_graph_8_gpu(self):
"""Tests Keras model in legacy graph mode with 8 GPUs."""
self._setup()
self._run_and_report_benchmark(
experiment_name='benchmark_graph_8_gpu',
num_gpus=8,
distribution_strategy='mirrored',
per_replica_batch_size=128)
def benchmark_graph_xla_8_gpu(self):
"""Tests Keras model in legacy graph mode with XLA and 8 GPUs."""
self._setup()
self._run_and_report_benchmark(
experiment_name='benchmark_graph_xla_8_gpu',
num_gpus=8,
enable_xla=True,
distribution_strategy='mirrored',
per_replica_batch_size=128)
def benchmark_2x2_tpu_fp16(self): def benchmark_2x2_tpu_fp16(self):
"""Test Keras model with 2x2 TPU, fp16.""" """Test Keras model with 2x2 TPU, fp16."""
self._setup() self._setup()
...@@ -762,18 +721,6 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark): ...@@ -762,18 +721,6 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
FLAGS.batch_size = 128 FLAGS.batch_size = 128
self._run_and_report_benchmark() self._run_and_report_benchmark()
def benchmark_graph_1_gpu_no_dist_strat(self):
"""Test Keras model in legacy graph mode with 1 GPU, no dist strat."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.enable_eager = False
FLAGS.distribution_strategy = 'off'
FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu_no_dist_strat')
FLAGS.batch_size = 96 # BatchNorm is less efficient in legacy graph mode
# due to its reliance on v1 cond.
self._run_and_report_benchmark()
def benchmark_1_gpu(self): def benchmark_1_gpu(self):
"""Test Keras model with 1 GPU.""" """Test Keras model with 1 GPU."""
self._setup() self._setup()
...@@ -890,69 +837,6 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark): ...@@ -890,69 +837,6 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
FLAGS.loss_scale = 'dynamic' FLAGS.loss_scale = 'dynamic'
self._run_and_report_benchmark() self._run_and_report_benchmark()
def benchmark_graph_1_gpu(self):
"""Test Keras model in legacy graph mode with 1 GPU."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.enable_eager = False
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()
def benchmark_graph_xla_1_gpu(self):
"""Test Keras model in legacy graph mode with XLA and 1 GPU."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.enable_eager = False
FLAGS.enable_xla = True
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()
def benchmark_graph_1_gpu_fp16(self):
"""Test Keras model in legacy graph mode with 1 GPU and fp16."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.dtype = 'fp16'
FLAGS.enable_eager = False
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()
def benchmark_graph_xla_1_gpu_fp16(self):
"""Test Keras model in legacy graph mode with 1 GPU, fp16 and XLA."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.dtype = 'fp16'
FLAGS.enable_eager = False
FLAGS.enable_xla = True
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()
def benchmark_graph_xla_1_gpu_fp16_tweaked(self):
"""Test Keras model in legacy graph with 1 GPU, fp16, XLA, and tuning."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.enable_eager = False
FLAGS.enable_xla = True
FLAGS.distribution_strategy = 'one_device'
FLAGS.model_dir = self._get_model_dir(
'benchmark_graph_xla_1_gpu_fp16_tweaked')
FLAGS.dtype = 'fp16'
FLAGS.batch_size = 256
FLAGS.tf_gpu_thread_mode = 'gpu_private'
self._run_and_report_benchmark()
def benchmark_8_gpu(self): def benchmark_8_gpu(self):
"""Test Keras model with 8 GPUs.""" """Test Keras model with 8 GPUs."""
self._setup() self._setup()
...@@ -1052,6 +936,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark): ...@@ -1052,6 +936,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16_tweaked') FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16_tweaked')
FLAGS.batch_size = 256 * 8 # 8 GPUs FLAGS.batch_size = 256 * 8 # 8 GPUs
FLAGS.tf_gpu_thread_mode = 'gpu_private' FLAGS.tf_gpu_thread_mode = 'gpu_private'
FLAGS.dataset_num_private_threads = 40
self._run_and_report_benchmark() self._run_and_report_benchmark()
def benchmark_8_gpu_fp16_dynamic_tweaked(self): def benchmark_8_gpu_fp16_dynamic_tweaked(self):
...@@ -1067,6 +952,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark): ...@@ -1067,6 +952,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
FLAGS.batch_size = 256 * 8 # 8 GPUs FLAGS.batch_size = 256 * 8 # 8 GPUs
FLAGS.loss_scale = 'dynamic' FLAGS.loss_scale = 'dynamic'
FLAGS.tf_gpu_thread_mode = 'gpu_private' FLAGS.tf_gpu_thread_mode = 'gpu_private'
FLAGS.dataset_num_private_threads = 40
self._run_and_report_benchmark() self._run_and_report_benchmark()
def benchmark_xla_8_gpu_fp16(self): def benchmark_xla_8_gpu_fp16(self):
...@@ -1113,6 +999,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark): ...@@ -1113,6 +999,7 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
'benchmark_xla_8_gpu_fp16_tweaked_delay_measure') 'benchmark_xla_8_gpu_fp16_tweaked_delay_measure')
FLAGS.batch_size = 256 * 8 FLAGS.batch_size = 256 * 8
FLAGS.tf_gpu_thread_mode = 'gpu_private' FLAGS.tf_gpu_thread_mode = 'gpu_private'
FLAGS.datasets_num_private_threads = 48
FLAGS.train_steps = 310 FLAGS.train_steps = 310
self._run_and_report_benchmark() self._run_and_report_benchmark()
...@@ -1133,132 +1020,6 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark): ...@@ -1133,132 +1020,6 @@ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
FLAGS.datasets_num_private_threads = 48 FLAGS.datasets_num_private_threads = 48
self._run_and_report_benchmark() self._run_and_report_benchmark()
def benchmark_graph_8_gpu(self):
"""Test Keras model in legacy graph mode with 8 GPUs."""
self._setup()
FLAGS.num_gpus = 8
FLAGS.enable_eager = False
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()
def benchmark_graph_xla_8_gpu(self):
"""Test Keras model in legacy graph mode with XLA and 8 GPUs."""
self._setup()
FLAGS.num_gpus = 8
FLAGS.enable_eager = False
FLAGS.enable_xla = True
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()
def benchmark_graph_8_gpu_fp16(self):
"""Test Keras model in legacy graph mode with 8 GPUs and fp16."""
self._setup()
FLAGS.num_gpus = 8
FLAGS.dtype = 'fp16'
FLAGS.enable_eager = False
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()
def benchmark_graph_xla_8_gpu_fp16(self):
"""Test Keras model in legacy graph mode with XLA, 8 GPUs and fp16."""
self._setup()
FLAGS.num_gpus = 8
FLAGS.dtype = 'fp16'
FLAGS.enable_eager = False
FLAGS.enable_xla = True
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()
def benchmark_graph_8_gpu_fp16_tweaked(self):
"""Test Keras model in legacy graph mode, tuning, 8 GPUs, and FP16."""
self._setup()
FLAGS.num_gpus = 8
FLAGS.dtype = 'fp16'
FLAGS.enable_eager = False
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.tf_gpu_thread_mode = 'gpu_private'
self._run_and_report_benchmark()
def benchmark_graph_xla_8_gpu_fp16_tweaked(self):
"""Test Keras model in legacy graph tuning, XLA_FP16, 8 GPUs and fp16."""
self._setup()
FLAGS.num_gpus = 8
FLAGS.dtype = 'fp16'
FLAGS.enable_eager = False
FLAGS.enable_xla = True
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
FLAGS.tf_gpu_thread_mode = 'gpu_private'
self._run_and_report_benchmark()
def benchmark_graph_xla_8_gpu_fp16_tweaked_delay_measure(self):
"""Test in legacy graph mode with manual config tuning, XLA, 8 GPUs, fp16.
Delay performance measurement for stable performance on 96 vCPU platforms.
"""
self._setup()
FLAGS.num_gpus = 8
FLAGS.dtype = 'fp16'
FLAGS.enable_eager = False
FLAGS.enable_xla = True
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
FLAGS.tf_gpu_thread_mode = 'gpu_private'
FLAGS.train_steps = 310
self._run_and_report_benchmark()
def benchmark_graph_8_gpu_fp16_dynamic_tweaked(self):
"""Test graph Keras with config tuning, 8 GPUs and dynamic fp16."""
self._setup()
FLAGS.num_gpus = 8
FLAGS.dtype = 'fp16'
FLAGS.enable_eager = False
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
FLAGS.loss_scale = 'dynamic'
FLAGS.tf_gpu_thread_mode = 'gpu_private'
self._run_and_report_benchmark()
def benchmark_graph_xla_8_gpu_fp16_dynamic_tweaked(self):
"""Test graph Keras with config tuning, XLA, 8 GPUs and dynamic fp16."""
self._setup()
FLAGS.num_gpus = 8
FLAGS.dtype = 'fp16'
FLAGS.enable_eager = False
FLAGS.enable_xla = True
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
FLAGS.loss_scale = 'dynamic'
FLAGS.tf_gpu_thread_mode = 'gpu_private'
self._run_and_report_benchmark()
def benchmark_2x2_tpu_fp16(self): def benchmark_2x2_tpu_fp16(self):
"""Test Keras model with 2x2 TPU, fp16.""" """Test Keras model with 2x2 TPU, fp16."""
self._setup() self._setup()
...@@ -1410,19 +1171,6 @@ class Resnet50KerasBenchmarkRemoteData(Resnet50KerasBenchmarkBase): ...@@ -1410,19 +1171,6 @@ class Resnet50KerasBenchmarkRemoteData(Resnet50KerasBenchmarkBase):
self._override_flags_to_run_test_shorter() self._override_flags_to_run_test_shorter()
self._run_and_report_benchmark() self._run_and_report_benchmark()
def benchmark_graph_1_gpu_no_dist_strat(self):
"""Test Keras model in legacy graph mode with 1 GPU, no dist strat."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.enable_eager = False
FLAGS.distribution_strategy = 'off'
FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu_no_dist_strat')
FLAGS.batch_size = 96 # BatchNorm is less efficient in legacy graph mode
# due to its reliance on v1 cond.
self._override_flags_to_run_test_shorter()
self._run_and_report_benchmark()
def benchmark_1_gpu(self): def benchmark_1_gpu(self):
"""Test Keras model with 1 GPU.""" """Test Keras model with 1 GPU."""
self._setup() self._setup()
...@@ -1548,74 +1296,6 @@ class Resnet50KerasBenchmarkRemoteData(Resnet50KerasBenchmarkBase): ...@@ -1548,74 +1296,6 @@ class Resnet50KerasBenchmarkRemoteData(Resnet50KerasBenchmarkBase):
self._override_flags_to_run_test_shorter() self._override_flags_to_run_test_shorter()
self._run_and_report_benchmark() self._run_and_report_benchmark()
def benchmark_graph_1_gpu(self):
"""Test Keras model in legacy graph mode with 1 GPU."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.enable_eager = False
FLAGS.distribution_strategy = 'one_device'
FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu')
FLAGS.batch_size = 128
self._override_flags_to_run_test_shorter()
self._run_and_report_benchmark()
def benchmark_graph_xla_1_gpu(self):
"""Test Keras model in legacy graph mode with XLA and 1 GPU."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.enable_eager = False
FLAGS.enable_xla = True
FLAGS.distribution_strategy = 'one_device'
FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_1_gpu')
FLAGS.batch_size = 128
self._override_flags_to_run_test_shorter()
self._run_and_report_benchmark()
def benchmark_graph_1_gpu_fp16(self):
"""Test Keras model in legacy graph mode with 1 GPU and fp16."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.dtype = 'fp16'
FLAGS.enable_eager = False
FLAGS.distribution_strategy = 'one_device'
FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu_fp16')
FLAGS.batch_size = 256
self._override_flags_to_run_test_shorter()
self._run_and_report_benchmark()
def benchmark_graph_xla_1_gpu_fp16(self):
"""Test Keras model in legacy graph mode with 1 GPU, fp16 and XLA."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.dtype = 'fp16'
FLAGS.enable_eager = False
FLAGS.enable_xla = True
FLAGS.distribution_strategy = 'one_device'
FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_1_gpu_fp16')
FLAGS.batch_size = 256
self._override_flags_to_run_test_shorter()
self._run_and_report_benchmark()
def benchmark_graph_xla_1_gpu_fp16_tweaked(self):
"""Test Keras model in legacy graph with 1 GPU, fp16, XLA, and tuning."""
self._setup()
FLAGS.num_gpus = 1
FLAGS.enable_eager = False
FLAGS.enable_xla = True
FLAGS.distribution_strategy = 'one_device'
FLAGS.model_dir = self._get_model_dir(
'benchmark_graph_xla_1_gpu_fp16_tweaked')
FLAGS.dtype = 'fp16'
FLAGS.batch_size = 256
FLAGS.tf_gpu_thread_mode = 'gpu_private'
self._override_flags_to_run_test_shorter()
self._run_and_report_benchmark()
@benchmark_wrappers.enable_runtime_flags @benchmark_wrappers.enable_runtime_flags
def _run_and_report_benchmark(self): def _run_and_report_benchmark(self):
if FLAGS.num_gpus == 1 or FLAGS.run_eagerly: if FLAGS.num_gpus == 1 or FLAGS.run_eagerly:
......
official/benchmark/resnet_ctl_imagenet_benchmark.py
View file @ 0672bdcf
...@@ -144,8 +144,6 @@ class Resnet50CtlAccuracy(CtlBenchmark): ...@@ -144,8 +144,6 @@ class Resnet50CtlAccuracy(CtlBenchmark):
FLAGS.epochs_between_evals = 10 FLAGS.epochs_between_evals = 10
FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu') FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
FLAGS.dtype = 'fp32' FLAGS.dtype = 'fp32'
# Add some thread tunings to improve performance.
FLAGS.datasets_num_private_threads = 14
self._run_and_report_benchmark() self._run_and_report_benchmark()
def benchmark_8_gpu_fp16(self): def benchmark_8_gpu_fp16(self):
...@@ -158,8 +156,6 @@ class Resnet50CtlAccuracy(CtlBenchmark): ...@@ -158,8 +156,6 @@ class Resnet50CtlAccuracy(CtlBenchmark):
FLAGS.epochs_between_evals = 10 FLAGS.epochs_between_evals = 10
FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16') FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16')
FLAGS.dtype = 'fp16' FLAGS.dtype = 'fp16'
# Add some thread tunings to improve performance.
FLAGS.datasets_num_private_threads = 14
self._run_and_report_benchmark() self._run_and_report_benchmark()
def benchmark_8_gpu_amp(self): def benchmark_8_gpu_amp(self):
...@@ -173,8 +169,6 @@ class Resnet50CtlAccuracy(CtlBenchmark): ...@@ -173,8 +169,6 @@ class Resnet50CtlAccuracy(CtlBenchmark):
FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_amp') FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_amp')
FLAGS.dtype = 'fp16' FLAGS.dtype = 'fp16'
FLAGS.fp16_implementation = 'graph_rewrite' FLAGS.fp16_implementation = 'graph_rewrite'
# Add some thread tunings to improve performance.
FLAGS.datasets_num_private_threads = 14
self._run_and_report_benchmark() self._run_and_report_benchmark()
@benchmark_wrappers.enable_runtime_flags @benchmark_wrappers.enable_runtime_flags
......
official/modeling/hyperparams/base_config.py
View file @ 0672bdcf
...@@ -259,7 +259,6 @@ class RuntimeConfig(Config): ...@@ -259,7 +259,6 @@ class RuntimeConfig(Config):
distribution_strategy: e.g. 'mirrored', 'tpu', etc. distribution_strategy: e.g. 'mirrored', 'tpu', etc.
enable_xla: Whether or not to enable XLA. enable_xla: Whether or not to enable XLA.
per_gpu_thread_count: thread count per GPU. per_gpu_thread_count: thread count per GPU.
gpu_threads_enabled: Whether or not GPU threads are enabled.
gpu_thread_mode: Whether and how the GPU device uses its own threadpool. gpu_thread_mode: Whether and how the GPU device uses its own threadpool.
dataset_num_private_threads: Number of threads for a private threadpool dataset_num_private_threads: Number of threads for a private threadpool
created for all datasets computation. created for all datasets computation.
...@@ -278,7 +277,6 @@ class RuntimeConfig(Config): ...@@ -278,7 +277,6 @@ class RuntimeConfig(Config):
""" """
distribution_strategy: str = 'mirrored' distribution_strategy: str = 'mirrored'
enable_xla: bool = False enable_xla: bool = False
gpu_threads_enabled: bool = False
gpu_thread_mode: Optional[str] = None gpu_thread_mode: Optional[str] = None
dataset_num_private_threads: Optional[int] = None dataset_num_private_threads: Optional[int] = None
per_gpu_thread_count: int = 0 per_gpu_thread_count: int = 0
......
official/vision/image_classification/classifier_trainer.py
View file @ 0672bdcf
...@@ -234,7 +234,7 @@ def initialize(params: base_configs.ExperimentConfig, ...@@ -234,7 +234,7 @@ def initialize(params: base_configs.ExperimentConfig,
"""Initializes backend related initializations.""" """Initializes backend related initializations."""
keras_utils.set_session_config( keras_utils.set_session_config(
enable_xla=params.runtime.enable_xla) enable_xla=params.runtime.enable_xla)
if params.runtime.gpu_threads_enabled: if params.runtime.gpu_thread_mode:
keras_utils.set_gpu_thread_mode_and_count( keras_utils.set_gpu_thread_mode_and_count(
per_gpu_thread_count=params.runtime.per_gpu_thread_count, per_gpu_thread_count=params.runtime.per_gpu_thread_count,
gpu_thread_mode=params.runtime.gpu_thread_mode, gpu_thread_mode=params.runtime.gpu_thread_mode,
......
official/vision/image_classification/classifier_trainer_test.py
View file @ 0672bdcf
...@@ -320,7 +320,6 @@ class UtilTests(parameterized.TestCase, tf.test.TestCase): ...@@ -320,7 +320,6 @@ class UtilTests(parameterized.TestCase, tf.test.TestCase):
runtime=base_configs.RuntimeConfig( runtime=base_configs.RuntimeConfig(
run_eagerly=False, run_eagerly=False,
enable_xla=False, enable_xla=False,
gpu_threads_enabled=True,
per_gpu_thread_count=1, per_gpu_thread_count=1,
gpu_thread_mode='gpu_private', gpu_thread_mode='gpu_private',
num_gpus=1, num_gpus=1,
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment