def _run_and_report_benchmark(self):
start_time_sec = time.time()
stats = keras_imagenet_main.run(FLAGS)
wall_time_sec = time.time() - start_time_sec
super(TrivialKerasBenchmarkReal, self)._report_benchmark(
stats,
wall_time_sec,
total_batch_size=FLAGS.batch_size,
log_steps=FLAGS.log_steps)
def _run_and_report_benchmark(self):
start_time_sec = time.time()
stats = keras_imagenet_main.run(FLAGS)
wall_time_sec = time.time() - start_time_sec
super(TrivialKerasBenchmarkReal,
self)._report_benchmark(stats,
wall_time_sec,
total_batch_size=FLAGS.batch_size,
log_steps=FLAGS.log_steps)
def benchmark_graph_8_gpu(self):
"""Test Keras model with Keras fit/dist_strat and 8 GPUs."""
self._setup()
FLAGS.num_gpus = 8
FLAGS.data_dir = DATA_DIR
FLAGS.batch_size = 128 * 8
FLAGS.train_epochs = 90
FLAGS.model_dir = self._get_model_dir('keras_resnet50_8_gpu')
FLAGS.dtype = 'fp32'
stats = keras_imagenet_main.run(FLAGS)
self._fill_report_object(stats, FLAGS.batch_size)
def keras_resnet50_8_gpu(self):
"""Test Keras model with Keras fit/dist_strat and 8 GPUs."""
self._setup()
flags.FLAGS.num_gpus = 8
flags.FLAGS.data_dir = DATA_DIR
flags.FLAGS.batch_size = 64 * 8
flags.FLAGS.train_epochs = 90
flags.FLAGS.model_dir = self._get_model_dir('keras_resnet50_8_gpu')
flags.FLAGS.dtype = 'fp32'
stats = keras_imagenet_main.run(flags.FLAGS)
self._fill_report_object(stats)
def benchmark_graph_8_gpu(self):
"""Test Keras model with Keras fit/dist_strat and 8 GPUs."""
self._setup()
FLAGS.num_gpus = 8
FLAGS.data_dir = DATA_DIR
FLAGS.batch_size = 128*8
FLAGS.train_epochs = 90
FLAGS.model_dir = self._get_model_dir('keras_resnet50_8_gpu')
FLAGS.dtype = 'fp32'
stats = keras_imagenet_main.run(FLAGS)
self._fill_report_object(stats, FLAGS.batch_size)
def _run_and_report_benchmark(self):
start_time_sec = time.time()
stats = keras_imagenet_main.run(flags.FLAGS)
wall_time_sec = time.time() - start_time_sec
super(Resnet50KerasAccuracy, self)._report_benchmark(
stats,
wall_time_sec,
top_1_min=MIN_TOP_1_ACCURACY,
top_1_max=MAX_TOP_1_ACCURACY,
total_batch_size=FLAGS.batch_size,
log_steps=100)
def _run_and_report_benchmark(self):
start_time_sec = time.time()
stats = keras_imagenet_main.run(flags.FLAGS)
wall_time_sec = time.time() - start_time_sec
super(Resnet50KerasAccuracy, self)._report_benchmark(
stats,
wall_time_sec,
top_1_min=MIN_TOP_1_ACCURACY,
top_1_max=MAX_TOP_1_ACCURACY,
total_batch_size=FLAGS.batch_size,
log_steps=100)
def _run_and_report_benchmark(self):
start_time_sec = time.time()
stats = keras_imagenet_main.run(FLAGS)
wall_time_sec = time.time() - start_time_sec
# Number of logged step time entries that are excluded in performance
# report. We keep results from last 100 batches in this case.
warmup = (FLAGS.train_steps - 100) // FLAGS.log_steps
super(Resnet50KerasBenchmarkBase,
self)._report_benchmark(stats,
wall_time_sec,
total_batch_size=FLAGS.batch_size,
log_steps=FLAGS.log_steps,
warmup=warmup)
def _run_benchmark(self):
stats = keras_imagenet_main.run(FLAGS)
self.fill_report_object(stats)
def _run_benchmark(self): stats = keras_imagenet_main.run(FLAGS) self.fill_report_object(stats)
