def ProgramSchedule(self):
p = program_lib.MLPerfProgramScheduleForTask(
train_dataset_name='Train',
train_steps_per_loop=20,
decode_dataset_name='Test',
decode_steps_per_loop=1,
num_epochs_per_session_run=4,
warmup_seconds=0)
p.train_executions_per_eval = 1
p.ml_perf.benchmark_name = 'transformer'
p.ml_perf.steps_per_epoch = 1
p.ml_perf.decoder_metric_name = 'ml_perf_bleu'
# Dummy value just to see run_stop/success.
p.ml_perf.decoder_metric_success_threshold = -10.0
p.ml_perf.max_sequence_length = 80
p.ml_perf.global_batch_size = 64
p.ml_perf.optimizer_name = 'adam'
p.ml_perf.opt_adam_beta_1 = 0.9
p.ml_perf.opt_adam_beta_2 = 0.98
p.ml_perf.opt_adam_epsilon = 1e-9
p.ml_perf.base_learning_rate = 2.0
p.ml_perf.warmup_steps = self.WARMUP_STEPS
p.ml_perf.train_samples = 566340
p.ml_perf.eval_samples = 3003
return p
def ProgramSchedule(self):
p = program_lib.MLPerfProgramScheduleForTask(
train_dataset_name='Train',
train_steps_per_loop=1107,
decode_dataset_name='Test',
decode_steps_per_loop=1,
num_epochs_per_session_run=4)
p.train_executions_per_eval = 1
# For compliance logging.
p.ml_perf.benchmark_name = 'transformer'
p.ml_perf.steps_per_epoch = self.STEPS_PER_EPOCH
p.ml_perf.decoder_metric_name = 'ml_perf_bleu'
p.ml_perf.decoder_metric_success_threshold = 0.25
p.ml_perf.max_sequence_length = 80
p.ml_perf.global_batch_size = 512
p.ml_perf.optimizer_name = 'adam'
p.ml_perf.opt_adam_beta_1 = 0.9
p.ml_perf.opt_adam_beta_2 = 0.98
p.ml_perf.opt_adam_epsilon = 1e-9
p.ml_perf.base_learning_rate = self.LEARNING_RATE
p.ml_perf.warmup_steps = self.WARMUP_STEPS
p.ml_perf.train_samples = 566340
p.ml_perf.eval_samples = 3003
return p
