def ProgramSchedule(self):
return program.SimpleProgramScheduleForTask(
train_dataset_name='Train',
train_steps_per_loop=50,
eval_dataset_names=['Test'],
eval_steps_per_loop=5,
decode_steps_per_loop=0)
def ProgramSchedule(self):
"""Returns a schedule for the Executor."""
return program_lib.SimpleProgramScheduleForTask(
'Train',
train_steps_per_loop=self.Task().train.tpu_steps_per_loop,
eval_dataset_names=[],
eval_steps_per_loop=0,
decode_steps_per_loop=0)
def ProgramSchedule(self):
# Only needed if --use_tpu_executor.
p = program.SimpleProgramScheduleForTask(train_dataset_name='Train',
train_steps_per_loop=1000,
eval_dataset_names=[],
eval_steps_per_loop=0,
decode_steps_per_loop=0)
p.train_executions_per_eval = 0
return p
def ProgramSchedule(self):
return program.SimpleProgramScheduleForTask(
train_dataset_name='Train',
train_steps_per_loop=100,
# I want to compute WER...
eval_dataset_names=['Dev'],
eval_steps_per_loop=1,
decode_steps_per_loop=1,
)
def ProgramSchedule(self):
p = program.SimpleProgramScheduleForTask(
train_dataset_name='Train',
train_steps_per_loop=1000,
eval_dataset_names=['Dev', 'Test'],
eval_steps_per_loop=1,
decode_steps_per_loop=1)
p.train_executions_per_eval = 0
return p
def ProgramSchedule(self):
p = program.SimpleProgramScheduleForTask(
train_dataset_name='Train',
train_steps_per_loop=train_steps_per_loop,
eval_dataset_names=['Test'],
eval_steps_per_loop=eval_decode_steps_per_loop,
decode_steps_per_loop=eval_decode_steps_per_loop)
if max_train_steps == 0:
p.train_executions_per_eval = 0
return p
def ProgramSchedule(self):
p = program.SimpleProgramScheduleForTask(
train_dataset_name='Train',
train_steps_per_loop=100,
eval_dataset_names=['Train'],
eval_steps_per_loop=100,
decode_steps_per_loop=0,
)
p.train_program.spmd = True
# every 5K steps
p.train_executions_per_eval = 5
return p
def ProgramSchedule(self):
p = program.SimpleProgramScheduleForTask(
train_dataset_name='Train',
train_steps_per_loop=self.TRAIN_STEPS_PER_LOOP,
eval_dataset_names=['Test'],
eval_steps_per_loop=10,
decode_steps_per_loop=0,
)
p.train_program.spmd = True
p.train_executions_per_eval = self.TRAIN_EXES_PER_EVAL
# For compliance logging.
p.ml_perf.benchmark_name = 'bert'
p.ml_perf.submission_metadata = {
'global_batch_size': self.BATCH_SIZE,
'submission_org': 'Google',
'submission_platform': 'tpu-v4-4096',
'submission_division': 'open',
'submission_status': 'cloud',
'submission_benchmark': p.ml_perf.benchmark_name,
'submission_model': 'lingvo',
'cache_clear': None,
'train_samples': 156725653,
'eval_samples': 10000
}
# For BERT, we log the number of examples as the epoch.
# epoch_num = global_step / steps_per_epoch
# epoch_num = num_examples_trained = global_step * examples_per_step
# steps_per_epoch = global_step / (global_step * examples_per_step)
# steps_per_epoch = 1 / examples_per_step
examples_per_step = self.BATCH_SIZE
p.ml_perf.steps_per_epoch = 1 / examples_per_step
p.ml_perf.decoder_metric_name = 'acc1'
p.ml_perf.decoder_metric_success_threshold = 0.6
p.ml_perf.max_steps_to_train = 31790
return p
