def Params(cls):
"""Params for a MLPerfProgramSchedule."""
p = hyperparams.InstantiableParams(cls)
p.Define('task_dict', None, 'dataset_name -> task params')
p.Define('task_name', None, 'High level task name')
p.Define('logdir', None, 'Log directory')
p.Define('train_program', None, 'Train program params')
p.Define('train_executions_per_eval', 1, '')
p.Define('dataset_names', [], 'List of all dataset names.')
p.Define('num_splits_per_client', None, '')
p.Define('ml_perf', hyperparams.Params(), 'MlPerf configuration.')
mlp = p.ml_perf
mlp.Define('benchmark_name', None, 'Benchmark name for compliance log.')
mlp.Define('decoder_metric_name', None,
'Name of the decoder metric to report for compliance log.')
mlp.Define('decoder_metric_success_threshold', None,
'Benchmark run must exceed this value to succeeed.')
mlp.Define('steps_per_epoch', None, 'Number of training steps per epoch.')
mlp.Define('global_batch_size', None, 'Global batch size.')
mlp.Define('max_sequence_length', None, 'Maximum sequence length.')
mlp.Define('optimizer_name', None, 'Optimizer used.')
mlp.Define('opt_adam_beta_1', None, 'beta_1 used by Adam optimizer.')
mlp.Define('opt_adam_beta_2', None, 'beta_2 used by Adam optimizer.')
mlp.Define('opt_adam_epsilon', None, 'epsilon used by Adam optimizer.')
mlp.Define('base_learning_rate', None, 'Base learning rate.')
mlp.Define('warmup_steps', None, 'Number of warm-up steps.')
mlp.Define('train_samples', None, 'Number of train samples.')
mlp.Define('eval_samples', None, 'Number of eval samples.')
return p
def _JobSpec(cls, replicas):
"""Construct a job spec param with the given number of replicas."""
p = hyperparams.Params()
# By default, we use /job:localhost so that most of tests can just
# work out of the box. trainer.py will then set job names accordingly.
p.Define('name', '/job:localhost',
'TensorFlow job spec, e.g., /job:trainer, /job:ps')
p.Define('replicas', replicas, 'The number of tasks of a job.')
p.Define(
'targets', '', 'The target network address(es) to which we can '
'create tf sessions. E.g., a single ip:port, or a list of '
'comma-separated grpc://ip:port, etc.')
p.Define('cpus_per_replica', 1, 'The number of CPU devices to use per '
'replica.')
p.Define('gpus_per_replica', 0, 'The number of GPU devices to use per '
'replica.')
p.Define(
'devices_per_split', 1, 'Devices of a replica are grouped into '
'splits. Each split contains these many devices. One split is a '
'group of devices on which the computation nodes of a graph is '
'placed upon.E.g., one can place the forward lstm on device 0 of '
'a split and place the backward lstm on device 1. etc.')
p.Define('tpus_per_replica', 0,
'The number of tpu cores to use per replica.')
p.Define('num_tpu_hosts', 0, 'The number of tpu hosts.')
return p
def Params(cls):
p = hyperparams.Params()
p.Define('name', 'EarlyStop', '')
p.Define('metric_history', MetricHistory.Params(),
'Metric history params.')
p.Define(
'tolerance', 0.0, 'Minimum significant difference in metric; '
'useful if progress is asymptotic.')
p.Define('window', 0,
'Maximum number of steps between best and current.')
p.Define('verbose', True, 'Log early-stop checks.')
p.Define('min_steps', 0, 'Minimum number of steps before stopping.')
return p
def Params(cls):
"""Default parameters for a trial."""
p = hyperparams.Params()
p.Define(
'report_interval_seconds', 600,
'Interval between reporting trial results and checking for early '
'stopping.')
p.Define(
'vizier_objective_metric_key', 'loss',
'Which eval metric to use as the "objective value" for tuning.')
p.Define(
'report_during_training', False,
'Whether to report objective metrics during the training process.')
return p
def Params(cls):
"""Params for a SimpleProgramSchedule."""
p = hyperparams.InstantiableParams(cls)
p.Define('task_dict', None, 'dataset_name -> task params')
p.Define('task_name', None, 'High level task name')
p.Define('logdir', None, 'Log directory')
p.Define('train_program', None, 'Train program params')
p.Define('train_executions_per_eval', 1, '')
p.Define('eval_programs', [], 'List of eval program params.')
p.Define('num_splits_per_client', None, '')
p.Define('dataset_names', [], 'List of all dataset names.')
# TODO(blee): Clean these up.
p.Define('ml_perf', hyperparams.Params(), 'MlPerf configuration.')
mlp = p.ml_perf
mlp.Define('benchmark_name', None, 'Benchmark name for compliance log.')
return p
def BuildData():
"""Returns a hyperparam recording build information of this py binary."""
p = hyperparams.Params()
p.Define('timestamp', build_data.TimestampAscii(),
'Build timestamp as a string.')
p.Define('info', build_data.BuildInfo(),
'User, host, and directory of builder.')
p.Define('target', build_data.Target(), 'Build target.')
p.Define('id', build_data.BuildID(), 'Build id.')
p.Define('changelist', build_data.Changelist(), 'Build CL.')
p.Define('client_info', build_data.ClientInfo(),
'Perforce client changelist and status as descriptive string.')
p.Define('label', build_data.BuildLabel(),
'Build label (passed to make-{opt,dbg} -l).')
p.Define('platform', build_data.Platform(), 'Google platform.')
p.Define('tool', build_data.BuildTool(), 'Build tool.')
p.Define('paropts', build_data.ParOptions(), 'Par options.')
return p
def Params(cls):
p = hyperparams.Params()
p.Define('name', 'MetricHistory',
'Used by SetLogdirInMetricHistories.')
p.Define('jobname', 'eval_dev',
'Job and dataset to which metric applies.')
p.Define('metric', 'log_pplx', 'Metric to record.')
p.Define(
'minimize', True,
'If True, training minimizes the metric. If False, training '
'maximizes the metric.')
p.Define('logdir', '', 'Root dir for BF logs.')
p.Define(
'tfevent_file', False, 'If True, read the metric from '
'events.out.tfevents.* files in the job dir instead of '
'maintaining a history file.')
p.Define('local_filesystem', False,
'Logdir is on local filesystem (needed for unit test).')
return p
def Params(cls):
"""Defaults params for input generators."""
p = super(BaseInputGenerator, cls).Params()
p.name = 'input'
p.Define(
'batch_size', 0, 'Batch size for a device split. This will be '
'scaled to match the accelarator hardware topology.')
p.Define(
'num_samples', 0,
'If non-zero, the dataset contains these many samples. '
'For test/eval dataset, if we want the test/evel job evaluate '
'the whole dataset, this param must be set precisely. Otherwise, '
'this param is optional.')
# TPU related infeed tuning.
p.Define('use_per_host_infeed', False,
'Whether run infeed op on each host.')
p.Define(
'tpu_infeed_parallelism', 1,
'Uses these many python threads to drive infeed concurrently.')
p.Define('use_partitioned_infeed_queue', False,
'Use partitioned infeed')
p.Define('num_partitions', None, 'Num partitions')
p.Define('remote', hyperparams.Params(),
'Params to configure remote input policy.')
pp = p.remote
pp.Define(
'shardable_batch', True,
'True if and only if this input generates simple batches whose 1st '
'dimension of every tensor in a batch is the batch dimension, and '
'other dimensions are always the same.')
pp.Define(
'max_inflights_per_target', 32, 'The maximum number of '
'concurrent inflight remote input fetches per remote target.')
return p
def Test(self):
"""Returns Params for the testing dataset."""
return hyperparams.Params()
def Dev(self):
"""Returns Params for the development dataset."""
return hyperparams.Params()
def Train(self):
"""Returns Params for the training dataset."""
return hyperparams.Params()
def Params(cls):
p = super(QuantizableLayer, cls).Params()
p.Define('qdomain', hyperparams.Params(),
'Container for quantization domains.')
p.qdomain.Define('default', None, 'Default quantization domain.')
return p