def __new__(cls,
iterations_per_loop=2,
num_shards=2,
per_host_input_for_training=True):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
per_host_input_for_training=per_host_input_for_training)
def __new__(cls,
iterations_per_loop=2,
num_shards=2,
per_host_input_for_training=False):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
per_host_input_for_training=per_host_input_for_training)
def __new__(cls,
iterations_per_loop=2,
num_shards=2,
per_host_input_for_training=True,
tpu_job_name=None,
initial_infeed_sleep_secs=None):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
# Check initial_infeed_sleep_secs.
if initial_infeed_sleep_secs:
util_lib.check_positive_integer(initial_infeed_sleep_secs,
'TPUConfig initial_infeed_sleep_secs')
tpu_job_name = tpu_job_name or _get_tpu_job_name_from_tf_config()
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
per_host_input_for_training=per_host_input_for_training,
tpu_job_name=tpu_job_name,
initial_infeed_sleep_secs=initial_infeed_sleep_secs)
def __new__(cls,
iterations_per_loop=2,
num_shards=2,
per_host_input_for_training=True,
tpu_job_name=None,
initial_infeed_sleep_secs=None):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
# Check initial_infeed_sleep_secs.
if initial_infeed_sleep_secs:
util_lib.check_positive_integer(
initial_infeed_sleep_secs,
'TPUConfig initial_infeed_sleep_secs')
tpu_job_name = tpu_job_name or _get_tpu_job_name_from_tf_config()
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
per_host_input_for_training=per_host_input_for_training,
tpu_job_name=tpu_job_name,
initial_infeed_sleep_secs=initial_infeed_sleep_secs)
def __new__(cls,
iterations_per_loop=2,
num_shards=2,
computation_shape=None,
per_host_input_for_training=True,
tpu_job_name=None,
initial_infeed_sleep_secs=None):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
# Check computation_shape
if computation_shape is not None and len(computation_shape) != 3:
raise ValueError(
'computation_shape must be a list with length 3 or None; got {}'.
format(str(computation_shape)))
if computation_shape is not None:
computation_shape_array = np.asarray(computation_shape, dtype=np.int32)
# This prevents any computation being replicated across multiple hosts, so
# that each host feeds the same number of computations.
if any(computation_shape_array < 1) or any(computation_shape_array > 2):
raise ValueError('computation_shape elements can only be 1 or 2; got '
'computation_shape={}'.format(computation_shape))
max_replicas_per_host = (
_NUM_CORES_PER_HOST // np.prod(computation_shape_array))
if num_shards > max_replicas_per_host and (
num_shards % max_replicas_per_host != 0):
raise ValueError(
'{0} shards can not be evenly distributed across'
' multiple hosts. Each shard needs {1} cores and each'
' host has {2} cores. Thus {0} shards needs {3} hosts.'
' Please adjust num shards so that num_shards is'
' divisible by {4} or <= {4}.'.format(
num_shards, np.prod(computation_shape), _NUM_CORES_PER_HOST,
num_shards / max_replicas_per_host, max_replicas_per_host))
# Check initial_infeed_sleep_secs.
if initial_infeed_sleep_secs:
util_lib.check_positive_integer(initial_infeed_sleep_secs,
'TPUConfig initial_infeed_sleep_secs')
tpu_job_name = tpu_job_name or _get_tpu_job_name_from_tf_config()
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
computation_shape=computation_shape,
per_host_input_for_training=per_host_input_for_training,
tpu_job_name=tpu_job_name,
initial_infeed_sleep_secs=initial_infeed_sleep_secs)
def __new__(cls,
iterations_per_loop=2,
num_shards=None,
computation_shape=None,
per_host_input_for_training=True,
tpu_job_name=None,
initial_infeed_sleep_secs=None):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
if num_shards is not None:
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
# Check computation_shape
if computation_shape is not None and len(computation_shape) != 3:
raise ValueError(
'computation_shape must be a list with length 3 or None; got {}'
.format(str(computation_shape)))
if computation_shape is not None:
computation_shape_array = np.asarray(computation_shape,
dtype=np.int32)
# This prevents any computation being replicated across multiple hosts, so
# that each host feeds the same number of computations.
if any(computation_shape_array < 1) or any(
computation_shape_array > 2):
raise ValueError(
'computation_shape elements can only be 1 or 2; got '
'computation_shape={}'.format(computation_shape))
# per_host_input_for_training may be True, False, or integer in [1..3].
# Map legacy values (True, False) to numeric values.
if per_host_input_for_training is False:
per_host_input_for_training = InputPipelineConfig.PER_SHARD_V1
elif per_host_input_for_training is True:
per_host_input_for_training = InputPipelineConfig.PER_HOST_V1
# Check initial_infeed_sleep_secs.
if initial_infeed_sleep_secs:
util_lib.check_positive_integer(
initial_infeed_sleep_secs,
'TPUConfig initial_infeed_sleep_secs')
tpu_job_name = tpu_job_name or _get_tpu_job_name_from_tf_config()
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
computation_shape=computation_shape,
per_host_input_for_training=per_host_input_for_training,
tpu_job_name=tpu_job_name,
initial_infeed_sleep_secs=initial_infeed_sleep_secs)
def __new__(cls,
iterations_per_loop=2,
num_shards=None,
computation_shape=None,
per_host_input_for_training=True,
tpu_job_name=None,
initial_infeed_sleep_secs=None):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
if num_shards is not None:
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
# Check computation_shape
if computation_shape is not None and len(computation_shape) != 3:
raise ValueError(
'computation_shape must be a list with length 3 or None; got {}'.
format(str(computation_shape)))
if computation_shape is not None:
computation_shape_array = np.asarray(computation_shape, dtype=np.int32)
# This prevents any computation being replicated across multiple hosts, so
# that each host feeds the same number of computations.
if any(computation_shape_array < 1) or any(computation_shape_array > 2):
raise ValueError('computation_shape elements can only be 1 or 2; got '
'computation_shape={}'.format(computation_shape))
# per_host_input_for_training may be True, False, or integer in [1..3].
# Map legacy values (True, False) to numeric values.
if per_host_input_for_training is False:
per_host_input_for_training = InputPipelineConfig.PER_SHARD_V1
elif per_host_input_for_training is True:
per_host_input_for_training = InputPipelineConfig.PER_HOST_V1
# Check initial_infeed_sleep_secs.
if initial_infeed_sleep_secs:
util_lib.check_positive_integer(initial_infeed_sleep_secs,
'TPUConfig initial_infeed_sleep_secs')
tpu_job_name = tpu_job_name or _get_tpu_job_name_from_tf_config()
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
computation_shape=computation_shape,
per_host_input_for_training=per_host_input_for_training,
tpu_job_name=tpu_job_name,
initial_infeed_sleep_secs=initial_infeed_sleep_secs)
def __new__(cls,
iterations_per_loop=2,
num_shards=None,
num_cores_per_replica=None,
per_host_input_for_training=True,
tpu_job_name=None,
initial_infeed_sleep_secs=None):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
if num_shards is not None:
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
# Parse computation_shape
if num_cores_per_replica is not None:
if num_cores_per_replica not in [1, 2, 4, 8]:
raise ValueError(
'num_cores_per_replica must be 1, 2, 4, or 8; got {}'.
format(str(num_cores_per_replica)))
# per_host_input_for_training may be True, False, or integer in [1..3].
# Map legacy values (True, False) to numeric values.
if per_host_input_for_training is False:
per_host_input_for_training = InputPipelineConfig.PER_SHARD_V1
elif per_host_input_for_training is True:
per_host_input_for_training = InputPipelineConfig.PER_HOST_V1
# Check initial_infeed_sleep_secs.
if initial_infeed_sleep_secs:
util_lib.check_positive_integer(
initial_infeed_sleep_secs,
'TPUConfig initial_infeed_sleep_secs')
tpu_job_name = tpu_job_name or _get_tpu_job_name_from_tf_config()
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
num_cores_per_replica=num_cores_per_replica,
per_host_input_for_training=per_host_input_for_training,
tpu_job_name=tpu_job_name,
initial_infeed_sleep_secs=initial_infeed_sleep_secs)
def __new__(cls,
iterations_per_loop=2,
num_shards=None,
num_cores_per_replica=None,
per_host_input_for_training=True,
tpu_job_name=None,
initial_infeed_sleep_secs=None):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
if num_shards is not None:
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
# Parse computation_shape
if num_cores_per_replica is not None:
if num_cores_per_replica not in [1, 2, 4, 8]:
raise ValueError(
'num_cores_per_replica must be 1, 2, 4, or 8; got {}'.format(
str(num_cores_per_replica)))
# per_host_input_for_training may be True, False, or integer in [1..3].
# Map legacy values (True, False) to numeric values.
if per_host_input_for_training is False:
per_host_input_for_training = InputPipelineConfig.PER_SHARD_V1
elif per_host_input_for_training is True:
per_host_input_for_training = InputPipelineConfig.PER_HOST_V1
# Check initial_infeed_sleep_secs.
if initial_infeed_sleep_secs:
util_lib.check_positive_integer(initial_infeed_sleep_secs,
'TPUConfig initial_infeed_sleep_secs')
tpu_job_name = tpu_job_name or _get_tpu_job_name_from_tf_config()
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
num_cores_per_replica=num_cores_per_replica,
per_host_input_for_training=per_host_input_for_training,
tpu_job_name=tpu_job_name,
initial_infeed_sleep_secs=initial_infeed_sleep_secs)
def __new__(cls,
iterations_per_loop=2,
num_shards=None,
num_cores_per_replica=None,
per_host_input_for_training=True,
tpu_job_name=None,
initial_infeed_sleep_secs=None,
input_partition_dims=None):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
if num_shards is not None:
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
if input_partition_dims is not None:
if len(input_partition_dims) != 1 and len(
input_partition_dims) != 2:
raise ValueError(
'input_partition_dims must be a list/tuple with one or two'
' elements.')
if per_host_input_for_training is not InputPipelineConfig.PER_HOST_V2:
raise ValueError(
'input_partition_dims is only supported in PER_HOST_V2 mode.'
)
if num_cores_per_replica is None:
raise ValueError(
'input_partition_dims requires setting num_cores_per_replica.'
)
# Check num_cores_per_replica
if num_cores_per_replica is not None:
if num_cores_per_replica not in [1, 2, 4, 8]:
raise ValueError(
'num_cores_per_replica must be 1, 2, 4, or 8; got {}'.
format(str(num_cores_per_replica)))
# per_host_input_for_training may be True, False, or integer in [1..3].
# Map legacy values (True, False) to numeric values.
if per_host_input_for_training is False:
per_host_input_for_training = InputPipelineConfig.PER_SHARD_V1
elif per_host_input_for_training is True:
per_host_input_for_training = InputPipelineConfig.PER_HOST_V1
# Check initial_infeed_sleep_secs.
if initial_infeed_sleep_secs:
util_lib.check_positive_integer(
initial_infeed_sleep_secs,
'TPUConfig initial_infeed_sleep_secs')
tpu_job_name = tpu_job_name or _get_tpu_job_name_from_tf_config()
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
num_cores_per_replica=num_cores_per_replica,
per_host_input_for_training=per_host_input_for_training,
tpu_job_name=tpu_job_name,
initial_infeed_sleep_secs=initial_infeed_sleep_secs,
input_partition_dims=input_partition_dims)
def __new__(cls,
iterations_per_loop=2,
num_shards=2,
computation_shape=None,
per_host_input_for_training=True,
tpu_job_name=None,
initial_infeed_sleep_secs=None):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
# Check computation_shape
if computation_shape is not None and len(computation_shape) != 3:
raise ValueError(
'computation_shape must be a list with length 3 or None; got {}'
.format(str(computation_shape)))
if computation_shape is not None:
computation_shape_array = np.asarray(computation_shape,
dtype=np.int32)
# This prevents any computation being replicated across multiple hosts, so
# that each host feeds the same number of computations.
if any(computation_shape_array < 1) or any(
computation_shape_array > 2):
raise ValueError(
'computation_shape elements can only be 1 or 2; got '
'computation_shape={}'.format(computation_shape))
max_replicas_per_host = (_NUM_CORES_PER_HOST //
np.prod(computation_shape_array))
if num_shards > max_replicas_per_host and (
num_shards % max_replicas_per_host != 0):
raise ValueError(
'{0} shards can not be evenly distributed across'
' multiple hosts. Each shard needs {1} cores and each'
' host has {2} cores. Thus {0} shards needs {3} hosts.'
' Please adjust num shards so that num_shards is'
' divisible by {4} or <= {4}.'.format(
num_shards, np.prod(computation_shape),
_NUM_CORES_PER_HOST,
num_shards / max_replicas_per_host,
max_replicas_per_host))
# Check initial_infeed_sleep_secs.
if initial_infeed_sleep_secs:
util_lib.check_positive_integer(
initial_infeed_sleep_secs,
'TPUConfig initial_infeed_sleep_secs')
tpu_job_name = tpu_job_name or _get_tpu_job_name_from_tf_config()
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
computation_shape=computation_shape,
per_host_input_for_training=per_host_input_for_training,
tpu_job_name=tpu_job_name,
initial_infeed_sleep_secs=initial_infeed_sleep_secs)
def __new__(cls,
iterations_per_loop=2,
num_shards=None,
num_cores_per_replica=None,
per_host_input_for_training=True,
tpu_job_name=None,
initial_infeed_sleep_secs=None,
input_partition_dims=None):
# Check iterations_per_loop.
util_lib.check_positive_integer(iterations_per_loop,
'TPUConfig iterations_per_loop')
# Check num_shards.
if num_shards is not None:
util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards')
if input_partition_dims is not None:
if len(input_partition_dims) != 1 and len(input_partition_dims) != 2:
raise ValueError(
'input_partition_dims must be a list/tuple with one or two'
' elements.')
if per_host_input_for_training is not InputPipelineConfig.PER_HOST_V2:
raise ValueError(
'input_partition_dims is only supported in PER_HOST_V2 mode.')
if num_cores_per_replica is None:
raise ValueError(
'input_partition_dims requires setting num_cores_per_replica.')
# Check num_cores_per_replica
if num_cores_per_replica is not None:
if num_cores_per_replica not in [1, 2, 4, 8]:
raise ValueError(
'num_cores_per_replica must be 1, 2, 4, or 8; got {}'.format(
str(num_cores_per_replica)))
# per_host_input_for_training may be True, False, or integer in [1..3].
# Map legacy values (True, False) to numeric values.
if per_host_input_for_training is False:
per_host_input_for_training = InputPipelineConfig.PER_SHARD_V1
elif per_host_input_for_training is True:
per_host_input_for_training = InputPipelineConfig.PER_HOST_V1
# Check initial_infeed_sleep_secs.
if initial_infeed_sleep_secs:
util_lib.check_positive_integer(initial_infeed_sleep_secs,
'TPUConfig initial_infeed_sleep_secs')
tpu_job_name = tpu_job_name or _get_tpu_job_name_from_tf_config()
return super(TPUConfig, cls).__new__(
cls,
iterations_per_loop=iterations_per_loop,
num_shards=num_shards,
num_cores_per_replica=num_cores_per_replica,
per_host_input_for_training=per_host_input_for_training,
tpu_job_name=tpu_job_name,
initial_infeed_sleep_secs=initial_infeed_sleep_secs,
input_partition_dims=input_partition_dims)
