def hvd_rotation_profiler_config_parser(config_folder, monkeypatch):
config_path = os.path.join(config_folder,
"hvd_rotation_profiler_config_parser.json")
monkeypatch.setenv("SMPROFILER_CONFIG_PATH", config_path)
return ProfilerConfigParser()
def native_tf2_pyinstrument_profiler_config_parser(config_folder, monkeypatch):
config_path = os.path.join(
config_folder,
"test_native_tf2_pyinstrument_profiler_config_parser.json")
monkeypatch.setenv("SMPROFILER_CONFIG_PATH", config_path)
return ProfilerConfigParser(FRAMEWORK.TENSORFLOW)
def case_insensitive_profiler_config_parser(config_folder, monkeypatch):
config_path = os.path.join(config_folder,
"case_insensitive_profiler_config_parser.json")
monkeypatch.setenv("SMPROFILER_CONFIG_PATH", config_path)
return ProfilerConfigParser()
def user_disabled_profiler_config_parser(config_folder, monkeypatch):
config_path = os.path.join(config_folder,
"user_disabled_profile_config_parser.json")
monkeypatch.setenv("SMPROFILER_CONFIG_PATH", config_path)
return ProfilerConfigParser()
def invalid_string_data_profiler_config_parser(config_folder, monkeypatch):
config_path = os.path.join(
config_folder, "invalid_string_data_profiler_config_parser.json")
monkeypatch.setenv("SMPROFILER_CONFIG_PATH", config_path)
return ProfilerConfigParser()
def missing_config_profiler_config_parser(config_folder, monkeypatch):
config_path = os.path.join(
config_folder, "missing_profile_config_parser.json") # doesn't exist
monkeypatch.setenv("SMPROFILER_CONFIG_PATH", config_path)
return ProfilerConfigParser()
def tf2_profiler_config_parser_by_time(config_folder, monkeypatch):
config_path = os.path.join(config_folder,
"test_tf2_profiler_config_parser_by_time.json")
monkeypatch.setenv("SMPROFILER_CONFIG_PATH", config_path)
return ProfilerConfigParser()
# smdistributed.dataparallel should be invoked via `mpirun`.
# It supports EC2 machines with 8 GPUs per machine.
smdataparallel = None
if check_smdataparallel_env():
try:
import smdistributed.dataparallel.torch.distributed as smdataparallel
except ImportError:
pass
DEFAULT_INCLUDE_COLLECTIONS = [CollectionKeys.LOSSES]
python_profiler = None
# Enable python profiling if profiling is enabled.
profiler_config_parser = ProfilerConfigParser()
if profiler_config_parser.profiling_enabled:
config = profiler_config_parser.config
if config.python_profiling_config.is_enabled():
python_profiler = PythonProfiler.get_python_profiler(config, "pytorch")
python_profiler.start_profiling(StepPhase.START)
class Hook(CallbackHook):
"""
The _TraceEventData is similar to a structure that contains the event data to be written in the event file.
It contains the following metadata:
training_phase such as "X", "M", "I" etc.
start time, duration and end time in microseconds
pid is the process id that wants to record this event.
and event arguments.
def __init__(
self,
collection_manager: CollectionManager,
default_include_collections: List[str],
init_step: int = 0,
out_dir: Optional[str] = None,
export_tensorboard: bool = False,
tensorboard_dir: Optional[str] = None,
dry_run: bool = False,
reduction_config: Optional[ReductionConfig] = None,
save_config: Optional[Union[SaveConfig, Dict[ModeKeys,
SaveConfigMode]]] = None,
include_regex: Optional[List[str]] = None,
include_collections: Optional[List[str]] = None,
save_all: bool = False,
include_workers: str = "one",
profiler_config_parser: Optional[ProfilerConfigParser] = None,
):
"""
A class used to represent the hook which gets attached to the
training process. This takes the form appropriate for the framework
such as tf.train.SessionRunHook for TF, Callback for keras...
...
Attributes
----------
out_dir : str
represents a path into which outputs will be written to. The hook raises error if the 'out_dir' already
exists. The implementation does not support merging the tensors generated in current job with tensors
from previous job. Hence, ensure that the 'out_dir' does not exist.
dry_run : bool
when dry run is set, behavior is only described in the log file.
tensors are not actually saved.
save_config: SaveConfig object
Takes save config object which is applied as default for all included tensors.
A collection can optionally have its own saveconfig object
which overrides this for its tensors.
reduction_config: ReductionConfig object
if passed, this reduction config object is used
as default for all tensors included.
A collection has its own saveconfig object
which overrides this for its tensors. if this is not passed,
tensor is saved in full.
include_regex: list of str
takes as input the list of string representing regular expressions. Tensors whose names match
these regular expressions will be saved. These tensors will be available as part of the `default`
collection.
include_collections: list of str representing collection names
takes as input the collections which should be saved.
if this is empty, it defaults to including all collections from code
save_all: bool
a shortcut for saving all tensors in the model.
they are all saved in the collection `all`
include_workers: str
makes the hook save data from all workers
profiler_config_parser: ProfilerConfigParser object
if passed, use this profiler configuration. by default, set up a new profiler configuration here.
"""
self.out_dir = verify_and_get_out_dir(out_dir)
self.tensorboard_dir = get_tensorboard_dir(
export_tensorboard=export_tensorboard,
tensorboard_dir=tensorboard_dir,
out_dir=self.out_dir,
)
self.dry_run = dry_run
self.worker = None
# when smdebug is used during an unsupported dist training process
# we write data only from the process that has self.first_process set to True.
self.first_process = None
self.save_all_workers = True if include_workers == "all" else False
self.chief_worker = DEFAULT_WORKER_NAME
if include_collections is None:
include_collections = default_include_collections
else:
include_collections = flatten(include_collections)
self.include_collections = list(
set(include_collections).union(set(default_include_collections)))
self.save_all = save_all
self.save_config = SaveConfig.parse(save_config)
if reduction_config is None:
reduction_config = ReductionConfig(save_raw_tensor=True)
self.reduction_config = reduction_config
self.include_regex = include_regex
self.collection_manager = collection_manager
self.init_step = init_step
# The written_tensor_name_for_step dictionary stores
# the names of each tensor saved for every step.
# This is to detect name clashes.
# If a name clash is detected, it is avoided by appending
# an index to the tensor name.
self.written_tensor_name_for_step = defaultdict(int)
self.logger = logger
if self.tensorboard_dir is None:
self.logger.info(
f"tensorboard_dir has not been set for the hook. SMDebug will not be exporting tensorboard summaries."
)
if include_regex is not None:
collection_manager.get(
CollectionKeys.DEFAULT).include(include_regex)
if CollectionKeys.DEFAULT not in self.include_collections:
self.include_collections.append(CollectionKeys.DEFAULT)
self.save_all = save_all
if self.save_all:
collection_manager.get(CollectionKeys.ALL).include(".*")
if CollectionKeys.ALL not in self.include_collections:
self.include_collections.append(CollectionKeys.ALL)
if (CollectionKeys.DEFAULT not in self.include_collections and
collection_manager.get(CollectionKeys.DEFAULT).include_regex):
self.logger.warn("The `default` collection was not passed to "
"include_collections. So it is not being saved")
self._collections_to_save = set()
self._collections_to_save_for_step = None
self.prepared_collections = False
self.tensor_to_collections = {}
self.step = init_step
self.last_saved_step = None
self.mode = ModeKeys.GLOBAL
self.mode_steps = {ModeKeys.GLOBAL: init_step}
self.writer = None
if profiler_config_parser is None:
profiler_config_parser = ProfilerConfigParser()
profiler_config_parser.load_config()
self.profiler_config_parser = profiler_config_parser
self.timeline_writer = TimelineFileWriter(
profiler_config_parser=profiler_config_parser)
self.hvd_reader = None
self.is_smdataparallel_profiling = False
if is_sagemaker_job() and SageMakerFileMetricsWriter is not None:
self.metrics_writer = SageMakerFileMetricsWriter()
else:
self.metrics_writer = None
# Maps ModeKeys to FileWriter objects
self.tb_writers = {}
# Cache scalars that are being saved through save_scalar() calls
self.scalar_cache = []
self.logger.info("Saving to {}".format(self.out_dir))
atexit.register(self._cleanup)
# Check if there is any last saved state. Initialize the hook based last saved state.
self.training_run = 0
self._initialize_to_last_saved_state()
self.custom_tensors_to_save = dict()
