def get_temp_path(file_path):
directory = os.path.dirname(file_path)
if is_sagemaker_job():
temp_path = file_path + SAGEMAKER_TEMP_PATH_SUFFIX
else:
if len(file_path) > 0 and file_path[0] == "/":
file_path = file_path[1:]
temp_path = os.path.join(NON_SAGEMAKER_TEMP_PATH_PREFIX, file_path)
return temp_path
def check_metrics_file(save_steps, saved_scalars=None):
"""
Check the SageMaker metrics file to ensure that all the scalars saved using
save_scalar(sm_metrics=True) or mentioned through SM_METRICS collections, have been saved.
"""
# need this to seek to the right file offset for test output verification
metrics_file_position = 0
if is_sagemaker_job():
METRICS_DIR = os.environ.get(DEFAULT_SAGEMAKER_METRICS_PATH)
if not METRICS_DIR:
logging.warning("SageMaker Metric Directory not specified")
return
file_name = "{}/{}.json".format(METRICS_DIR, str(os.getpid()))
scalarnames = set()
import collections
train_metric = collections.defaultdict(list)
eval_metric = collections.defaultdict(list)
with open(file_name) as fp:
# since SM metrics expects all metrics to be written in 1 file, seeking to
# the right offset for the purpose of this test - so that the metrics logged in
# the corresponding test are verified
fp.seek(metrics_file_position)
for line in fp:
data = json.loads(line)
assert data[
"IterationNumber"] != -1 # iteration number should not be -1
metric_name = data["MetricName"]
if "TRAIN" in metric_name:
train_metric[metric_name].append(data["IterationNumber"])
scalarnames.add(metric_name.rstrip("_TRAIN"))
elif "EVAL" in metric_name:
eval_metric[metric_name].append(data["IterationNumber"])
scalarnames.add(metric_name.rstrip("_EVAL"))
else:
scalarnames.add(
metric_name.rstrip("_GLOBAL")
) # check the scalar saved using save_scalar()
metrics_file_position = fp.tell()
assert scalarnames
if saved_scalars:
assert len(set(saved_scalars) & set(scalarnames)) > 0
# check if all metrics have been written at the expected step number
for train_data in train_metric:
assert len(
set(save_steps["TRAIN"])
& set(train_metric[train_data])) == len(save_steps["TRAIN"])
for eval_data in eval_metric:
assert len(set(save_steps["EVAL"])
& set(eval_metric[eval_data])) == len(
save_steps["EVAL"])
def create_hook_from_json_config(hook_cls,
json_config_path,
default_values=None):
"""Returns a SessionHook object corresponding to either TF, PT, or MXNet.
If json_config_path is None, an environment variable must be set.
Here we compare HookParameters with CollectionConfiguration and set all the defaults.
"""
params_dict = get_json_config_as_dict(json_config_path=json_config_path)
hook_params = collect_hook_config_params(params_dict)
out_dir = hook_params.get("out_dir")
dry_run = hook_params.get("dry_run", False)
reduction_config = hook_params.get(CONFIG_RDN_CFG_KEY, None)
save_config = SaveConfig.from_dict(hook_params.get("save_config_modes"),
default_values)
include_regex = hook_params.get(CONFIG_INCLUDE_REGEX_KEY)
include_collections = get_include_collections(params_dict)
save_all = hook_params.get(CONFIG_SAVE_ALL_KEY, False)
include_workers = hook_params.get(CONFIG_INCLUDE_WORKERS_KEY, "one")
# If Sagemaker, emit TB only if JSON file exists
if is_sagemaker_job():
tensorboard_dir = get_tensorboard_dir_from_json_config()
export_tensorboard = bool(tensorboard_dir is not None)
# Otherwise, place TB artifacts in out_dir
else:
tensorboard_dir = hook_params[TENSORBOARD_DIR_KEY]
export_tensorboard = hook_params[EXPORT_TENSORBOARD_KEY]
hook = hook_cls(
out_dir=out_dir,
export_tensorboard=export_tensorboard,
tensorboard_dir=tensorboard_dir,
dry_run=dry_run,
reduction_config=reduction_config,
save_config=save_config,
include_regex=include_regex,
include_collections=include_collections,
include_workers=include_workers,
save_all=save_all,
)
add_collections_to_manager(hook.collection_manager, params_dict,
hook_params)
return hook
def check_metrics_file(saved_scalars):
"""
Check the SageMaker metrics file to ensure that all the scalars saved using
save_scalar(sm_metrics=True) or mentioned through SM_METRICS collections, have been saved.
"""
if is_sagemaker_job():
METRICS_DIR = os.environ.get(DEFAULT_SAGEMAKER_METRICS_PATH)
if not METRICS_DIR:
logging.warning("SageMaker Metric Directory not specified")
return
file_name = "{}/{}.json".format(METRICS_DIR, str(os.getpid()))
scalarnames = set()
with open(file_name) as fp:
for line in fp:
data = json.loads(line)
scalarnames.add(data["MetricName"])
assert scalarnames
assert len(set(saved_scalars) & set(scalarnames)) > 0
def training_has_ended(trial_prefix):
# Emit the end of training file only if the job is not running under SageMaker.
if is_sagemaker_job():
logger.info(
f"The end of training job file will not be written for jobs running under SageMaker."
)
return
try:
check_dir_exists(trial_prefix)
# if path does not exist, then we don't need to write a file
except RuntimeError:
# dir exists
pass
file_path = os.path.join(trial_prefix, END_OF_JOB_FILENAME)
s3, bucket_name, key_name = is_s3(file_path)
if s3:
writer = TSAccessS3(bucket_name, key_name, binary=False)
else:
writer = TSAccessFile(file_path, "a+")
writer.flush()
try:
writer.close()
except OSError:
"""
def __init__(
self,
collection_manager: CollectionManager,
default_include_collections: List[str],
profiler_config_parser: ProfilerConfigParser,
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",
):
"""
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.
"""
error_handling_agent.set_hook(
self) # This should be the first line in the constructor.
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
self.profiler_config_parser = profiler_config_parser
self.profiler_config_parser.load_config()
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()
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",
):
"""
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
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
"""
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
self.save_all_workers = True if include_workers == "all" else False
self.chief_worker = CONFIG_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.collection_manager.set_num_workers(self._get_num_workers())
self.init_step = init_step
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 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()
