def dummy_trial_creator(trial_dir, num_workers, job_ended):
Path(trial_dir).mkdir(parents=True, exist_ok=True)
cm = CollectionManager()
for i in range(num_workers):
collection_file_name = f"worker_{i}_collections.json"
cm.export(trial_dir, collection_file_name)
if job_ended:
Path(os.path.join(trial_dir, "training_job_end.ts")).touch()
def test_collection_defaults_to_hook_config():
"""Test that hook save_configs propagate to collection defaults.
For example, if we set ModeKeys.TRAIN: save_interval=10 in the hook
and ModeKeys.EVAL: save_interval=20 in a collection, we would like the collection to
be finalized as {ModeKeys.TRAIN: save_interval=10, ModeKeys.EVAL: save_interval=20}.
"""
cm = CollectionManager()
cm.create_collection("foo")
cm.get("foo").include_regex = "*"
cm.get("foo").save_config = {
ModeKeys.EVAL: SaveConfigMode(save_interval=20)
}
hook = Hook(
out_dir="/tmp/test_collections/" + str(datetime.datetime.now()),
save_config={ModeKeys.TRAIN: SaveConfigMode(save_interval=10)},
include_collections=["foo"],
reduction_config=ReductionConfig(save_raw_tensor=True),
)
hook.collection_manager = cm
assert cm.get("foo").save_config.mode_save_configs[ModeKeys.TRAIN] is None
assert cm.get("foo").reduction_config is None
hook._prepare_collections()
assert cm.get("foo").save_config.mode_save_configs[
ModeKeys.TRAIN].save_interval == 10
assert cm.get("foo").reduction_config.save_raw_tensor is True
def test_json_params():
params_dict = get_json_config_as_dict(
json_config_path="tests/core/json_configs/all_params.json")
hook_params = collect_hook_config_params(params_dict)
include_collections = get_include_collections(params_dict)
coll_manager = CollectionManager()
add_collections_to_manager(coll_manager, params_dict, hook_params)
assert hook_params["include_workers"] == "one"
assert hook_params["save_all"] is True
assert coll_manager.get("weights").save_histogram is False
assert coll_manager.get("gradients").save_histogram is False
assert "weights" in include_collections
assert "gradients" in include_collections
assert len(include_collections) == 2
assert hook_params["export_tensorboard"] == True
assert hook_params["tensorboard_dir"] == "/tmp/tensorboard"
def _read_collections(self, collection_files):
first_collection_file = collection_files[0] # First Collection File
key = os.path.join(first_collection_file)
collections_req = ReadObjectRequest(self._get_s3_location(key))
obj_data = self.s3_handler.get_objects([collections_req])[0]
obj_data = obj_data.decode("utf-8")
self.collection_manager = CollectionManager.load_from_string(obj_data)
self.num_workers = self.collection_manager.get_num_workers()
def help_test_multiple_trials(num_steps=20, num_tensors=10):
trial_name = str(uuid.uuid4())
bucket = "smdebug-testing"
path = "s3://" + os.path.join(bucket, "outputs/")
c = CollectionManager()
c.add("default")
c.get("default").tensor_names = [
"foo_" + str(i) for i in range(num_tensors)
]
c.export(path + trial_name, DEFAULT_COLLECTIONS_FILE_NAME)
c.export(path + trial_name, DEFAULT_COLLECTIONS_FILE_NAME)
for i in range(num_steps):
generate_data(
path=path,
trial=trial_name,
num_tensors=num_tensors,
step=i,
tname_prefix="foo",
worker="algo-1",
shape=(3, 3, 3),
rank=0,
)
_, bucket, prefix = is_s3(os.path.join(path, trial_name))
trial_obj = S3Trial(name=prefix, bucket_name=bucket, prefix_name=prefix)
return trial_obj, trial_name
def test_invalid_collection_config_exception():
cm = CollectionManager()
cm.create_collection("foo")
hook = Hook(
out_dir="/tmp/test_collections/" + str(datetime.datetime.now()),
save_config={ModeKeys.TRAIN: SaveConfigMode(save_interval=10)},
include_collections=["foo"],
reduction_config=ReductionConfig(save_raw_tensor=True),
)
hook.collection_manager = cm
try:
hook._prepare_collections()
except InvalidCollectionConfiguration:
pass
else:
assert False, "Invalid Collection Name did not raise error"
cm.get("foo").include_regex = "*"
try:
hook._prepare_collections()
except InvalidCollectionConfiguration:
assert False, "Valid Collection Name raised an error"
def generate_data(
path,
trial,
step,
tname_prefix,
num_tensors,
worker,
shape,
dtype=np.float32,
rank=None,
mode=None,
mode_step=None,
export_colls=True,
data=None,
):
with FileWriter(trial_dir=os.path.join(path, trial),
step=step,
worker=worker) as fw:
for i in range(num_tensors):
if data is None:
data = np.ones(shape=shape, dtype=dtype) * step
fw.write_tensor(tdata=data,
tname=f"{tname_prefix}_{i}",
mode=mode,
mode_step=mode_step)
if export_colls:
c = CollectionManager()
c.add("default")
c.get("default").tensor_names = [
f"{tname_prefix}_{i}" for i in range(num_tensors)
]
c.add("gradients")
c.get("gradients").tensor_names = [
f"{tname_prefix}_{i}" for i in range(num_tensors)
]
c.export(os.path.join(path, trial), DEFAULT_COLLECTIONS_FILE_NAME)
def test_mode_data():
run_id = "trial_" + datetime.now().strftime("%Y%m%d-%H%M%S%f")
trial_dir = "/tmp/ts_outputs/" + run_id
c = CollectionManager()
c.add("default")
c.get("default").tensor_names = ["arr_1"]
c.get("default").tensor_names = ["arr_2"]
c.export(trial_dir, DEFAULT_COLLECTIONS_FILE_NAME)
trial = create_trial(trial_dir)
worker = socket.gethostname()
for s in range(0, 10):
fw = FileWriter(trial_dir=trial_dir, step=s, worker=worker)
if s % 2 == 0:
fw.write_tensor(
tdata=np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32),
tname="arr_1",
mode=modes.TRAIN,
mode_step=s // 2,
)
else:
fw.write_tensor(
tdata=np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32),
tname="arr_2",
mode=modes.EVAL,
mode_step=s // 2,
)
fw.close()
assert trial.tensor_names() == ["arr_1", "arr_2"]
assert trial.tensor_names(step=0) == ["arr_1"]
assert trial.tensor_names(step=1) == ["arr_2"]
assert trial.tensor_names(step=0, mode=modes.TRAIN) == ["arr_1"]
assert trial.tensor_names(step=0, mode=modes.EVAL) == ["arr_2"]
assert trial.tensor_names(mode=modes.TRAIN) == ["arr_1"]
assert trial.tensor_names(mode=modes.EVAL) == ["arr_2"]
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()
def write_dummy_collection_file(trial):
cm = CollectionManager()
cm.create_collection("default")
cm.add(Collection(trial))
cm.export(trial, DEFAULT_COLLECTIONS_FILE_NAME)
def test_manager():
cm = CollectionManager()
cm.create_collection("default")
cm.get("default").include("loss")
cm.get("default").add_tensor_name("assaas")
cm.add(Collection("trial1"))
cm.add("trial2")
cm.get("trial2").include("total_loss")
assert len(cm.collections) == 3
assert cm.get("default") == cm.collections["default"]
assert "loss" in cm.get("default").include_regex
assert len(cm.get("default").tensor_names) > 0
assert "total_loss" in cm.collections["trial2"].include_regex
def test_manager_export_load():
cm = CollectionManager()
cm.create_collection("default")
cm.get("default").include("loss")
cm.add(Collection("trial1"))
cm.add("trial2")
cm.get("trial2").include("total_loss")
cm.export("/tmp/dummy_trial", DEFAULT_COLLECTIONS_FILE_NAME)
cm2 = CollectionManager.load(
os.path.join(get_path_to_collections("/tmp/dummy_trial"),
DEFAULT_COLLECTIONS_FILE_NAME))
assert cm == cm2
def _read_collections(self, collection_files):
first_collection_file = collection_files[0] # First Collection File
self.collection_manager = CollectionManager.load(first_collection_file)
self.num_workers = self.collection_manager.get_num_workers()
def test_mode_data():
run_id = "trial_" + datetime.now().strftime("%Y%m%d-%H%M%S%f")
trial_dir = "/tmp/ts_outputs/" + run_id
c = CollectionManager()
c.add("default")
c.get("default").tensor_names = ["arr"]
c.export(trial_dir, DEFAULT_COLLECTIONS_FILE_NAME)
tr = create_trial(trial_dir)
worker = socket.gethostname()
for s in range(0, 10):
fw = FileWriter(trial_dir=trial_dir, step=s, worker=worker)
if s % 2 == 0:
fw.write_tensor(
tdata=np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32),
tname="arr",
mode=modes.TRAIN,
mode_step=s // 2,
)
else:
fw.write_tensor(
tdata=np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32),
tname="arr",
mode=modes.EVAL,
mode_step=s // 2,
)
fw.close()
if s % 2 == 0:
assert tr.has_passed_step(s // 2,
mode=modes.TRAIN) == StepState.AVAILABLE
assert tr.has_passed_step(
s // 2, mode=modes.EVAL) == StepState.NOT_YET_AVAILABLE
else:
assert tr.has_passed_step(s // 2,
mode=modes.EVAL) == StepState.AVAILABLE
assert tr.has_passed_step(s) == StepState.AVAILABLE
assert tr.has_passed_step(s + 1) == StepState.NOT_YET_AVAILABLE
assert tr.has_passed_step(
s + 1, mode=modes.TRAIN) == StepState.NOT_YET_AVAILABLE
assert len(tr.tensor_names()) == 1
assert len(tr.steps()) == 10
assert len(tr.steps(mode=modes.TRAIN)) == 5
assert len(tr.steps(mode=modes.EVAL)) == 5
assert len(tr.modes()) == 2
for i in range(10):
if i % 2 == 0:
assert tr.mode(i) == modes.TRAIN
else:
assert tr.mode(i) == modes.EVAL
assert tr.mode_step(i) == i // 2
for i in range(5):
assert tr.global_step(modes.TRAIN, i) == (i * 2)
assert tr.global_step(modes.EVAL, i) == (i * 2) + 1
assert len(tr.tensor("arr").steps()) == 10
assert len(tr.tensor("arr").steps(mode=modes.TRAIN)) == 5
assert len(tr.tensor("arr").steps(mode=modes.EVAL)) == 5
for i in range(10):
assert tr.tensor("arr").value(i) is not None
if i < 5:
assert tr.tensor("arr").value(i, mode=modes.TRAIN) is not None
assert tr.tensor("arr").value(i, mode=modes.EVAL) is not None
shutil.rmtree("/tmp/ts_outputs/" + run_id)
