def load_model(experiment_id: Optional[str] = None,
operator_id: Optional[str] = None) -> Dict[str, object]:
"""Retrieves a model from object storage.
Args:
experiment_id (str, optional): the experiment uuid. Defaults to None.
operator_id (str, optional): the operator uuid. Defaults to None.
Returns:
dict: A dictionary of models.
Raises:
TypeError: when experiment_id is undefined in args and env.
TypeError: when operator_id is undefined in args and env.
"""
if experiment_id is None:
experiment_id = get_experiment_id()
if operator_id is None:
operator_id = get_operator_id()
try:
object_name = f"{PREFIX_1}/{experiment_id}/{PREFIX_2}/{operator_id}/{MODEL_FILE}"
data = MINIO_CLIENT.get_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
)
except (NoSuchBucket, NoSuchKey):
return {}
buffer = BytesIO(data.read())
return load(buffer)
def list_metrics(experiment_id: Optional[str] = None,
operator_id: Optional[str] = None,
run_id: Optional[str] = None) -> List[Dict[str, object]]:
"""Lists metrics from object storage.
Args:
experiment_id (str, optional): the experiment uuid. Defaults to None.
operator_id (str, optional): the operator uuid. Defaults to None.
run_id (str, optional): the run id. Defaults to None.
Returns:
list: A list of metrics.
Raises:
TypeError: when experiment_id is undefined in args and env.
TypeError: when operator_id is undefined in args and env.
"""
if experiment_id is None:
experiment_id = get_experiment_id()
if operator_id is None:
operator_id = get_operator_id()
# ensures MinIO bucket exists
make_bucket(BUCKET_NAME)
if run_id is None:
# gets run_id from env variable
# Attention: returns None if env is unset
run_id = get_run_id()
elif run_id == "latest":
try:
metadata = stat_metadata(experiment_id, operator_id)
run_id = metadata.get("run_id")
except FileNotFoundError:
return []
try:
object_name = operator_filepath(METRICS_FILE, experiment_id, operator_id, run_id)
data = MINIO_CLIENT.get_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
)
except (NoSuchBucket, NoSuchKey):
raise FileNotFoundError(f"No such file or directory: '{experiment_id}'")
return load(data)
def save_metrics(experiment_id: Optional[str] = None,
operator_id: Optional[str] = None,
run_id: Optional[str] = None,
**kwargs):
"""Saves metrics of an experiment to the object storage.
Args:
experiment_id (str, optional): the experiment uuid. Defaults to None
operator_id (str, optional): the operator uuid. Defaults to None
run_id (str, optional): the run id. Defaults to None.
**kwargs: the metrics dict.
Raises:
TypeError: when experiment_id is undefined in args and env.
TypeError: when operator_id is undefined in args and env.
"""
if experiment_id is None:
experiment_id = get_experiment_id()
if operator_id is None:
operator_id = get_operator_id()
if run_id is None:
# gets run_id from env variables
# Attention: returns None if env is unset
run_id = get_run_id()
# ensures MinIO bucket exists
make_bucket(BUCKET_NAME)
if run_id:
metadata = {}
try:
metadata = stat_metadata(experiment_id, operator_id)
if run_id == "latest":
run_id = metadata.get("run_id")
except FileNotFoundError:
pass
metadata["run_id"] = run_id
# encodes metadata to JSON format and uploads to MinIO
buffer = BytesIO(dumps(metadata).encode())
MINIO_CLIENT.put_object(
bucket_name=BUCKET_NAME,
object_name=f'experiments/{experiment_id}/operators/{operator_id}/.metadata',
data=buffer,
length=buffer.getbuffer().nbytes,
)
object_name = operator_filepath(METRICS_FILE, experiment_id, operator_id, run_id)
encoded_metrics = []
# retrieves the metrics saved previosuly
try:
data = MINIO_CLIENT.get_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
)
encoded_metrics = loads(data.read())
except NoSuchKey:
pass
# appends new metrics
encoded_metrics.extend(_encode_metrics(kwargs))
# puts metrics into buffer
buffer = BytesIO(dumps(encoded_metrics).encode())
# uploads metrics to MinIO
MINIO_CLIENT.put_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
data=buffer,
length=buffer.getbuffer().nbytes,
)
def load_dataset(
name: str,
run_id: Optional[str] = None,
operator_id: Optional[str] = None,
page: Optional[int] = None,
page_size: Optional[int] = None) -> Union[pd.DataFrame, BinaryIO]:
"""Retrieves the contents of a dataset.
If run_id exists, then loads the dataset from the specified run.
If the dataset does not exist for given run_id/operator_id return the
'original' dataset
Args:
name (str): the dataset name.
run_id (str, optional): the run id of training pipeline. Defaults to None.
operator_id (str, optional): the operator uuid. Defaults to None.
Returns:
The contents of a dataset. Either a `pandas.DataFrame` or an `BinaryIO` buffer.
Raises:
FileNotFoundError: If dataset does not exist in the object storage.
"""
if run_id is None:
# gets run_id from env variable
# Attention: returns None if env is unset
run_id = get_run_id()
elif run_id == "latest":
metadata = stat_dataset(name)
run_id = metadata.get("run_id")
# when the dataset does not exist for given run_id/operator_id
# must return the 'original' dataset
# unset run_id so data_filepath points to the 'original' dataset
if run_id and operator_id:
try:
metadata = stat_dataset(name, run_id, operator_id)
except FileNotFoundError:
run_id = None
elif run_id:
try:
run_metadata = stat_dataset(name, run_id)
operator_id = run_metadata.get("operator_id")
except FileNotFoundError:
run_id = None
# builds the path to the dataset file
path = _data_filepath(name, run_id, operator_id)
if page_size and page_size > 0:
nrows = page_size
else:
nrows = None
if page and page > 0:
skiprows = (page - 1) * page_size
else:
skiprows = None
try:
metadata = stat_dataset(name, run_id, operator_id)
dataset = pd.read_csv(S3FS.open(path),
header=0,
index_col=False,
nrows=nrows,
skiprows=skiprows)
dtypes = dict((column, "object") for column, ftype in zip(
metadata["columns"], metadata["featuretypes"])
if ftype in [CATEGORICAL, DATETIME])
dataset = dataset.astype(dtypes)
except (UnicodeDecodeError, pd.errors.EmptyDataError,
pd.errors.ParserError):
# reads the raw file
data = MINIO_CLIENT.get_object(
bucket_name=BUCKET_NAME,
object_name=path.lstrip(f"{BUCKET_NAME}/"),
)
return BytesIO(data.read())
except KeyError:
# metadata file does not contains "columns" or "featuretypes"
# ignore this error and return dataset without cast its type
pass
except FileNotFoundError:
raise FileNotFoundError("The specified dataset does not exist")
return dataset
def stat_dataset(name: str,
run_id: Optional[str] = None,
operator_id: Optional[str] = None) -> Dict[str, str]:
"""Retrieves the metadata of a dataset.
Args:
name (str): the dataset name.
run_id (str, optional): the run id of trainning pipeline. Defaults to None.
operator_id (str, optional): the operator uuid. Defaults to None.
Returns:
dict: The metadata.
Raises:
FileNotFoundError: If dataset does not exist in the object storage.
"""
metadata = {}
# remove /tmp/data/ from dataset name
# because in jupyter we use dataset with full path
name = name.replace("/tmp/data/", "")
if run_id == "latest":
metadata = stat_dataset(name)
run_id = metadata.get("run_id")
if run_id is None:
# gets run_id from env variables
# Attention: returns None if env is unset
run_id = get_run_id()
if run_id and operator_id:
# get metadata for a specific operator of a run, if exists
object_name = _metadata_filepath(name, run_id, operator_id)
elif run_id:
# if no metadata was generated by the operator,
# get the last one generated by the pipeline flow
object_name = _metadata_filepath(name, run_id)
if not metadata_exists(name, run_id):
# if it is at the beginning of a run,
# there will be no metadata generated by run_id
object_name = _metadata_filepath(name)
else:
# unable to get run_id automatically,
# this function is probably being called out of a run
object_name = _metadata_filepath(name)
else:
# get path according to received parameters
run_id = None if run_id == "root" else run_id
object_name = _metadata_filepath(name, run_id, operator_id)
try:
# reads the .metadata file
data = MINIO_CLIENT.get_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
)
# decodes the metadata (which is in JSON format)
metadata = loads(data.read())
except (NoSuchBucket, NoSuchKey):
raise FileNotFoundError("The specified dataset does not exist")
return metadata