def save(self, path, dependency_path, aws_credentials = {}):
""" Save predictive object to the given path
Parameters
----------
path : str
The location to save the predictive object to
"""
# only support saving to local or S3 for now
if (not fu.is_s3_path(path)) and (not fu.is_local_path(path)):
raise RuntimeError("Only save to local and S3 path is supported, cannot \
save predictive object to path %s. " % path)
if (fu.is_s3_path(path)):
self._save_s3(path, dependency_path, aws_credentials)
else:
if os.path.exists(path):
if os.path.isfile(path):
__logger__.warning("Overwriting existing file '%s' when saving predictive object" % path)
else:
raise RuntimeError("Path %s already exists, please remove that and save again" % path)
self._save_local(path, dependency_path, aws_credentials)
tracker = _mt._get_metric_tracker()
tracker.track('deploy.predictive_service.predictive_object',
value=1,
properties={
'type': self.__class__.__name__,
'dependencies': len(self.dependencies) if self.dependencies else 0}
)
def _load_local(cls, path):
path = fu.expand_full_path(path)
if not os.path.isfile(path):
raise RuntimeError("path %s is not a file, a tar.gz file is expected." % path)
rm_tmp_tree = True
try:
po_path = tempfile.mkdtemp()
fu.unzip_directory(path, po_path)
po_object_path = os.path.join(po_path, 'definition')
# load the describe
describe = cls._load_object(cls._get_describe_path(po_object_path))
po_schema_version = describe.get('schema_version', PREDICTIVE_OBJECT_SCHEMA_VERSION)
if po_schema_version > PREDICTIVE_OBJECT_SCHEMA_VERSION:
raise RuntimeError("Your GraphLab Create only supports Predictive "
"Objects with schema version up to '%s', the Predictive Object "
"you are trying to load has schema version '%s'. Please upgrade your "
"GraphLab Create version to most up-to-date one." % \
(PREDICTIVE_OBJECT_SCHEMA_VERSION, po_schema_version))
po_obj = None
# load the instance
try:
with open(cls._get_definition_path(po_object_path), 'rb') as f:
po_obj = _pickle.load(f)
except Exception as e:
__logger__.error('Unable to load object: %s' % (e.message))
raise
po_obj.description = str(describe['description'])
po_obj.schema_version = int(describe['schema_version'])
__logger__.info('Loaded predictive object of type: %s' % (type(po_obj)))
# load all dependencies
dependencies = describe['dependencies']
gl_objects = {}
for (uri, gl_obj_info) in dependencies.iteritems():
gl_obj_type = gl_obj_info['type']
gl_obj_path = gl_obj_info['path']
# v1.0 format is different from v1.1
# v1.0 we save all dependent objects together with the PO definition and
# remember relative path of the dependent object
# v1.1 we save absolute path of the GraphLab objects so the objects are
# directly loaded from s3
if fu.is_local_path(gl_obj_path) and not os.path.isabs(gl_obj_path):
rm_tmp_tree = False
gl_obj_path = os.path.join(po_object_path, 'GraphLabObjects', uri)
__logger__.info("Loading old format Predictive Object dependencies, changed \
path from '%s' to '%s'" % (gl_obj_info['path'], gl_obj_path))
gl_objects[uri] = cls._load_graphlab_object(gl_obj_type, gl_obj_path)
po_obj.dependencies = gl_objects
finally:
# For v1.0 since we are downloading all GraphLab objects and load GraphLab
# objects from there, we cannot really remove the folder. For v1.1 it is
# fine because GraphLab server loads all GraphLab objects from S3 and manages
# temp objects life time
if rm_tmp_tree:
shutil.rmtree(po_path)
return po_obj
def _save_imp(self, po_path, dependency_path, aws_credentials):
'''Save the predictive object to a directory
The files for a predictive object are laid out the following way:
po_path/definition/meta -- serialized json file about the predictive
object, including: description, dependencies, etc.
po_path/definition/definition -- cloudpickle-serialized PredictiveObject
dependency_path -- all dependent GraphLab objects, each in its
own directory:
dependency_path/uri1/ -- serialized GraphLab object with uri1
dependency_path/uri2/ -- serialized GraphLab object with uri2
'''
fu.create_directory(po_path)
describe = {
'description': self.description,
'dependencies': {},
'schema_version' : self.schema_version
}
for (uri, gl_obj) in self.dependencies.iteritems():
# If it isn't already saved, save it.
temp_path = None
try:
if not fu.is_path(gl_obj):
obj_type = self._get_graphlab_object_type(gl_obj)
temp_path = tempfile.mkdtemp()
__logger__.info("Saving dependent GraphLab %s (%s) locally to '%s' " % (obj_type, uri, temp_path))
gl_obj.save(temp_path)
gl_obj = temp_path
else:
obj_type = get_graphlab_object_type(gl_obj)
# Copy the saved object without loading it.
save_path = os.path.join(dependency_path, uri)
__logger__.info("Copying dependent GraphLab %s(%s) from '%s' to '%s' " % (obj_type, uri, gl_obj, save_path))
if fu.is_s3_path(gl_obj) and fu.is_s3_path(save_path):
fu.intra_s3_copy_model(gl_obj, save_path, aws_credentials)
elif fu.is_local_path(gl_obj) and fu.is_s3_path(save_path):
fu.s3_copy_model(gl_obj, save_path, aws_credentials)
elif fu.is_local_path(gl_obj) and fu.is_local_path(save_path):
# Useful for unit tests
shutil.copytree(gl_obj, save_path)
else:
raise RuntimeError("Copy GraphLab object from S3 to local path is not supported. GraphLab object path: %s, save path: %s" % (gl_obj, save_path))
finally:
if temp_path:
shutil.rmtree(temp_path)
# add to the global describe dictionary
describe['dependencies'][uri] = {
'path': save_path,
'type': obj_type
}
# persist the global description
describe_path = self._get_describe_path(po_path)
self._save_object(describe_path, describe)
# persist the definition of myself
definition_path = self._get_definition_path(po_path)
try:
with open(definition_path, 'wb') as f:
_cloudpickle.dump(self, f)
except Exception as e:
__logger__.error('Unable to save object: %s' % (e.message))
raise e
