def load_from_checkpoint(filename: str,
dataset=None,
use_tmp_log_folder=True,
device="cpu") -> "KgeModel":
"""Loads a model from a checkpoint file of a training job.
If dataset is specified, associates this dataset with the model. Otherwise uses
the dataset used to train the model.
If `use_tmp_log_folder` is set, the logs and traces are written to a temporary
file. Otherwise, the files `kge.log` and `trace.yaml` will be created (or
appended to) in the checkpoint's folder.
"""
checkpoint = torch.load(filename, map_location=device)
original_config = checkpoint["config"]
config = Config() # round trip to handle deprecated configs
config.load_options(original_config.options)
config.set("job.device", device)
if use_tmp_log_folder:
import tempfile
config.log_folder = tempfile.mkdtemp(prefix="kge-")
else:
config.log_folder = os.path.dirname(filename)
if not config.log_folder:
config.log_folder = "."
if dataset is None:
dataset = Dataset.load(config, preload_data=False)
model = KgeModel.create(config, dataset)
model.load(checkpoint["model"])
model.eval()
return model
def create_default(
model: Optional[str] = None,
dataset: Optional[Union[Dataset, str]] = None,
options: Dict[str, Any] = {},
folder: Optional[str] = None,
) -> "KgeModel":
"""Utility method to create a model, including configuration and dataset.
`model` is the name of the model (takes precedence over
``options["model"]``), `dataset` a dataset name or `Dataset` instance (takes
precedence over ``options["dataset.name"]``), and options arbitrary other
configuration options.
If `folder` is ``None``, creates a temporary folder. Otherwise uses the
specified folder.
"""
# load default model config
if model is None:
model = options["model"]
default_config_file = filename_in_module(kge.model,
"{}.yaml".format(model))
config = Config()
config.load(default_config_file, create=True)
# apply specified options
config.set("model", model)
if isinstance(dataset, Dataset):
config.set("dataset.name", dataset.config.get("dataset.name"))
elif isinstance(dataset, str):
config.set("dataset.name", dataset)
config.set_all(new_options=options)
# create output folder
if folder is None:
config.folder = tempfile.mkdtemp("{}-{}-".format(
config.get("dataset.name"), config.get("model")))
else:
config.folder = folder
# create dataset and model
if not isinstance(dataset, Dataset):
dataset = Dataset.load(config)
model = KgeModel.create(config, dataset)
return model
def main():
# default config
config = Config()
# now parse the arguments
parser = create_parser(config)
args, unknown_args = parser.parse_known_args()
# If there where unknown args, add them to the parser and reparse. The correctness
# of these arguments will be checked later.
if len(unknown_args) > 0:
parser = create_parser(
config, filter(lambda a: a.startswith("--"), unknown_args)
)
args = parser.parse_args()
# process meta-commands
process_meta_command(args, "create", {"command": "start", "run": False})
process_meta_command(args, "eval", {"command": "resume", "job.type": "eval"})
process_meta_command(
args, "test", {"command": "resume", "job.type": "eval", "eval.split": "test"}
)
process_meta_command(
args, "valid", {"command": "resume", "job.type": "eval", "eval.split": "valid"}
)
# dump command
if args.command == "dump":
dump(args)
exit()
# start command
if args.command == "start":
# use toy config file if no config given
if args.config is None:
args.config = kge_base_dir() + "/" + "examples/toy-complex-train.yaml"
print("WARNING: No configuration specified; using " + args.config)
print("Loading configuration {}...".format(args.config))
config.load(args.config)
# resume command
if args.command == "resume":
if os.path.isdir(args.config) and os.path.isfile(args.config + "/config.yaml"):
args.config += "/config.yaml"
print("Resuming from configuration {}...".format(args.config))
config.load(args.config)
config.folder = os.path.dirname(args.config)
if not config.folder:
config.folder = "."
if not os.path.exists(config.folder):
raise ValueError(
"{} is not a valid config file for resuming".format(args.config)
)
# overwrite configuration with command line arguments
for key, value in vars(args).items():
if key in [
"command",
"config",
"run",
"folder",
"checkpoint",
"abort_when_cache_outdated",
]:
continue
if value is not None:
if key == "search.device_pool":
value = "".join(value).split(",")
try:
if isinstance(config.get(key), bool):
value = argparse_bool_type(value)
except KeyError:
pass
config.set(key, value)
if key == "model":
config._import(value)
# initialize output folder
if args.command == "start":
if args.folder is None: # means: set default
config_name = os.path.splitext(os.path.basename(args.config))[0]
config.folder = os.path.join(
kge_base_dir(),
"local",
"experiments",
datetime.datetime.now().strftime("%Y%m%d-%H%M%S") + "-" + config_name,
)
else:
config.folder = args.folder
# catch errors to log them
try:
if args.command == "start" and not config.init_folder():
raise ValueError("output folder {} exists already".format(config.folder))
config.log("Using folder: {}".format(config.folder))
# determine checkpoint to resume (if any)
if hasattr(args, "checkpoint"):
if args.checkpoint == "default":
if config.get("job.type") in ["eval", "valid"]:
checkpoint_file = config.checkpoint_file("best")
else:
checkpoint_file = None # means last
elif is_number(args.checkpoint, int) or args.checkpoint == "best":
checkpoint_file = config.checkpoint_file(args.checkpoint)
else:
# otherwise, treat it as a filename
checkpoint_file = args.checkpoint
# disable processing of outdated cached dataset files globally
Dataset._abort_when_cache_outdated = args.abort_when_cache_outdated
# log configuration
config.log("Configuration:")
config.log(yaml.dump(config.options), prefix=" ")
config.log("git commit: {}".format(get_git_revision_short_hash()), prefix=" ")
# set random seeds
if config.get("random_seed.python") > -1:
import random
random.seed(config.get("random_seed.python"))
if config.get("random_seed.torch") > -1:
import torch
torch.manual_seed(config.get("random_seed.torch"))
if config.get("random_seed.numpy") > -1:
import numpy.random
numpy.random.seed(config.get("random_seed.numpy"))
# let's go
if args.command == "start" and not args.run:
config.log("Job created successfully.")
else:
# load data
dataset = Dataset.load(config)
# let's go
job = Job.create(config, dataset)
if args.command == "resume":
job.resume(checkpoint_file)
job.run()
except BaseException as e:
tb = traceback.format_exc()
config.log(tb, echo=False)
raise e from None
