def test_upgrade_to_latest(self):
for p in glob.iglob(
os.path.join(os.path.dirname(__file__), "json_config/*.json")):
print("Trying to upgrade file:" + p)
with PathManager.open(p) as f:
test_data = json.load(f)
for test_case in test_data:
# make sure the config can be instantiated, don't need return value
pytext_config_from_json(test_case["original"])
def load_v1(state):
config = pytext_config_from_json(state[CONFIG_JSON])
task = create_task(config.task,
metadata=state[DATA_STATE],
model_state=state[MODEL_STATE])
return task, config
def load_v1(state):
config = pytext_config_from_json(state[CONFIG_JSON])
# importing in file level generates circular import/dependency failures,
# that need refator later to fix
from .task import create_task
task = create_task(config.task,
metadata=state[DATA_STATE],
model_state=state[MODEL_STATE])
return task, config
def load_v3(state, overwrite_config=None, rank=0, world_size=1):
saved_config = pytext_config_from_json(state[CONFIG_JSON])
if overwrite_config:
config = overwrite_config
print(f"Use config from current task")
else:
config = saved_config
print(f"Use config saved in snapshot")
model_state = state[MODEL_STATE]
training_state = state[TRAINING_STATE]
if training_state and training_state.tensorizers:
tensorizers = training_state.tensorizers
else:
tensorizers = state[TENSORIZERS]
# importing in file level generates circular import/dependency failures,
# that need refator later to fix
from .task import create_task
task = create_task(
config.task,
metadata=state[DATA_STATE],
model_state=model_state,
tensorizers=tensorizers,
rank=rank,
world_size=world_size,
)
# TODO: T53664090 @stevenliu save & load state_dict() of optimizer and scheduler
if training_state:
if training_state.model is None and task.model:
training_state.model = task.model
if training_state.optimizer and task.trainer.optimizer:
"""
https://pytorch.org/tutorials/beginner/saving_loading_models.html
Unpickling optimizer object from checkpoint could result in a
different parameter copy from model parameters. Especially in
mixied precision training, which optimizer param_groups maintains
master weights copy instead of the model parameters.
The suggested loading mechanism is
model = TheModelClass(*args, **kwargs)
optimizer = TheOptimizerClass(model.parameters(), *args, **kwargs)
checkpoint = torch.load(PATH)
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
"""
optimizer = task.trainer.optimizer
optimizer.load_state_dict(training_state.optimizer.state_dict())
training_state.optimizer = optimizer
return task, config, training_state
def load_v2(state):
config = pytext_config_from_json(state[CONFIG_JSON])
model_state = state[MODEL_STATE]
tensorizers = state[TENSORIZERS]
task = create_task(
config.task,
metadata=state[DATA_STATE],
model_state=model_state,
tensorizers=tensorizers,
)
return task, config
def load(load_path: str):
"""
Load task, will construct the task using the saved config then load metadata
and model state.
"""
if not (load_path and os.path.isfile(load_path)):
raise ValueError(f"Invalid snapshot path{load_path}")
print(f"Loading model from {load_path}...")
state = torch.load(load_path, map_location=lambda storage, loc: storage)
config = pytext_config_from_json(state[CONFIG_JSON])
task = create_task(config.task,
metadata=state[DATA_STATE],
model_state=state[MODEL_STATE])
return task, config
def load_v2(state):
config = pytext_config_from_json(state[CONFIG_JSON])
model_state = state[MODEL_STATE]
tensorizers = state[TENSORIZERS]
# importing in file level generates circular import/dependency failures,
# that need refator later to fix
from .task import create_task
task = create_task(
config.task,
metadata=state[DATA_STATE],
model_state=model_state,
tensorizers=tensorizers,
)
return task, config, None
