def test_unpack(self):
fsm = FileStructManager(base_dir=self.base_dir, is_continue=False)
cm = CheckpointsManager(fsm)
f = open(cm.weights_file(), 'w')
f.write('1')
f.close()
f = open(cm.optimizer_state_file(), 'w')
f.write('2')
f = open(cm.trainer_file(), 'w')
f.write('3')
f.close()
cm.pack()
try:
cm.unpack()
except Exception as err:
self.fail('Exception on unpacking')
for i, f in enumerate(
[cm.weights_file(),
cm.optimizer_state_file(),
cm.trainer_file()]):
if not (os.path.exists(f) and os.path.isfile(f)):
self.fail("File '{}' doesn't remove after pack".format(f))
with open(f, 'r') as file:
if file.read() != str(i + 1):
self.fail("File content corrupted")
def _save_state(self, ckpts_manager: CheckpointsManager, best_ckpts_manager: CheckpointsManager or None,
cur_best_state: float or None, epoch_idx: int) -> float or None:
"""
Internal method used for save states after epoch end
:param ckpts_manager: ordinal checkpoints manager
:param best_ckpts_manager: checkpoints manager, used for store best stages
:param cur_best_state: current best stage metric value
:return: new best stage metric value or None if it not update
"""
def save_trainer(ckp_manager):
with open(ckp_manager.trainer_file(), 'w') as out:
json.dump({'last_epoch': epoch_idx}, out)
if self._best_state_rule is not None:
new_best_state = self._best_state_rule()
if cur_best_state is None:
self._data_processor.save_state()
save_trainer(ckpts_manager)
ckpts_manager.pack()
return new_best_state
else:
if new_best_state <= cur_best_state:
self._data_processor.set_checkpoints_manager(best_ckpts_manager)
self._data_processor.save_state()
save_trainer(best_ckpts_manager)
best_ckpts_manager.pack()
self._data_processor.set_checkpoints_manager(ckpts_manager)
return new_best_state
self._data_processor.save_state()
save_trainer(ckpts_manager)
ckpts_manager.pack()
return None
def __init__(self,
model: Model,
fsm: FileStructManager,
device: torch.device = None):
self._fsm = fsm
self.__data_processor = DataProcessor(model, device=device)
checkpoint_manager = CheckpointsManager(self._fsm)
self.__data_processor.set_checkpoints_manager(checkpoint_manager)
checkpoint_manager.unpack()
self.__data_processor.load()
checkpoint_manager.pack()
def __init__(self,
model: Model,
fsm: FileStructManager,
from_best_state: bool = False):
self._fsm = fsm
self._data_processor = DataProcessor(model)
checkpoint_manager = CheckpointsManager(
self._fsm, prefix='best' if from_best_state else None)
self._data_processor.set_checkpoints_manager(checkpoint_manager)
checkpoint_manager.unpack()
self._data_processor.load()
checkpoint_manager.pack()
def _resume(self) -> int:
if self._resume_from == 'last':
ckpts_manager = self._checkpoint_manager
elif self._checkpoint_manager == 'best':
ckpts_manager = CheckpointsManager(self._fsm, 'best')
else:
raise NotImplementedError("Resume parameter may be only 'last' or 'best' not {}".format(self._resume_from))
ckpts_manager.unpack()
self._data_processor.load()
with open(ckpts_manager.trainer_file(), 'r') as file:
start_epoch_idx = json.load(file)['last_epoch'] + 1
ckpts_manager.pack()
return start_epoch_idx
def test_pack(self):
fsm = FileStructManager(base_dir=self.base_dir, is_continue=False)
cm = CheckpointsManager(fsm)
with self.assertRaises(CheckpointsManager.SMException):
cm.pack()
os.mkdir(cm.weights_file())
os.mkdir(cm.optimizer_state_file())
with self.assertRaises(CheckpointsManager.SMException):
cm.pack()
shutil.rmtree(cm.weights_file())
shutil.rmtree(cm.optimizer_state_file())
f = open(cm.weights_file(), 'w')
f.close()
f = open(cm.optimizer_state_file(), 'w')
f.close()
f = open(cm.trainer_file(), 'w')
f.close()
try:
cm.pack()
except Exception as err:
self.fail('Exception on packing files: [{}]'.format(err))
for f in [cm.weights_file(), cm.optimizer_state_file()]:
if os.path.exists(f) and os.path.isfile(f):
self.fail("File '{}' doesn't remove after pack".format(f))
result = os.path.join(
fsm.get_path(cm, check=False, create_if_non_exists=False),
'last_checkpoint.zip')
self.assertTrue(os.path.exists(result) and os.path.isfile(result))
f = open(cm.weights_file(), 'w')
f.close()
f = open(cm.optimizer_state_file(), 'w')
f.close()
f = open(cm.trainer_file(), 'w')
f.close()
try:
cm.pack()
result = os.path.join(
fsm.get_path(cm, check=False, create_if_non_exists=False),
'last_checkpoint.zip.old')
self.assertTrue(os.path.exists(result) and os.path.isfile(result))
except Exception as err:
self.fail('Fail to pack with existing previous state file')