def test_continue_from_checkpoint(self):
def on_node(n1, n2):
self.assertTrue(np.array_equal(n1.numpy(), n2.numpy()))
model = SimpleModel().train()
loss = SimpleLoss()
for optim in [
torch.optim.SGD(model.parameters(), lr=0.1),
torch.optim.Adam(model.parameters(), lr=0.1)
]:
train_config = TrainConfig([], loss, optim)
dp_before = TrainDataProcessor(model=model,
train_config=train_config)
before_state_dict = model.state_dict().copy()
dp_before.update_lr(0.023)
with self.assertRaises(Model.ModelException):
dp_before.save_state()
try:
fsm = FileStructManager(base_dir=self.base_dir,
is_continue=False)
dp_before.set_checkpoints_manager(CheckpointsManager(fsm))
dp_before.save_state()
except:
self.fail(
"Exception on saving state when 'CheckpointsManager' specified"
)
fsm = FileStructManager(base_dir=self.base_dir, is_continue=True)
dp_after = TrainDataProcessor(model=model,
train_config=train_config)
with self.assertRaises(Model.ModelException):
dp_after.load()
try:
cm = CheckpointsManager(fsm)
dp_after.set_checkpoints_manager(cm)
dp_after.load()
except:
self.fail('DataProcessor initialisation raises exception')
after_state_dict = model.state_dict().copy()
dict_pair_recursive_bypass(before_state_dict, after_state_dict,
on_node)
self.assertEqual(dp_before.get_lr(), dp_after.get_lr())
shutil.rmtree(self.base_dir)
class Trainer:
"""
Class, that run drive process.
Trainer get list of training stages and every epoch loop over it.
Training process looks like:
.. highlight:: python
.. code-block:: python
for epoch in epochs_num:
for stage in training_stages:
stage.run()
monitor_hub.update_metrics(stage.metrics_processor().get_metrics())
save_state()
on_epoch_end_callback()
:param model: model for training
:param train_config: :class:`TrainConfig` object
:param fsm: :class:`FileStructManager` object
:param device: device for training process
"""
class TrainerException(Exception):
def __init__(self, msg):
super().__init__()
self._msg = msg
def __str__(self):
return self._msg
def __init__(self,
model: Module,
train_config: TrainConfig,
fsm: FileStructManager,
device: torch.device = None):
self._fsm = fsm
self.monitor_hub = MonitorHub()
self._checkpoint_manager = CheckpointsManager(self._fsm)
self.__epoch_num = 100
self._resume_from = None
self._on_epoch_end = []
self._best_state_rule = None
self.__train_config = train_config
self._device = device
self._data_processor = TrainDataProcessor(model, self.__train_config, self._device) \
.set_checkpoints_manager(self._checkpoint_manager)
self._lr = LearningRate(self._data_processor.get_lr())
def set_epoch_num(self, epoch_number: int) -> 'Trainer':
"""
Define number of epoch for training. One epoch - one iteration over all train stages
:param epoch_number: number of training epoch
:return: self object
"""
self.__epoch_num = epoch_number
return self
def resume(self, from_best_checkpoint: bool) -> 'Trainer':
"""
Resume train from last checkpoint
:param from_best_checkpoint: is need to continue from best checkpoint
:return: self object
"""
self._resume_from = 'last' if from_best_checkpoint is False else 'best'
return self
def enable_lr_decaying(self, coeff: float, patience: int,
target_val_clbk: callable) -> 'Trainer':
"""
Enable rearing rate decaying. Learning rate decay when `target_val_clbk` returns doesn't update
minimum for `patience` steps
:param coeff: lr decay coefficient
:param patience: number of steps
:param target_val_clbk: callback which returns the value that is used for lr decaying
:return: self object
"""
self._lr = DecayingLR(self._data_processor.get_lr(), coeff, patience,
target_val_clbk)
return self
def train(self) -> None:
"""
Run training process
"""
if len(self.__train_config.stages()) < 1:
raise self.TrainerException("There's no sages for training")
best_checkpoints_manager = None
cur_best_state = None
if self._best_state_rule is not None:
best_checkpoints_manager = CheckpointsManager(self._fsm, 'best')
start_epoch_idx = 1
if self._resume_from is not None:
start_epoch_idx += self._resume()
self.monitor_hub.add_monitor(ConsoleMonitor())
with self.monitor_hub:
for epoch_idx in range(start_epoch_idx,
self.__epoch_num + start_epoch_idx):
self.monitor_hub.set_epoch_num(epoch_idx)
for stage in self.__train_config.stages():
stage.run(self._data_processor)
if stage.metrics_processor() is not None:
self.monitor_hub.update_metrics(
stage.metrics_processor().get_metrics())
new_best_state = self._save_state(self._checkpoint_manager,
best_checkpoints_manager,
cur_best_state, epoch_idx)
if new_best_state is not None:
cur_best_state = new_best_state
self._data_processor.update_lr(self._lr.value())
for clbk in self._on_epoch_end:
clbk()
self._update_losses()
self.__iterate_by_stages(lambda s: s.on_epoch_end())
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 _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 _update_losses(self) -> None:
"""
Update loses procedure
"""
losses = {}
for stage in self.__train_config.stages():
if stage.get_losses() is not None:
losses[stage.name()] = stage.get_losses()
self.monitor_hub.update_losses(losses)
def data_processor(self) -> TrainDataProcessor:
"""
Get data processor object
:return: data processor
"""
return self._data_processor
def enable_best_states_saving(self, rule: callable) -> 'Trainer':
"""
Enable best states saving
Best stages will save when return of `rule` update minimum
:param rule: callback which returns the value that is used for define when need store best metric
:return: self object
"""
self._best_state_rule = rule
return self
def disable_best_states_saving(self) -> 'Trainer':
"""
Enable best states saving
:return: self object
"""
self._best_state_rule = None
return self
def add_on_epoch_end_callback(self, callback: callable) -> 'Trainer':
"""
Add callback, that will be called after every epoch end
:param callback: method, that will be called. This method may not get any parameters
:return: self object
"""
self._on_epoch_end.append(callback)
return self
def __iterate_by_stages(self, func: callable) -> None:
"""
Internal method, that used for iterate by stages
:param func: callback, that calls for every stage
"""
for stage in self.__train_config.stages():
func(stage)