def __init__(self,
fsm: FileStructManager,
is_continue: bool,
network_name: str = None):
super().__init__()
self.__writer = None
self.__txt_log_file = None
fsm.register_dir(self)
dir = fsm.get_path(self)
if dir is None:
return
dir = os.path.join(dir,
network_name) if network_name is not None else dir
if not (fsm.in_continue_mode()
or is_continue) and os.path.exists(dir) and os.path.isdir(dir):
idx = 0
tmp_dir = dir + "_v{}".format(idx)
while os.path.exists(tmp_dir) and os.path.isdir(tmp_dir):
idx += 1
tmp_dir = dir + "_v{}".format(idx)
dir = tmp_dir
os.makedirs(dir, exist_ok=True)
self.__writer = SummaryWriter(dir)
self.__txt_log_file = open(os.path.join(dir, "log.txt"),
'a' if is_continue else 'w')
def test_savig_best_states(self):
fsm = FileStructManager(base_dir=self.base_dir, is_continue=False)
model = SimpleModel()
metrics_processor = MetricsProcessor()
stages = [TrainStage(TestDataProducer([[{'data': torch.rand(1, 3), 'target': torch.rand(1)}
for _ in list(range(20))]]), metrics_processor)]
trainer = Trainer(TrainConfig(model, stages, SimpleLoss(), torch.optim.SGD(model.parameters(), lr=0.1)),
fsm).set_epoch_num(3).enable_best_states_saving(lambda: np.mean(stages[0].get_losses()))
checkpoint_file = os.path.join(self.base_dir, 'checkpoints', 'last', 'last_checkpoint.zip')
best_checkpoint_file = os.path.join(self.base_dir, 'checkpoints', 'best', 'best_checkpoint.zip')
class Val:
def __init__(self):
self.v = None
first_val = Val()
def on_epoch_end(val):
if val.v is not None and np.mean(stages[0].get_losses()) < val.v:
self.assertTrue(os.path.exists(best_checkpoint_file))
os.remove(best_checkpoint_file)
val.v = np.mean(stages[0].get_losses())
return
val.v = np.mean(stages[0].get_losses())
self.assertTrue(os.path.exists(checkpoint_file))
self.assertFalse(os.path.exists(best_checkpoint_file))
os.remove(checkpoint_file)
trainer.add_on_epoch_end_callback(lambda: on_epoch_end(first_val))
trainer.train()
def test_base_ops(self):
fsm = FileStructManager(base_dir=self.base_dir, is_continue=False)
model = SimpleModel()
trainer = Trainer(TrainConfig(model, [], torch.nn.L1Loss(), torch.optim.SGD(model.parameters(), lr=1)),
fsm)
with self.assertRaises(Trainer.TrainerException):
trainer.train()
def test_train(self):
fsm = FileStructManager(base_dir=self.base_dir, is_continue=False)
model = SimpleModel()
metrics_processor = MetricsProcessor()
stages = [TrainStage(TestDataProducer([[{'data': torch.rand(1, 3), 'target': torch.rand(1)}
for _ in list(range(20))]]), metrics_processor),
ValidationStage(TestDataProducer([[{'data': torch.rand(1, 3), 'target': torch.rand(1)}
for _ in list(range(20))]]), metrics_processor)]
Trainer(TrainConfig(model, stages, SimpleLoss(), torch.optim.SGD(model.parameters(), lr=1)), fsm) \
.set_epoch_num(1).train()
def test_train_stage(self):
data_producer = DataProducer([[{'data': torch.rand(1, 3), 'target': torch.rand(1)} for _ in list(range(20))]])
metrics_processor = FakeMetricsProcessor()
train_stage = TrainStage(data_producer, metrics_processor).enable_hard_negative_mining(0.1)
fsm = FileStructManager(base_dir=self.base_dir, is_continue=False)
model = SimpleModel()
Trainer(TrainConfig(model, [train_stage], SimpleLoss(), torch.optim.SGD(model.parameters(), lr=1)), fsm) \
.set_epoch_num(1).train()
self.assertEqual(metrics_processor.call_num, len(data_producer))
def test_savig_states(self):
fsm = FileStructManager(base_dir=self.base_dir, is_continue=False)
model = SimpleModel()
metrics_processor = MetricsProcessor()
stages = [TrainStage(TestDataProducer([[{'data': torch.rand(1, 3), 'target': torch.rand(1)}
for _ in list(range(20))]]), metrics_processor)]
trainer = Trainer(TrainConfig(model, stages, SimpleLoss(), torch.optim.SGD(model.parameters(), lr=0.1)),
fsm).set_epoch_num(3)
checkpoint_file = os.path.join(self.base_dir, 'checkpoints', 'last', 'last_checkpoint.zip')
def on_epoch_end():
self.assertTrue(os.path.exists(checkpoint_file))
os.remove(checkpoint_file)
trainer.add_on_epoch_end_callback(on_epoch_end)
trainer.train()
def train():
model = resnet18(classes_num=1, in_channels=3, pretrained=True)
train_config = TrainConfig(model, [train_stage, val_stage], torch.nn.BCEWithLogitsLoss(),
torch.optim.Adam(model.parameters(), lr=1e-4))
file_struct_manager = FileStructManager(base_dir='data', is_continue=False)
trainer = Trainer(train_config, file_struct_manager, torch.device('cuda:0')).set_epoch_num(2)
tensorboard = TensorboardMonitor(file_struct_manager, is_continue=False, network_name='PortraitSegmentation')
log = LogMonitor(file_struct_manager).write_final_metrics()
trainer.monitor_hub.add_monitor(tensorboard).add_monitor(log)
trainer.enable_best_states_saving(lambda: np.mean(train_stage.get_losses()))
trainer.enable_lr_decaying(coeff=0.5, patience=10, target_val_clbk=lambda: np.mean(train_stage.get_losses()))
trainer.add_on_epoch_end_callback(lambda: tensorboard.update_scalar('params/lr', trainer.data_processor().get_lr()))
trainer.train()
def test_lr_decaying(self):
fsm = FileStructManager(base_dir=self.base_dir, is_continue=False)
model = SimpleModel()
metrics_processor = MetricsProcessor()
stages = [TrainStage(TestDataProducer([[{'data': torch.rand(1, 3), 'target': torch.rand(1)}
for _ in list(range(20))]]), metrics_processor),
ValidationStage(TestDataProducer([[{'data': torch.rand(1, 3), 'target': torch.rand(1)}
for _ in list(range(20))]]), metrics_processor)]
trainer = Trainer(TrainConfig(model, stages, SimpleLoss(), torch.optim.SGD(model.parameters(), lr=0.1)),
fsm).set_epoch_num(10)
def target_value_clbk() -> float:
return 1
trainer.enable_lr_decaying(0.5, 3, target_value_clbk)
trainer.train()
self.assertAlmostEqual(trainer.data_processor().get_lr(), 0.1 * (0.5 ** 3), delta=1e-6)
def train():
train_config = PoseNetTrainConfig()
file_struct_manager = FileStructManager(
base_dir=PoseNetTrainConfig.experiment_dir, is_continue=False)
trainer = Trainer(train_config, file_struct_manager, torch.device('cuda'))
trainer.set_epoch_num(EPOCH_NUM)
tensorboard = TensorboardMonitor(file_struct_manager, is_continue=False)
log = LogMonitor(file_struct_manager).write_final_metrics()
trainer.monitor_hub.add_monitor(tensorboard).add_monitor(log)
trainer.enable_best_states_saving(
lambda: np.mean(train_config.val_stage.get_losses()))
trainer.enable_lr_decaying(
coeff=0.5,
patience=10,
target_val_clbk=lambda: np.mean(train_config.val_stage.get_losses()))
trainer.add_on_epoch_end_callback(
lambda: tensorboard.update_scalar('params/lr',
trainer.data_processor().get_lr()))
trainer.train()