def __init__(self, config, train, val):
validate_config(config)
self.config = config
self.model = get_net(config["model"])
self.train_dataset = train
self.val_dataset = val
self.warmup_epochs = config.get("warmup_num", 0)
# build experiment directory name
self.experiment_base_dir = "_".join(
map(
str,
[
"experiment",
self.config["experiment"]["folder"],
self.config["experiment"]["name"],
self.config["train"]["transform"]["size"],
self.config["model"]["arch"],
self.config["num_epochs"],
self.config["batch_size"],
str(int(datetime.timestamp(datetime.now()))),
],
))
# create experiment directory
os.makedirs(self.experiment_base_dir, exist_ok=True)
self.metric_counter = get_metric_counter(config,
self.experiment_base_dir)
self.steps_per_epoch = config.get("steps_per_epoch",
len(self.train_dataset))
self.validation_steps = config.get("validation_steps",
len(self.val_dataset))
self.sample_input_shape = config.get("sample_input_shape",
(1, 3, 32, 32))
self.criterion = get_loss(self.config["model"]["loss"])
self.optimizer = get_optimizer(
config=config,
params=filter(lambda p: p.requires_grad, self.model.parameters()),
)
self.scheduler = get_scheduler(config, self.optimizer)
self.early_stopping = EarlyStopping(
patience=self.config["early_stopping"])
self.model_adapter = get_model_adapter(self.config["model"])
self.monitor = TrainingMonitor(config["training_monitor"]["method"],
config["training_monitor"]["interval"])
# setup logging to log file
logger.addHandler(
logging.FileHandler(self.experiment_base_dir + "/experiment.log"))
# copy config file to experiment directory
with open(osp.join(self.experiment_base_dir, "config.yaml"),
"w") as outfile:
yaml.dump(self.config, outfile, default_flow_style=False)
def test_dice_loss(tensor1, tensor2, tensor3, tensor4):
dice_loss = loss.get_loss({"name": "dice_loss"})
assert dice_loss.forward(tensor1, tensor1).data.numpy() == pytest.approx(
0.71844566, standard_approx)
assert dice_loss.forward(tensor2, tensor2).data.numpy() == pytest.approx(
0.71844566, standard_approx)
assert dice_loss.forward(tensor1, tensor2).data.numpy() == pytest.approx(
0.54923755, standard_approx)
assert dice_loss.forward(tensor3, tensor4).data.numpy() == pytest.approx(
0.57556236, standard_approx)
def test_mixed_loss(tensor1, tensor2, tensor3, tensor4):
mixed_loss = loss.get_loss({"name": "mixed_loss"})
assert mixed_loss.forward(tensor1, tensor1).data.numpy() == pytest.approx(
1.3103894, standard_approx)
assert mixed_loss.forward(tensor2, tensor2).data.numpy() == pytest.approx(
1.3103894, standard_approx)
assert mixed_loss.forward(tensor1, tensor2).data.numpy() == pytest.approx(
4.975, standard_approx)
assert mixed_loss.forward(tensor3, tensor4).data.numpy() == pytest.approx(
3.230158, standard_approx)
def test_iou_loss(tensor1, tensor2, tensor3, tensor4):
iou_loss = loss.get_loss({"name": "iou_loss"})
assert iou_loss.iou_metric(tensor1, tensor1) == 1.0
assert iou_loss.iou_metric(tensor2, tensor2) == 1.0
# almost zero
iou_result = iou_loss.iou_metric(tensor1, tensor2)
assert abs(iou_result) < standard_approx
assert iou_loss.iou_metric(tensor3, tensor4).data.numpy() == pytest.approx(
0.3333333, standard_approx)
def _init_params(self):
""" Initializes Trainer
Initialized attributes:
- criterion: loss to be used during training
- optimizer: model optimizer
- scheduler: scheduler for optimizer
- early_stopping: Early Stopping technique, which stops training if validation loss doesn't improve
after a given patience
- model_adapter: adapter for a given model
"""
self.criterion = get_loss(self.config['model']['loss'])
self.optimizer = self._get_optim(
filter(lambda p: p.requires_grad, self.model.parameters()))
self.scheduler = self._get_scheduler(self.optimizer)
self.early_stopping = EarlyStopping(
patience=self.config['early_stopping'])
self.model_adapter = get_model_adapter(self.config)
os.makedirs(osp.join(self.config['experiment']['folder'],
self.config['experiment']['name']),
exist_ok=True)
def test_getting_incorrect_loss():
with pytest.raises(ValueError):
assert loss.get_loss(
{"name": "this_is_some_incorrect_name_for_loss_function"})
def test_getting_correct_loss():
assert loss.get_loss({"name": "lovasz"})