class MetricsTracker:
def __init__(self, vh: VisdomHandler, single_stage: bool):
self.__vh = vh
self.__train_loss = LossTracker() if single_stage else MultipleLossTracker()
self.__val_loss = LossTracker() if single_stage else MultipleLossTracker()
self.__best_val_loss = 100.0
self.__best_mean, self.__best_median, self.__best_trimean = 100.0, 100.0, 100.0
self.__best_bst25, self.__best_wst25, self.__best_pct95 = 100.0, 100.0, 100.0
def update_train_loss(self, loss: Union[float, list], epoch: int):
self.__train_loss.update(loss)
self.__train_loss.update_visualization(self.__vh, epoch, name="Training loss")
def update_val_loss(self, loss: Union[float, list], epoch: int):
self.__val_loss.update(loss)
self.__train_loss.update_visualization(self.__vh, epoch, name="Validation loss")
def get_train_loss_value(self) -> float:
return self.__train_loss.get_angular_loss()
def get_val_loss_value(self) -> float:
return self.__val_loss.get_angular_loss()
def get_train_loss_tracker(self) -> Union[LossTracker, MultipleLossTracker]:
return self.__train_loss
def get_val_loss_tracker(self) -> Union[LossTracker, MultipleLossTracker]:
return self.__val_loss
def get_best_val_loss(self) -> float:
return self.__best_val_loss
def reset_losses(self):
self.__train_loss.reset()
self.__val_loss.reset()
def update_metrics(self, metrics: dict):
self.__best_val_loss = self.__val_loss.get_angular_loss()
self.__best_mean = metrics["mean"]
self.__best_median = metrics["median"]
self.__best_trimean = metrics["trimean"]
self.__best_bst25 = metrics["bst25"]
self.__best_wst25 = metrics["wst25"],
self.__best_pct95 = metrics["pct95"]
def log_metrics(self, log_name: str, epoch: int):
log_table = {
"train_loss": self.__train_loss.get_angular_loss(),
"val_loss": self.__val_loss.get_angular_loss(),
"best_val_loss": self.__best_val_loss,
"mean": self.__best_mean,
"median": self.__best_median,
"trimean": self.__best_trimean,
"bst25": self.__best_bst25,
"wst25": self.__best_wst25,
"pct95": self.__best_pct95
}
open(log_name, 'a').write("Stats for epoch {}:\n {} \n".format(epoch, json.dumps(log_table, indent=2)))
class MultipleLossTracker:
def __init__(self):
self.__loss = LossTracker()
self.__loss1 = LossTracker()
self.__loss2 = LossTracker()
self.__loss3 = LossTracker()
def reset(self):
self.__loss.reset()
self.__loss1.reset()
self.__loss2.reset()
self.__loss3.reset()
def update(self, losses: list):
self.__loss.update(losses[0])
self.__loss1.update(losses[1])
self.__loss2.update(losses[2])
self.__loss3.update(losses[3])
def update_visualization(self, vh: VisdomHandler, epoch: int, name: str):
self.__loss.update_visualization(vh, epoch, name)
self.__loss1.update_visualization(vh, epoch, name)
self.__loss2.update_visualization(vh, epoch, name)
self.__loss3.update_visualization(vh, epoch, name)
def get_step_loss(self, step: int) -> float:
if step == 1:
return self.__loss1.get_angular_loss()
if step == 2:
return self.__loss2.get_angular_loss()
if step == 3:
return self.__loss3.get_angular_loss()
raise ValueError(
"{} is not a valid step, must be either 1, 2 or 3!".format(step))
def get_angular_loss(self):
return self.__loss.get_angular_loss()