|
from torch.optim.lr_scheduler import ReduceLROnPlateau |
|
from sklearn.metrics import accuracy_score |
|
|
|
class EarlyStopping: |
|
def __init__(self, patience=5, verbose=False, delta=0): |
|
self.patience = patience |
|
self.verbose = verbose |
|
self.counter = 0 |
|
self.best_score = None |
|
self.early_stop = False |
|
self.delta = delta |
|
|
|
def __call__(self, val_loss, model): |
|
score = -val_loss |
|
|
|
if self.best_score is None: |
|
self.best_score = score |
|
self.save_checkpoint(val_loss, model) |
|
elif score < self.best_score + self.delta: |
|
self.counter += 1 |
|
if self.verbose: |
|
print(f'EarlyStopping counter: {self.counter} out of {self.patience}') |
|
if self.counter >= self.patience: |
|
self.early_stop = True |
|
else: |
|
self.best_score = score |
|
self.save_checkpoint(val_loss, model) |
|
self.counter = 0 |
|
|
|
def save_checkpoint(self, val_loss, model): |
|
if self.verbose: |
|
print(f'Validation loss decreased ({self.best_score:.6f} --> {val_loss:.6f}). Saving model ...') |
|
torch.save(model.state_dict(), 'checkpoint.pt') |
|
|