def test_auroc(pos_label):
auroc = AUROC(pos_label=pos_label)
assert auroc.name == 'auroc'
pred, target = torch.tensor([1, 2, 3, 4]), torch.tensor([1, 2, 0, 1])
area = auroc(pred=pred, target=target, sample_weight=[0.1, 0.2, 0.3, 0.4])
assert isinstance(area, torch.Tensor)
def validation_epoch_end(self, outputs):
avg_loss = torch.stack([x["val_loss"] for x in outputs]).mean()
y = torch.cat([x["y"] for x in outputs], dim=0).flatten()
y_hat = torch.cat([x["y_hat"] for x in outputs], dim=0).flatten()
avg_acc = (y_hat.round() == y).float().mean()
auc = AUROC()(y_hat,
y.float()) if y.float().mean() > 0 else torch.tensor(0.5)
metrics = {"val_loss": avg_loss, "val_acc": avg_acc, "val_auc": auc}
self.logger.log_metrics(metrics, step=self.current_epoch)
return metrics
def validation_epoch_end(self, outputs):
avg_loss = torch.stack([x["val_loss"] for x in outputs]).mean()
y = torch.cat([x["y"] for x in outputs])
y_hat = torch.cat([x["y_hat"] for x in outputs])
auc = (
AUROC()(pred=y_hat, target=y) if y.float().mean() > 0 else 0.5
) # skip sanity check
acc = (y_hat.round() == y).float().mean().item()
print(f"Epoch {self.current_epoch} acc:{acc} auc:{auc}")
tensorboard_logs = {"val_loss": avg_loss, "val_auc": auc, "val_acc": acc}
return {
"avg_val_loss": avg_loss,
"val_auc": auc,
"val_acc": acc,
"log": tensorboard_logs,
}
def validation_epoch_end(self, outputs): # 한 에폭이 끝났을 때 실행
avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
y = torch.cat([x['y'] for x in outputs])
y_hat = torch.cat([x['y_hat'] for x in outputs])
auc = AUROC()(y_hat,
y) if y.float().mean() > 0 else 0.5 # skip sanity check
acc = (y_hat.round() == y).float().mean().item()
print(f"Epoch {self.current_epoch} acc:{acc} auc:{auc}")
tensorboard_logs = {
'val_loss': avg_loss,
'val_auc': auc,
'val_acc': acc
}
return {
'avg_val_loss': avg_loss,
'val_auc': auc,
'val_acc': acc,
'log': tensorboard_logs
}
def validation_epoch_end(self, outputs):
avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
LOGITS = torch.cat([x['logits'] for x in outputs])
LABELS = torch.cat([x['labels'] for x in outputs])
# Skip Sanity Check
auc = AUROC()(pred=LOGITS.detach(), target=LABELS.detach()) if LABELS.float().mean() > 0 else 0.5
logs = {'val/epoch_loss': avg_loss.item(), 'val/epoch_auc': auc}
# Log loss, auc
self.experiment.log_metrics(logs, step=self.epoch_num)
# Update Epoch Num
self.epoch_num += 1
# Save Weights
if self.best_loss > avg_loss:
self.best_loss = avg_loss
filename = f'{self.cfg.exp.exp_name}_epoch_{self.epoch_num}_loss_{self.best_loss:.3f}_auc_{auc:.3f}.pth'
torch.save(self.net.state_dict(), filename)
self.experiment.log_model(name=filename, file_or_folder='./'+filename)
os.remove(filename)
self.best_auc = auc
return {'avg_val_loss': avg_loss}