我尝试使用ModelCheckpoint来保存每个时期验证损失中的最佳性能模型。

class model(pl.lightningModule)
   :
   :
   :

   def validation_step(self, batch, batch_idx):
        if batch_idx == 0:
            self.totalValLoss = 0
            self.totalValToken = 0
        batch = Batch(batch[0], batch[1])
        out = self(batch.src, batch.trg, batch.src_mask, batch.trg_mask)
        out = self.generator(out)
        criterion = LabelSmoothing(size=V, padding_idx=0, smoothing=0)
        loss = criterion(out.contiguous().view(-1, out.size(-1)), batch.trg_y.contiguous().view(-1)) / batch.ntokens
        self.totalValLoss += loss * batch.ntokens
        self.totalValToken += batch.ntokens
        if batch_idx == 99:
            self.totalValLoss = self.totalValLoss / self.totalValToken
            print(f"valLoss: {self.totalValLoss}")
        self.log("val_loss", self.totalValLoss)
        return {"val_loss": self.totalValLoss}

if __name__ == '__main__':

    if True:
        model = model(...)

        checkpoint_callback = 
        ModelCheckpoint(dirpath="D:/PycharmProjects/Transformer/Models", 
        save_top_k=2, monitor="val_loss")
        trainer = pl.Trainer(max_epochs=10, callbacks=[checkpoint_callback])
        trainer.fit(model)

在运行代码之后，我期望两个性能最好的模型会被保存到目录“D：/PycharmProjects/Transformer/Models”中，但是这并没有发生。并且在运行时也没有显示错误。