how to save the last epoch only?

[machin-x]

"monitor (Optional[str]) – quantity to monitor. By default it is None which saves a checkpoint only for the last epoch."
when i trainning a model, i set the 'monitor' to None, it should save the last epoch as the doc says. but it still save depend on the val_loss, it always save the model with lowest val_loss.

i also try another way, set the 'save_last' to True. while this needs to set a monitor. And if i set save_top_k to 0, it will save nothing; if set to 1, it will save 2 models, the best one and the last one. But i just want to save the last one.

is this a bug or i made sth wrong? is there a way to save model with epoch asigned myself? such as the last 3 epochs?

[awaelchli]

Hey! Have a look at this example:

import os
import torch
from torch.utils.data import Dataset
from pytorch_lightning import LightningModule, Trainer
from pytorch_lightning.callbacks import ModelCheckpoint


class RandomDataset(Dataset):

    def __init__(self, size, length):
        self.len = length
        self.data = torch.randn(length, size)

    def __getitem__(self, index):
        return self.data[index]

    def __len__(self):
        return self.len


class BoringModel(LightningModule):

    def __init__(self):
        super().__init__()
        self.layer = torch.nn.Linear(32, 2)

    def forward(self, x):
        return self.layer(x)

    def training_step(self, batch, batch_idx):
        loss = self(batch).sum()
        self.log("train_loss", loss)
        return {"loss": loss}

    def validation_step(self, batch, batch_idx):
        loss = self(batch).sum()
        self.log("valid_loss", loss, logger=False)
        return {"x": loss}

    def configure_optimizers(self):
        return torch.optim.SGD(self.layer.parameters(), lr=0.1)


def run():
    train_data = torch.utils.data.DataLoader(RandomDataset(32, 64), batch_size=2, num_workers=0)
    val_data = torch.utils.data.DataLoader(RandomDataset(32, 64), batch_size=2, num_workers=0)

    model = BoringModel()
    trainer = Trainer(
        default_root_dir=os.getcwd(),
        limit_train_batches=1,
        limit_val_batches=1,
        num_sanity_val_steps=0,
        max_epochs=5,  # this will save a checkpoint at epoch index 4 (last epoch)
        weights_summary=None,
        logger=False,
        callbacks=[ModelCheckpoint(dirpath="./checkpoints", monitor=None)]
    )
    trainer.fit(model, train_dataloader=train_data, val_dataloaders=val_data)


if __name__ == '__main__':
    run()

I'm choosing:
ModelCheckpoint(dirpath="./checkpoints", monitor=None)

That's all, it saves only one checkpoint, named epoch=4-step=4.ckpt, it corresponds to the last epoch being run.
Note: for backward compatibility, when monitor=None, we choose "val_loss" as the monitor when it is available. You should be able to avoid that by just renaming your validation loss to "valid_loss" or something else :)

[machin-x]

OK~ thx ! :D