---
title: ModelCheckpoint
description: >-
  Save the model after every epoch by monitoring a quantity. Every logged
  metrics are passed to the

  Logger for the version it gets saved in the same directory as the

  checkpoint.
canonical: >-
  https://lightning.ai/docs/pytorch/stable/optional-api/api/lightning.pytorch.callbacks.ModelCheckpoint
---
# ModelCheckpoint {#modelcheckpoint}

### `class lightning.pytorch.callbacks.ModelCheckpoint(dirpath=None, filename=None, monitor=None, verbose=False, save_last=None, save_top_k=1, save_on_exception=False, save_weights_only=False, mode='min', auto_insert_metric_name=True, every_n_train_steps=None, train_time_interval=None, every_n_epochs=None, save_on_train_epoch_end=None, enable_version_counter=True)` {#lightning.pytorch.callbacks.ModelCheckpoint}

Bases: `Checkpoint`

Save the model after every epoch by monitoring a quantity. Every logged metrics are passed to the [`Logger`](/docs/pytorch/2.6.6/optional-api/api/lightning.pytorch.loggers.logger.md#lightning.pytorch.loggers.logger.Logger "lightning.pytorch.loggers.logger.Logger") for the version it gets saved in the same directory as the checkpoint.

After training finishes, use `best_model_path` to retrieve the path to the best checkpoint file and `best_model_score` to get its score.

Note

When using manual optimization with `every_n_train_steps`, you should save the model state in your `training_step` before the optimizer step if you want the checkpoint to reflect the pre-optimization state. Example:

```python
def training_step(self, batch, batch_idx):
    # ... forward pass, loss calculation, backward pass ...

    # Save model state before optimization
    if not hasattr(self, 'saved_models'):
        self.saved_models = {}
    self.saved_models[batch_idx] = {
        k: v.detach().clone()
        for k, v in self.layer.state_dict().items()
    }

    # Then perform optimization
    optimizer.zero_grad()
    self.manual_backward(loss)
    optimizer.step()

    # Optional: Clean up old states to save memory
    if batch_idx > 10:  # Keep last 10 states
        del self.saved_models[batch_idx - 10]
```

Parameters:

*   **dirpath** (`Union`\[[`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)"), [`Path`](https://docs.python.org/3/library/pathlib.html#pathlib.Path "(in Python v3.15)"), [`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")\]) –
    
    Directory to save the model file. Example: `dirpath='my/path/'`.
    
    Warning
    
    In a distributed environment like DDP, it’s recommended to provide a dirpath to avoid race conditions. When using manual optimization with `every_n_train_steps`, make sure to save the model state in your training loop as shown in the example above.
    
    Can be remote file paths such as s3://mybucket/path/ or ‘hdfs://path/’ (default: `None`). If dirpath is `None`, we only keep the `k` best checkpoints in memory, and do not save anything to disk.
    
*   **filename** ([`Optional`](https://docs.python.org/3/library/typing.html#typing.Optional "(in Python v3.15)")\[[`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)")\]) – Checkpoint filename. Can contain named formatting options to be auto-filled. If no name is provided, it will be `None` and the checkpoint will be saved to `{epoch}`.and if the Trainer uses a logger, the path will also contain logger name and version.
    
*   **filename** –
    
    checkpoint filename. Can contain named formatting options to be auto-filled.
    
    Example:
    
    ```python
    # save any arbitrary metrics like `val_loss`, etc. in name
    # saves a file like: my/path/epoch=2-val_loss=0.02-other_metric=0.03.ckpt
    >>> checkpoint_callback = ModelCheckpoint(
    ...     dirpath='my/path',
    ...     filename='{epoch}-{val_loss:.2f}-{other_metric:.2f}'
    ... )
    ```
    
    By default, filename is `None` and will be set to `'{epoch}-{step}'`, where “epoch” and “step” match the number of finished epoch and optimizer steps respectively.
    
*   **monitor** ([`Optional`](https://docs.python.org/3/library/typing.html#typing.Optional "(in Python v3.15)")\[[`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)")\]) – quantity to monitor. By default it is `None` which saves a checkpoint only for the last epoch.
    
*   **verbose** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – verbosity mode. Default: `False`.
    
*   **save\_last** (`Union`\[[`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)"), [`Literal`](https://docs.python.org/3/library/typing.html#typing.Literal "(in Python v3.15)")\[`'link'`\], [`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")\]) – When `True`, saves a last.ckpt copy whenever a checkpoint file gets saved. Can be set to `'link'` on a local filesystem to create a symbolic link. This allows accessing the latest checkpoint in a deterministic manner. Default: `None`.
    
*   **save\_top\_k** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – if `save_top_k == k`, the best k models according to the quantity monitored will be saved. If `save_top_k == 0`, no models are saved. If `save_top_k == -1`, all models are saved. Please note that the monitors are checked every `every_n_epochs` epochs. If `save_top_k >= 2` and the callback is called multiple times inside an epoch, and the filename remains unchanged, the name of the saved file will be appended with a version count starting with `v1` to avoid collisions unless `enable_version_counter` is set to False. The version counter is unrelated to the top-k ranking of the checkpoint, and we recommend formatting the filename to include the monitored metric to avoid collisions.
    
*   **save\_on\_exception** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Whether to save a checkpoint when an exception is raised. Default: `False`.
    
*   **mode** ([`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)")) – one of {min, max}. If `save_top_k != 0`, the decision to overwrite the current save file is made based on either the maximization or the minimization of the monitored quantity. For `'val_acc'`, this should be `'max'`, for `'val_loss'` this should be `'min'`, etc.
    
*   **auto\_insert\_metric\_name** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – When `True`, the checkpoints filenames will contain the metric name. For example, `filename='checkpoint_{epoch:02d}-{acc:02.0f}` with epoch `1` and acc `1.12` will resolve to `checkpoint_epoch=01-acc=01.ckpt`. Is useful to set it to `False` when metric names contain `/` as this will result in extra folders. For example, `filename='epoch={epoch}-step={step}-val_acc={val/acc:.2f}', auto_insert_metric_name=False`
    
*   **save\_weights\_only** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – if `True`, then only the model’s weights will be saved. Otherwise, the optimizer states, lr-scheduler states, etc are added in the checkpoint too.
    
*   **every\_n\_train\_steps** ([`Optional`](https://docs.python.org/3/library/typing.html#typing.Optional "(in Python v3.15)")\[[`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")\]) –
    
    How many training steps to wait before saving a checkpoint. This does not take into account the steps of the current epoch. If `every_n_train_steps == None or every_n_train_steps == 0`, no checkpoints will be saved during training. Mutually exclusive with `train_time_interval` and `every_n_epochs`.
    
    Note
    
    When using with manual optimization, the checkpoint will be saved after the optimizer step by default. To save the model state before the optimizer step, you need to save the model state in your `training_step` before calling `optimizer.step()`. See the class docstring for an example.
    
*   **train\_time\_interval** ([`Optional`](https://docs.python.org/3/library/typing.html#typing.Optional "(in Python v3.15)")\[[`timedelta`](https://docs.python.org/3/library/datetime.html#datetime.timedelta "(in Python v3.15)")\]) – Checkpoints are monitored at the specified time interval. For all practical purposes, this cannot be smaller than the amount of time it takes to process a single training batch. This is not guaranteed to execute at the exact time specified, but should be close. This must be mutually exclusive with `every_n_train_steps` and `every_n_epochs`.
    
*   **every\_n\_epochs** ([`Optional`](https://docs.python.org/3/library/typing.html#typing.Optional "(in Python v3.15)")\[[`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")\]) – Number of epochs between checkpoints. This value must be `None` or non-negative. To disable saving top-k checkpoints, set `every_n_epochs = 0`. This argument does not impact the saving of `save_last=True` checkpoints. If all of `every_n_epochs`, `every_n_train_steps` and `train_time_interval` are `None`, we save a checkpoint at the end of every epoch (equivalent to `every_n_epochs = 1`). If `every_n_epochs == None` and either `every_n_train_steps != None` or `train_time_interval != None`, saving at the end of each epoch is disabled (equivalent to `every_n_epochs = 0`). This must be mutually exclusive with `every_n_train_steps` and `train_time_interval`. Setting both `ModelCheckpoint(..., every_n_epochs=V, save_on_train_epoch_end=False)` and `Trainer(max_epochs=N, check_val_every_n_epoch=M)` will only save checkpoints at epochs 0 < E <= N where both values for `every_n_epochs` and `check_val_every_n_epoch` evenly divide E.
    
*   **save\_on\_train\_epoch\_end** ([`Optional`](https://docs.python.org/3/library/typing.html#typing.Optional "(in Python v3.15)")\[[`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")\]) – Whether to run checkpointing at the end of the training epoch. If `True`, checkpoints are saved at the end of every training epoch. If `False`, checkpoints are saved at the end of validation. If `None` (default), checkpointing behavior is determined based on training configuration. If `val_check_interval` is a str, dict, or timedelta (time-based), checkpointing is performed after validation. If `check_val_every_n_epoch != 1`, checkpointing will not be performed at the end of every training epoch. If there are no validation batches of data, checkpointing will occur at the end of the training epoch. If there is a non-default number of validation runs per training epoch (`val_check_interval != 1`), checkpointing is performed after validation.
    
*   **enable\_version\_counter** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Whether to append a version to the existing file name. If `False`, then the checkpoint files will be overwritten.
    

Note

For extra customization, ModelCheckpoint includes the following attributes:

*   `CHECKPOINT_JOIN_CHAR = "-"`
    
*   `CHECKPOINT_EQUALS_CHAR = "="`
    
*   `CHECKPOINT_NAME_LAST = "last"`
    
*   `FILE_EXTENSION = ".ckpt"`
    
*   `STARTING_VERSION = 1`
    

For example, you can change the default last checkpoint name by doing `checkpoint_callback.CHECKPOINT_NAME_LAST = "{epoch}-last"`

If you want to checkpoint every N hours, every M train batches, and/or every K val epochs, then you should create multiple `ModelCheckpoint` callbacks.

If the checkpoint’s `dirpath` changed from what it was before while resuming the training, only `best_model_path` will be reloaded and a warning will be issued.

If you provide a `filename` on a mounted device where changing permissions is not allowed (causing `chmod` to raise a `PermissionError`), install fsspec>=2025.5.0. Then the error is caught, the file’s permissions remain unchanged, and the checkpoint is still saved. Otherwise, no checkpoint will be saved and training stops.

Raises:

*   **MisconfigurationException** – If `save_top_k` is smaller than `-1`, if `monitor` is `None` and `save_top_k` is none of `None`, `-1`, and `0`, or if `mode` is none of `"min"` or `"max"`.
    
*   [**ValueError**](https://docs.python.org/3/builtins/exceptions.html#ValueError "(in Python v3.15)") – If `trainer.save_checkpoint` is `None`.
    

Example:

```python
>>> from lightning.pytorch import Trainer
>>> from lightning.pytorch.callbacks import ModelCheckpoint

# saves checkpoints to 'my/path/' at every epoch
>>> checkpoint_callback = ModelCheckpoint(dirpath='my/path/')
>>> trainer = Trainer(callbacks=[checkpoint_callback])

# save epoch and val_loss in name
# saves a file like: my/path/sample-mnist-epoch=02-val_loss=0.32.ckpt
>>> checkpoint_callback = ModelCheckpoint(
...     monitor='val_loss',
...     dirpath='my/path/',
...     filename='sample-mnist-{epoch:02d}-{val_loss:.2f}'
... )

# save epoch and val_loss in name, but specify the formatting yourself (e.g. to avoid problems with Tensorboard
# or other loggers, due to the presence of characters like '=' or '/')
# saves a file like: my/path/sample-mnist-epoch02-val_loss0.32.ckpt
>>> checkpoint_callback = ModelCheckpoint(
...     monitor='val/loss',
...     dirpath='my/path/',
...     filename='sample-mnist-epoch{epoch:02d}-val_loss{val/loss:.2f}',
...     auto_insert_metric_name=False
... )

# retrieve the best checkpoint after training
>>> checkpoint_callback = ModelCheckpoint(dirpath='my/path/')
>>> trainer = Trainer(callbacks=[checkpoint_callback])
>>> model = ...  
>>> trainer.fit(model)  
>>> print(checkpoint_callback.best_model_path)  
```

Tip

Saving and restoring multiple checkpoint callbacks at the same time is supported under variation in the following arguments:

_monitor, mode, every\_n\_train\_steps, every\_n\_epochs, train\_time\_interval_

Read more: [Persisting Callback State](/docs/pytorch/2.6.6/reference/extensions/callbacks_state.md#save-callback-state)

#### `file_exists(filepath, trainer)` {#lightning.pytorch.callbacks.ModelCheckpoint.file_exists}

Checks if a file exists on rank 0 and synchronizes the result to all other ranks, preventing the internal state to diverge between ranks.

Return type:

[`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")

#### `format_checkpoint_name(metrics, filename=None, ver=None, prefix=None)` {#lightning.pytorch.callbacks.ModelCheckpoint.format_checkpoint_name}

Generate a filename according to the defined template.

Example:

```python
>>> tmpdir = os.path.dirname(__file__)
>>> ckpt = ModelCheckpoint(dirpath=tmpdir, filename='{epoch}')
>>> os.path.basename(ckpt.format_checkpoint_name(dict(epoch=0)))
'epoch=0.ckpt'
>>> ckpt = ModelCheckpoint(dirpath=tmpdir, filename='{epoch:03d}')
>>> os.path.basename(ckpt.format_checkpoint_name(dict(epoch=5)))
'epoch=005.ckpt'
>>> ckpt = ModelCheckpoint(dirpath=tmpdir, filename='{epoch}-{val_loss:.2f}')
>>> os.path.basename(ckpt.format_checkpoint_name(dict(epoch=2, val_loss=0.123456)))
'epoch=2-val_loss=0.12.ckpt'
>>> os.path.basename(ckpt.format_checkpoint_name(dict(epoch=2, val_loss=0.12), filename='{epoch:d}'))
'epoch=2.ckpt'
>>> ckpt = ModelCheckpoint(dirpath=tmpdir,
... filename='epoch={epoch}-validation_loss={val_loss:.2f}',
... auto_insert_metric_name=False)
>>> os.path.basename(ckpt.format_checkpoint_name(dict(epoch=2, val_loss=0.123456)))
'epoch=2-validation_loss=0.12.ckpt'
>>> ckpt = ModelCheckpoint(dirpath=tmpdir, filename='{missing:d}')
>>> os.path.basename(ckpt.format_checkpoint_name({}))
'missing=0.ckpt'
>>> ckpt = ModelCheckpoint(filename='{step}')
>>> os.path.basename(ckpt.format_checkpoint_name(dict(step=0)))
'step=0.ckpt'
```

Return type:

[`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)")

#### `load_state_dict(state_dict)` {#lightning.pytorch.callbacks.ModelCheckpoint.load_state_dict}

Called when loading a checkpoint, implement to reload callback state given callback’s `state_dict`.

Parameters:

**state\_dict** ([`dict`](https://docs.python.org/3/builtins/stdtypes.html#dict "(in Python v3.15)")\[[`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)"), [`Any`](https://docs.python.org/3/library/typing.html#typing.Any "(in Python v3.15)")\]) – the callback state returned by `state_dict`.

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `on_exception(trainer, pl_module, exception)` {#lightning.pytorch.callbacks.ModelCheckpoint.on_exception}

Save a checkpoint when an exception is raised.

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `on_train_batch_end(trainer, pl_module, outputs, batch, batch_idx)` {#lightning.pytorch.callbacks.ModelCheckpoint.on_train_batch_end}

Save checkpoint on train batch end if we meet the criteria for every\_n\_train\_steps

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `on_train_end(trainer, pl_module)` {#lightning.pytorch.callbacks.ModelCheckpoint.on_train_end}

Ensure save\_last=True is applied when training ends.

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `on_train_epoch_end(trainer, pl_module)` {#lightning.pytorch.callbacks.ModelCheckpoint.on_train_epoch_end}

Save a checkpoint at the end of the training epoch.

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `on_train_start(trainer, pl_module)` {#lightning.pytorch.callbacks.ModelCheckpoint.on_train_start}

Called when the train begins.

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `on_validation_end(trainer, pl_module)` {#lightning.pytorch.callbacks.ModelCheckpoint.on_validation_end}

Save a checkpoint at the end of the validation stage.

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `setup(trainer, pl_module, stage)` {#lightning.pytorch.callbacks.ModelCheckpoint.setup}

Called when fit, validate, test, predict, or tune begins.

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `state_dict()` {#lightning.pytorch.callbacks.ModelCheckpoint.state_dict}

Called when saving a checkpoint, implement to generate callback’s `state_dict`.

Return type:

[`dict`](https://docs.python.org/3/builtins/stdtypes.html#dict "(in Python v3.15)")\[[`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)"), [`Any`](https://docs.python.org/3/library/typing.html#typing.Any "(in Python v3.15)")\]

Returns:

A dictionary containing callback state.

#### `to_yaml(filepath=None)` {#lightning.pytorch.callbacks.ModelCheckpoint.to_yaml}

Saves the best\_k\_models dict containing the checkpoint paths with the corresponding scores to a YAML file.

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `property state_key: str` {#lightning.pytorch.callbacks.ModelCheckpoint.state_key}

Identifier for the state of the callback.

Used to store and retrieve a callback’s state from the checkpoint dictionary by `checkpoint["callbacks"][state_key]`. Implementations of a callback need to provide a unique state key if 1) the callback has state and 2) it is desired to maintain the state of multiple instances of that callback.

---

**Navigation**

- Parent: [Api](/docs/pytorch/2.6.6/optional-api/api.md)