Shortcuts

Callback


A callback is a self-contained program that can be reused across projects.

Lightning has a callback system to execute them when needed. Callbacks should capture NON-ESSENTIAL logic that is NOT required for your lightning module to run.

Here’s the flow of how the callback hooks are executed:

An overall Lightning system should have:

  1. Trainer for all engineering

  2. LightningModule for all research code.

  3. Callbacks for non-essential code.


Example:

from pytorch_lightning.callbacks import Callback


class MyPrintingCallback(Callback):
    def on_train_start(self, trainer, pl_module):
        print("Training is starting")

    def on_train_end(self, trainer, pl_module):
        print("Training is ending")


trainer = Trainer(callbacks=[MyPrintingCallback()])

We successfully extended functionality without polluting our super clean lightning module research code.



Built-in Callbacks

Lightning has a few built-in callbacks.

Note

For a richer collection of callbacks, check out our bolts library.

BackboneFinetuning

Finetune a backbone model based on a learning rate user-defined scheduling.

BaseFinetuning

This class implements the base logic for writing your own Finetuning Callback.

BasePredictionWriter

Base class to implement how the predictions should be stored.

BatchSizeFinder

The BatchSizeFinder callback tries to find the largest batch size for a given model that does not give an out of memory (OOM) error.

Callback

Abstract base class used to build new callbacks.

DeviceStatsMonitor

Automatically monitors and logs device stats during training, validation and testing stage.

EarlyStopping

Monitor a metric and stop training when it stops improving.

GradientAccumulationScheduler

Change gradient accumulation factor according to scheduling.

LambdaCallback

Create a simple callback on the fly using lambda functions.

LearningRateFinder

The LearningRateFinder callback enables the user to do a range test of good initial learning rates, to reduce the amount of guesswork in picking a good starting learning rate.

LearningRateMonitor

Automatically monitor and logs learning rate for learning rate schedulers during training.

ModelCheckpoint

Save the model periodically by monitoring a quantity.

ModelPruning

Model pruning Callback, using PyTorch's prune utilities.

ModelSummary

Generates a summary of all layers in a LightningModule.

ProgressBarBase

The base class for progress bars in Lightning.

QuantizationAwareTraining

Quantization allows speeding up inference and decreasing memory requirements by performing computations and storing tensors at lower bitwidths (such as INT8 or FLOAT16) than floating point precision.

RichModelSummary

Generates a summary of all layers in a LightningModule with rich text formatting.

RichProgressBar

Create a progress bar with rich text formatting.

StochasticWeightAveraging

Implements the Stochastic Weight Averaging (SWA) Callback to average a model.

Timer

The Timer callback tracks the time spent in the training, validation, and test loops and interrupts the Trainer if the given time limit for the training loop is reached.

TQDMProgressBar

This is the default progress bar used by Lightning.


Save Callback state

Some callbacks require internal state in order to function properly. You can optionally choose to persist your callback’s state as part of model checkpoint files using state_dict() and load_state_dict(). Note that the returned state must be able to be pickled.

When your callback is meant to be used only as a singleton callback then implementing the above two hooks is enough to persist state effectively. However, if passing multiple instances of the callback to the Trainer is supported, then the callback must define a state_key property in order for Lightning to be able to distinguish the different states when loading the callback state. This concept is best illustrated by the following example.

class Counter(Callback):
    def __init__(self, what="epochs", verbose=True):
        self.what = what
        self.verbose = verbose
        self.state = {"epochs": 0, "batches": 0}

    @property
    def state_key(self):
        # note: we do not include `verbose` here on purpose
        return self._generate_state_key(what=self.what)

    def on_train_epoch_end(self, *args, **kwargs):
        if self.what == "epochs":
            self.state["epochs"] += 1

    def on_train_batch_end(self, *args, **kwargs):
        if self.what == "batches":
            self.state["batches"] += 1

    def load_state_dict(self, state_dict):
        self.state.update(state_dict)

    def state_dict(self):
        return self.state.copy()


# two callbacks of the same type are being used
trainer = Trainer(callbacks=[Counter(what="epochs"), Counter(what="batches")])

A Lightning checkpoint from this Trainer with the two stateful callbacks will include the following information:

{
    "state_dict": ...,
    "callbacks": {
        "Counter{'what': 'batches'}": {"batches": 32, "epochs": 0},
        "Counter{'what': 'epochs'}": {"batches": 0, "epochs": 2},
        ...
    }
}

The implementation of a state_key is essential here. If it were missing, Lightning would not be able to disambiguate the state for these two callbacks, and state_key by default only defines the class name as the key, e.g., here Counter.


Best Practices

The following are best practices when using/designing callbacks.

  1. Callbacks should be isolated in their functionality.

  2. Your callback should not rely on the behavior of other callbacks in order to work properly.

  3. Do not manually call methods from the callback.

  4. Directly calling methods (eg. on_validation_end) is strongly discouraged.

  5. Whenever possible, your callbacks should not depend on the order in which they are executed.


Entry Points

Lightning supports registering Trainer callbacks directly through Entry Points. Entry points allow an arbitrary package to include callbacks that the Lightning Trainer can automatically use, without you having to add them to the Trainer manually. This is useful in production environments where it is common to provide specialized monitoring and logging callbacks globally for every application.

Here is a callback factory function that returns two special callbacks:

factories.py
def my_custom_callbacks_factory():
    return [MyCallback1(), MyCallback2()]

If we make this factories.py file into an installable package, we can define an entry point for this factory function. Here is a minimal example of the setup.py file for the package my-package:

setup.py
from setuptools import setup

setup(
    name="my-package",
    version="0.0.1",
    install_requires=["pytorch-lightning"],
    entry_points={
        "pytorch_lightning.callbacks_factory": [
            # The format here must be [any name]=[module path]:[function name]
            "monitor_callbacks=factories:my_custom_callbacks_factory"
        ]
    },
)

The group name for the entry points is pytorch_lightning.callbacks_factory and it contains a list of strings that specify where to find the function within the package.

Now, if you pip install -e . this package, it will register the my_custom_callbacks_factory function and Lightning will automatically call it to collect the callbacks whenever you run the Trainer!

To unregister the factory, simply uninstall the package with pip uninstall “my-package”.


Callback API

Here is the full API of methods available in the Callback base class.

The Callback class is the base for all the callbacks in Lightning just like the LightningModule is the base for all models. It defines a public interface that each callback implementation must follow, the key ones are:

Properties

state_key

Callback.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.

Return type

str

Hooks

setup

Callback.setup(trainer, pl_module, stage)[source]

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

Return type

None

teardown

Callback.teardown(trainer, pl_module, stage)[source]

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

Return type

None

on_fit_start

Callback.on_fit_start(trainer, pl_module)[source]

Called when fit begins.

Return type

None

on_fit_end

Callback.on_fit_end(trainer, pl_module)[source]

Called when fit ends.

Return type

None

on_sanity_check_start

Callback.on_sanity_check_start(trainer, pl_module)[source]

Called when the validation sanity check starts.

Return type

None

on_sanity_check_end

Callback.on_sanity_check_end(trainer, pl_module)[source]

Called when the validation sanity check ends.

Return type

None

on_train_batch_start

Callback.on_train_batch_start(trainer, pl_module, batch, batch_idx)[source]

Called when the train batch begins.

Return type

None

on_train_batch_end

Callback.on_train_batch_end(trainer, pl_module, outputs, batch, batch_idx)[source]

Called when the train batch ends.

Note

The value outputs["loss"] here will be the normalized value w.r.t accumulate_grad_batches of the loss returned from training_step.

Return type

None

on_train_epoch_start

Callback.on_train_epoch_start(trainer, pl_module)[source]

Called when the train epoch begins.

Return type

None

on_train_epoch_end

Callback.on_train_epoch_end(trainer, pl_module)[source]

Called when the train epoch ends.

To access all batch outputs at the end of the epoch, either:

  1. Implement training_epoch_end in the LightningModule and access outputs via the module OR

  2. Cache data across train batch hooks inside the callback implementation to post-process in this hook.

Return type

None

on_validation_epoch_start

Callback.on_validation_epoch_start(trainer, pl_module)[source]

Called when the val epoch begins.

Return type

None

on_validation_epoch_end

Callback.on_validation_epoch_end(trainer, pl_module)[source]

Called when the val epoch ends.

Return type

None

on_test_epoch_start

Callback.on_test_epoch_start(trainer, pl_module)[source]

Called when the test epoch begins.

Return type

None

on_test_epoch_end

Callback.on_test_epoch_end(trainer, pl_module)[source]

Called when the test epoch ends.

Return type

None

on_predict_epoch_start

Callback.on_predict_epoch_start(trainer, pl_module)[source]

Called when the predict epoch begins.

Return type

None

on_predict_epoch_end

Callback.on_predict_epoch_end(trainer, pl_module, outputs)[source]

Called when the predict epoch ends.

Return type

None

on_validation_batch_start

Callback.on_validation_batch_start(trainer, pl_module, batch, batch_idx, dataloader_idx)[source]

Called when the validation batch begins.

Return type

None

on_validation_batch_end

Callback.on_validation_batch_end(trainer, pl_module, outputs, batch, batch_idx, dataloader_idx)[source]

Called when the validation batch ends.

Return type

None

on_test_batch_start

Callback.on_test_batch_start(trainer, pl_module, batch, batch_idx, dataloader_idx)[source]

Called when the test batch begins.

Return type

None

on_test_batch_end

Callback.on_test_batch_end(trainer, pl_module, outputs, batch, batch_idx, dataloader_idx)[source]

Called when the test batch ends.

Return type

None

on_predict_batch_start

Callback.on_predict_batch_start(trainer, pl_module, batch, batch_idx, dataloader_idx)[source]

Called when the predict batch begins.

Return type

None

on_predict_batch_end

Callback.on_predict_batch_end(trainer, pl_module, outputs, batch, batch_idx, dataloader_idx)[source]

Called when the predict batch ends.

Return type

None

on_train_start

Callback.on_train_start(trainer, pl_module)[source]

Called when the train begins.

Return type

None

on_train_end

Callback.on_train_end(trainer, pl_module)[source]

Called when the train ends.

Return type

None

on_validation_start

Callback.on_validation_start(trainer, pl_module)[source]

Called when the validation loop begins.

Return type

None

on_validation_end

Callback.on_validation_end(trainer, pl_module)[source]

Called when the validation loop ends.

Return type

None

on_test_start

Callback.on_test_start(trainer, pl_module)[source]

Called when the test begins.

Return type

None

on_test_end

Callback.on_test_end(trainer, pl_module)[source]

Called when the test ends.

Return type

None

on_predict_start

Callback.on_predict_start(trainer, pl_module)[source]

Called when the predict begins.

Return type

None

on_predict_end

Callback.on_predict_end(trainer, pl_module)[source]

Called when predict ends.

Return type

None

on_exception

Callback.on_exception(trainer, pl_module, exception)[source]

Called when any trainer execution is interrupted by an exception.

Return type

None

state_dict

Callback.state_dict()[source]

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

Return type

Dict[str, Any]

Returns

A dictionary containing callback state.

on_save_checkpoint

Callback.on_save_checkpoint(trainer, pl_module, checkpoint)[source]

Called when saving a checkpoint to give you a chance to store anything else you might want to save.

Parameters
Return type

None

load_state_dict

Callback.load_state_dict(state_dict)[source]

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

Parameters

state_dict (Dict[str, Any]) – the callback state returned by state_dict.

Return type

None

on_load_checkpoint

Callback.on_load_checkpoint(trainer, pl_module, checkpoint)[source]

Called when loading a model checkpoint, use to reload state.

Parameters
Return type

None

on_before_backward

Callback.on_before_backward(trainer, pl_module, loss)[source]

Called before loss.backward().

Return type

None

on_after_backward

Callback.on_after_backward(trainer, pl_module)[source]

Called after loss.backward() and before optimizers are stepped.

Return type

None

on_before_optimizer_step

Callback.on_before_optimizer_step(trainer, pl_module, optimizer, opt_idx)[source]

Called before optimizer.step().

Return type

None

on_before_zero_grad

Callback.on_before_zero_grad(trainer, pl_module, optimizer)[source]

Called before optimizer.zero_grad().

Return type

None