Basic callbacks for Learner

For the actual fastai documentation, you should go to the Callbacks documentation. These are minimal docs simply to bring in the source code and related tests to ensure that minimal functionality is met

Events

Callbacks can occur at any of these times:: after_create before_fit before_epoch before_train before_batch after_pred after_loss before_backward before_step after_step after_cancel_batch after_batch after_cancel_train after_train before_validate after_cancel_validate after_validate after_cancel_epoch after_epoch after_cancel_fit after_fit.

class event[source]

event(*args, **kwargs)

All possible events as attributes to get tab-completion and typo-proofing

To ensure that you are referring to an event (that is, the name of one of the times when callbacks are called) that exists, and to get tab completion of event names, use event:

test_eq(event.before_step, 'before_step')

class Callback[source]

Callback(after_create=None, before_fit=None, before_epoch=None, before_train=None, before_batch=None, after_pred=None, after_loss=None, before_backward=None, before_step=None, after_cancel_step=None, after_step=None, after_cancel_batch=None, after_batch=None, after_cancel_train=None, after_train=None, before_validate=None, after_cancel_validate=None, after_validate=None, after_cancel_epoch=None, after_epoch=None, after_cancel_fit=None, after_fit=None) :: Stateful

Basic class handling tweaks of the training loop by changing a Learner in various events

The training loop is defined in Learner a bit below and consists in a minimal set of instructions: looping through the data we:

  • compute the output of the model from the input
  • calculate a loss between this output and the desired target
  • compute the gradients of this loss with respect to all the model parameters
  • update the parameters accordingly
  • zero all the gradients

Any tweak of this training loop is defined in a Callback to avoid over-complicating the code of the training loop, and to make it easy to mix and match different techniques (since they'll be defined in different callbacks). A callback can implement actions on the following events:

  • after_create: called after the Learner is created
  • before_fit: called before starting training or inference, ideal for initial setup.
  • before_epoch: called at the beginning of each epoch, useful for any behavior you need to reset at each epoch.
  • before_train: called at the beginning of the training part of an epoch.
  • before_batch: called at the beginning of each batch, just after drawing said batch. It can be used to do any setup necessary for the batch (like hyper-parameter scheduling) or to change the input/target before it goes in the model (change of the input with techniques like mixup for instance).
  • after_pred: called after computing the output of the model on the batch. It can be used to change that output before it's fed to the loss.
  • after_loss: called after the loss has been computed, but before the backward pass. It can be used to add any penalty to the loss (AR or TAR in RNN training for instance).
  • before_backward: called after the loss has been computed, but only in training mode (i.e. when the backward pass will be used)
  • before_step: called after the backward pass, but before the update of the parameters. It can be used to do any change to the gradients before said update (gradient clipping for instance).
  • after_step: called after the step and before the gradients are zeroed.
  • after_batch: called at the end of a batch, for any clean-up before the next one.
  • after_train: called at the end of the training phase of an epoch.
  • before_validate: called at the beginning of the validation phase of an epoch, useful for any setup needed specifically for validation.
  • after_validate: called at the end of the validation part of an epoch.
  • after_epoch: called at the end of an epoch, for any clean-up before the next one.
  • after_fit: called at the end of training, for final clean-up.

Callback.__call__[source]

Callback.__call__(event_name)

Call self.{event_name} if it's defined

One way to define callbacks is through subclassing:

class _T(Callback):
    def call_me(self): return "maybe"
test_eq(_T()("call_me"), "maybe")

Another way is by passing the callback function to the constructor:

def cb(self): return "maybe"
_t = Callback(before_fit=cb)
test_eq(_t(event.before_fit), "maybe")

Callbacks provide a shortcut to avoid having to write self.learn.bla for any bla attribute we seek; instead, just write self.bla. This only works for getting attributes, not for setting them.

mk_class('TstLearner', 'a')

class TstCallback(Callback):
    def batch_begin(self): print(self.a)

learn,cb = TstLearner(1),TstCallback()
cb.learn = learn
test_stdout(lambda: cb('batch_begin'), "1")

If you want to change the value of an attribute, you have to use self.learn.bla, no self.bla. In the example below, self.a += 1 creates an a attribute of 2 in the callback instead of setting the a of the learner to 2. It also issues a warning that something is probably wrong:

learn.a
1
class TstCallback(Callback):
    def batch_begin(self): self.a += 1

learn,cb = TstLearner(1),TstCallback()
cb.learn = learn
cb('batch_begin')
test_eq(cb.a, 2)
test_eq(cb.learn.a, 1)
/mnt/d/lib/python3.7/site-packages/ipykernel_launcher.py:22: UserWarning: You are shadowing an attribute (a) that exists in the learner. Use `self.learn.a` to avoid this

A proper version needs to write self.learn.a = self.a + 1:

class TstCallback(Callback):
    def batch_begin(self): self.learn.a = self.a + 1

learn,cb = TstLearner(1),TstCallback()
cb.learn = learn
cb('batch_begin')
test_eq(cb.learn.a, 2)

Callback.name[source]

Name of the Callback, camel-cased and with 'Callback' removed

test_eq(TstCallback().name, 'tst')
class ComplicatedNameCallback(Callback): pass
test_eq(ComplicatedNameCallback().name, 'complicated_name')

class TrainEvalCallback[source]

TrainEvalCallback(after_create=None, before_fit=None, before_epoch=None, before_train=None, before_batch=None, after_pred=None, after_loss=None, before_backward=None, before_step=None, after_cancel_step=None, after_step=None, after_cancel_batch=None, after_batch=None, after_cancel_train=None, after_train=None, before_validate=None, after_cancel_validate=None, after_validate=None, after_cancel_epoch=None, after_epoch=None, after_cancel_fit=None, after_fit=None) :: Callback

Callback that tracks the number of iterations done and properly sets training/eval mode

This Callback is automatically added in every Learner at initialization.

Attributes available to callbacks

When writing a callback, the following attributes of Learner are available:

  • model: the model used for training/validation
  • data: the underlying DataLoaders
  • loss_func: the loss function used
  • opt: the optimizer used to update the model parameters
  • opt_func: the function used to create the optimizer
  • cbs: the list containing all Callbacks
  • dl: current DataLoader used for iteration
  • x/xb: last input drawn from self.dl (potentially modified by callbacks). xb is always a tuple (potentially with one element) and x is detuplified. You can only assign to xb.
  • y/yb: last target drawn from self.dl (potentially modified by callbacks). yb is always a tuple (potentially with one element) and y is detuplified. You can only assign to yb.
  • pred: last predictions from self.model (potentially modified by callbacks)
  • loss: last computed loss (potentially modified by callbacks)
  • n_epoch: the number of epochs in this training
  • n_iter: the number of iterations in the current self.dl
  • epoch: the current epoch index (from 0 to n_epoch-1)
  • iter: the current iteration index in self.dl (from 0 to n_iter-1)

The following attributes are added by TrainEvalCallback and should be available unless you went out of your way to remove that callback:

  • train_iter: the number of training iterations done since the beginning of this training
  • pct_train: from 0. to 1., the percentage of training iterations completed
  • training: flag to indicate if we're in training mode or not

The following attribute is added by Recorder and should be available unless you went out of your way to remove that callback:

  • smooth_loss: an exponentially-averaged version of the training loss

Callbacks control flow

It happens that we may want to skip some of the steps of the training loop: in gradient accumulation, we don't always want to do the step/zeroing of the grads for instance. During an LR finder test, we don't want to do the validation phase of an epoch. Or if we're training with a strategy of early stopping, we want to be able to completely interrupt the training loop.

This is made possible by raising specific exceptions the training loop will look for (and properly catch).

class CancelStepException[source]

CancelStepException(*args, **kwargs) :: Exception

Skip stepping the optimizer

class CancelBatchException[source]

CancelBatchException(*args, **kwargs) :: Exception

Skip the rest of this batch and go to after_batch

class CancelTrainException[source]

CancelTrainException(*args, **kwargs) :: Exception

Skip the rest of the training part of the epoch and go to after_train

class CancelValidException[source]

CancelValidException(*args, **kwargs) :: Exception

Skip the rest of the validation part of the epoch and go to after_validate

class CancelEpochException[source]

CancelEpochException(*args, **kwargs) :: Exception

Skip the rest of this epoch and go to after_epoch

class CancelFitException[source]

CancelFitException(*args, **kwargs) :: Exception

Interrupts training and go to after_fit

You can detect one of those exceptions occurred and add code that executes right after with the following events:

  • after_cancel_batch: reached immediately after a CancelBatchException before proceeding to after_batch
  • after_cancel_train: reached immediately after a CancelTrainException before proceeding to after_epoch
  • after_cancel_valid: reached immediately after a CancelValidException before proceeding to after_epoch
  • after_cancel_epoch: reached immediately after a CancelEpochException before proceeding to after_epoch
  • after_cancel_fit: reached immediately after a CancelFitException before proceeding to after_fit

class GatherPredsCallback[source]

GatherPredsCallback(with_input=False, with_loss=False, save_preds=None, save_targs=None, concat_dim=0) :: Callback

Callback that saves the predictions and targets, optionally with_loss

class FetchPredsCallback[source]

FetchPredsCallback(ds_idx=1, dl=None, with_input=False, with_decoded=False, cbs=None, reorder=True) :: Callback

A callback to fetch predictions during the training loop