fastai provides a powerful callback system, which is documented on the
callbacks page; look on that page if you're just looking for how to use existing callbacks. If you want to create your own, you'll need to use the classes discussed below.
A key motivation for the callback system is that additional functionality can be entirely implemented in a single callback, so that it's easily read. By using this trick, we will have different methods categorized in different callbacks where we will find clearly stated all the interventions the method makes in training. For instance in the
LRFinder callback, on top of running the fit function with exponentially growing LRs, it needs to handle some preparation and clean-up, and all this code can be in the same callback so we know exactly what it is doing and where to look if we need to change something.
In addition, it allows our
fit function to be very clean and simple, yet still easily extended. So far in implementing a number of recent papers, we haven't yet come across any situation where we had to modify our training loop source code - we've been able to use callbacks every time.
Base class for callbacks that want to record values, dynamically change learner params, etc.
To create a new type of callback, you'll need to inherit from this class, and implement one or more methods as required for your purposes. Perhaps the easiest way to get started is to look at the source code for some of the pre-defined fastai callbacks. You might be surprised at how simple they are! For instance, here is the entire source code for
@dataclass class GradientClipping(LearnerCallback): clip:float def on_backward_end(self, **kwargs): if self.clip: nn.utils.clip_grad_norm_(self.learn.model.parameters(), self.clip)
You generally want your custom callback constructor to take a
Learner parameter, e.g.:
@dataclass class MyCallback(Callback): learn:Learner
Note that this allows the callback user to just pass your callback name to
callback_fns when constructing their
Learner, since that always passes
self when constructing callbacks from
callback_fns. In addition, by passing the learner, this callback will have access to everything: e.g all the inputs/outputs as they are calculated, the losses, and also the data loaders, the optimizer, etc. At any time:
- Changing self.learn.data.train_dl or self.data.valid_dl will change them inside the fit function (we just need to pass the
DataBunchobject to the fit function and not data.train_dl/data.valid_dl)
- Changing self.learn.opt.opt (We have an
OptimWrapperon top of the actual optimizer) will change it inside the fit function.
- Changing self.learn.data or self.learn.opt directly WILL NOT change the data or the optimizer inside the fit function.
In any of the callbacks you can unpack in the kwargs:
n_epochs, contains the number of epochs the training will take in total
epoch, contains the number of the current
iteration, contains the number of iterations done since the beginning of training
num_batch, contains the number of the batch we're at in the dataloader
last_input, contains the last input that got through the model (eventually updated by a callback)
last_target, contains the last target that got through the model (eventually updated by a callback)
last_output, contains the last output spitted by the model (eventually updated by a callback)
last_loss, contains the last loss computed (eventually updated by a callback)
smooth_loss, contains the smoothed version of the loss
last_metrics, contains the last validation loss and metrics computed
pbar, the progress bar
train, flag to know if we're in training mode or not
stop_training, that will stop the training at the end of the current epoch if True
stop_epoch, that will break the current epoch loop
skip_step, that will skip the next optimizer step
skip_zero, that will skip the next zero grad
All of these methods are optional; your subclass can handle as many or as few as you require.
Here we can initiliaze anything we need. The optimizer has now been initialized. We can change any hyper-parameters by typing, for instance:
self.opt.lr = new_lr self.opt.mom = new_mom self.opt.wd = new_wd self.opt.beta = new_beta
This is not technically required since we have
on_train_begin for epoch 0 and
on_epoch_end for all the other epochs,
yet it makes writing code that needs to be done at the beginning of every epoch easy and more readable.
Here is the perfect place to prepare everything before the model is called. Example: change the values of the hyperparameters (if we don't do it on_batch_end instead)
At the end of that event
yb will be set to
last_target of the state of the
Here is the place to run some code that needs to be executed after the output has been computed but before the loss computation. Example: putting the output back in FP32 when training in mixed precision.
At the end of that event the output will be set to
last_output of the state of the
Here is the place to run some code that needs to be executed after the loss has been computed but before the gradient computation.
reg_fn in RNNs.
At the end of that event the output will be set to
last_loss of the state of the
Here is the place to run some code that needs to be executed after the gradients have been computed but before the optimizer is called.
True at the end of this event, the optimizer step is skipped.
Here is the place to run some code that needs to be executed after the optimizer step but before the gradients are zeroed.
True at the end of this event, the gradients are not zeroed.
Here is the place to run some code that needs to be executed after a batch is fully done. Example: change the values of the hyperparameters (if we don't do it on_batch_begin instead)
True at the end of this event, the current epoch is interrupted (example: lr_finder stops the training when the loss explodes).
Here is the place to run some code that needs to be executed at the end of an epoch. Example: Save the model if we have a new best validation loss/metric.
True at the end of this event, the training stops (example: early stopping).
Here is the place to tidy everything. It's always executed even if there was an error during the training loop, and has an extra kwarg named exception to check if there was an exception or not. Examples: save log_files, load best model found during training
The following functions provide different annealing schedules. You probably won't need to call them directly, but would instead use them as part of a callback. Here's what each one looks like:
Manage all of the registered
metrics, smoothing loss by momentum
You probably won't need to use this class yourself. It's used by fastai to combine all the callbacks together and call any relevant callback functions for each training stage. The methods below simply call the equivalent method in each callback function in
This is a convenience class that provides a consistent API for getting and setting optimizer hyperparameters. For instance, for
optim.Adam the momentum parameter is actually
betas, whereas for
optim.SGD it's simply
momentum. As another example, the details of handling weight decay depend on whether you are using
true_wd or the traditional L2 regularization approach.
This class also handles setting different WD and LR for each layer group, for discriminative layer training.
lr. Set lr on
You don't call these yourself - they're called by fastai's
Callback system automatically to enable the class's functionality.