Callbacks implemented in the fastai library

List of callbacks

fastai's training loop is highly extensible, with a rich callback system. See the callback docs if you're interested in writing your own callback. See below for a list of callbacks that are provided with fastai, grouped by the module they're defined in.

Every callback that is passed to Learner with the callback_fns parameter will be automatically stored as an attribute. The attribute name is snake-cased, so for instance ActivationStats will appear as learn.activation_stats (assuming your object is named learn).

Callback

This sub-package contains more sophisticated callbacks that each are in their own module. They are (click the link for more details):

LRFinder

Use Leslie Smith's learning rate finder to find a good learning rate for training your model. Let's see an example of use on the MNIST dataset with a simple CNN.

path = untar_data(URLs.MNIST_SAMPLE)
data = ImageDataBunch.from_folder(path)
def simple_learner(): return Learner(data, simple_cnn((3,16,16,2)), metrics=[accuracy])
learn = simple_learner()

The fastai librairy already has a Learner method called lr_find that uses LRFinder to plot the loss as a function of the learning rate

learn.lr_find()

LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.

learn.recorder.plot()

In this example, a learning rate around 2e-2 seems like the right fit.

lr = 2e-2

OneCycleScheduler

Train with Leslie Smith's 1cycle annealing method. Let's train our simple learner using the one cycle policy.

learn.fit_one_cycle(3, lr)
Total time: 00:07

epoch train_loss valid_loss accuracy time
0 0.109439 0.059349 0.980864 00:02
1 0.039582 0.023152 0.992149 00:02
2 0.019009 0.021239 0.991659 00:02

The learning rate and the momentum were changed during the epochs as follows (more info on the dedicated documentation page).

learn.recorder.plot_lr(show_moms=True)

MixUpCallback

Data augmentation using the method from mixup: Beyond Empirical Risk Minimization. It is very simple to add mixup in fastai :

learn = Learner(data, simple_cnn((3, 16, 16, 2)), metrics=[accuracy]).mixup()

CSVLogger

Log the results of training in a csv file. Simply pass the CSVLogger callback to the Learner.

learn = Learner(data, simple_cnn((3, 16, 16, 2)), metrics=[accuracy, error_rate], callback_fns=[CSVLogger])

learn.fit(3)
Total time: 00:07

epoch train_loss valid_loss accuracy error_rate time
0 0.127259 0.098069 0.969578 0.030422 00:02
1 0.084601 0.068024 0.974975 0.025025 00:02
2 0.055074 0.047266 0.983317 0.016683 00:02

You can then read the csv.

learn.csv_logger.read_logged_file()
epoch train_loss valid_loss accuracy error_rate
0 0 0.127259 0.098069 0.969578 0.030422
1 1 0.084601 0.068024 0.974975 0.025025
2 2 0.055074 0.047266 0.983317 0.016683

GeneralScheduler

Create your own multi-stage annealing schemes with a convenient API. To illustrate, let's implement a 2 phase schedule.

def fit_odd_shedule(learn, lr):
    n = len(learn.data.train_dl)
    phases = [TrainingPhase(n).schedule_hp('lr', lr, anneal=annealing_cos), 
              TrainingPhase(n*2).schedule_hp('lr', lr, anneal=annealing_poly(2))]
    sched = GeneralScheduler(learn, phases)
    learn.callbacks.append(sched)
    total_epochs = 3
    learn.fit(total_epochs)

learn = Learner(data, simple_cnn((3,16,16,2)), metrics=accuracy)
fit_odd_shedule(learn, 1e-3)
Total time: 00:07

epoch train_loss valid_loss accuracy time
0 0.176607 0.157229 0.946025 00:02
1 0.140903 0.133690 0.954367 00:02
2 0.130910 0.131156 0.956820 00:02 
learn.recorder.plot_lr()

MixedPrecision

Use fp16 to take advantage of tensor cores on recent NVIDIA GPUs for a 200% or more speedup.

HookCallback

Convenient wrapper for registering and automatically deregistering PyTorch hooks. Also contains pre-defined hook callback: ActivationStats.

RNNTrainer

Callback taking care of all the tweaks to train an RNN.

TerminateOnNaNCallback

Stop training if the loss reaches NaN.

EarlyStoppingCallback

Stop training if a given metric/validation loss doesn't improve.

SaveModelCallback

Save the model at every epoch, or the best model for a given metric/validation loss.

learn = Learner(data, simple_cnn((3,16,16,2)), metrics=accuracy)
learn.fit_one_cycle(3,1e-4, callbacks=[SaveModelCallback(learn, every='epoch', monitor='accuracy')])
Total time: 00:07

epoch train_loss valid_loss accuracy time
0 0.679189 0.646599 0.804220 00:02
1 0.527475 0.497290 0.908243 00:02
2 0.464756 0.462471 0.917076 00:02 
!ls ~/.fastai/data/mnist_sample/models

best.pth	 bestmodel_2.pth  model_1.pth  model_4.pth    stage-1.pth
bestmodel_0.pth  bestmodel_3.pth  model_2.pth  model_5.pth    tmp.pth
bestmodel_1.pth  model_0.pth	  model_3.pth  one_epoch.pth  trained_model.pth

ReduceLROnPlateauCallback

Reduce the learning rate each time a given metric/validation loss doesn't improve by a certain factor.

PeakMemMetric

GPU and general RAM profiling callback

StopAfterNBatches

Stop training after n batches of the first epoch.

LearnerTensorboardWriter

Broadly useful callback for Learners that writes to Tensorboard. Writes model histograms, losses/metrics, embedding projector and gradient stats.

train and basic_train

Recorder

Track per-batch and per-epoch smoothed losses and metrics.

ShowGraph

Dynamically display a learning chart during training.

BnFreeze

Freeze batchnorm layer moving average statistics for non-trainable layers.

GradientClipping

Clips gradient during training.