Useful metrics for training

Training metrics

Metrics for training fastai models are simply functions that take input and target tensors, and return some metric of interest for training. You can write your own metrics by defining a function of that type, and passing it to Learner in the [code]metrics[/code] parameter, or use one of the following pre-defined functions.

Predefined metrics:


accuracy(`input`:Tensor, `targs`:Tensor) → Rank0Tensor

Compute accuracy with targs when input is bs * n_classes.


accuracy_thresh(`y_pred`:Tensor, `y_true`:Tensor, `thresh`:float=`0.5`, `sigmoid`:bool=`True`) → Rank0Tensor

Compute accuracy when y_pred and y_true are the same size.

Prediction are compared to thresh after sigmoid is maybe applied. Then we count the numbers that match the targets.


dice(`input`:Tensor, `targs`:Tensor, `iou`:bool=`False`) → Rank0Tensor

Dice coefficient metric for binary target. If iou=True, returns iou metric, classic for segmentation problems.


error_rate(`input`:Tensor, `targs`:Tensor) → Rank0Tensor

1 - accuracy


exp_rmspe(`pred`:Tensor, `targ`:Tensor) → Rank0Tensor

Exp RMSE between pred and targ.


fbeta(`y_pred`:Tensor, `y_true`:Tensor, `thresh`:float=`0.2`, `beta`:float=`2`, `eps`:float=`1e-09`, `sigmoid`:bool=`True`) → Rank0Tensor

Computes the f_beta between preds and targets

beta determines the value of the fbeta applied, eps is there for numeric stability. If sigmoid=True, a sigmoid is applied to the predictions before comparing them to thresh then to the targets. See the F1 score wikipedia page for details on the fbeta score.

class FBeta[source]

FBeta(`average`:Optional[str]=`'binary'`, `pos_label`:int=`1`, `eps`:float=`1e-09`, `beta`:float=`2`) :: CMScores

FBeta(average: Union[str, NoneType] = 'binary', pos_label: int = 1, eps: float = 1e-09, beta: float = 2)

Creating your own metric

Creating a new metric can be as simple as creating a new function. If your metric is an average over the total number of elements in your dataset, just write the function that will compute it on a batch (taking pred and targ as arguments). It will then be automatically averaged over the batches (taking their different sizes into acount).

Sometimes metrics aren't simple averages however. If we take the example of precision for instance, we have to divide the number of true positives by the number of predictions we made for that class. This isn't an average over the number of elements we have in the dataset, we only consider those where we made a positive prediction for a specific thing. Computing the precision for each batch, then averaging them will yield to a result that may be close to the real value, but won't be it exactly (and it really depends on how you deal with special case of 0 positive predictions).

This why in fastai, every metric is implemented as a callback. If you pass a regular function, the library transforms it to a proper callback called AverageCallback. The callback metrics are only called during the validation phase, and only for the following events:

  • on_epoch_begin (for initialization)
  • on_batch_begin (if we need to have a look at the input/target and maybe modify them)
  • on_batch_end (to analyze the last results and update our computation)
  • on_epoch_end(to wrap up the final result that should be stored in .metric)

As an example, the following code is the exact implementation of the AverageMetric callback that transforms a function like accuracy into a metric callback.

class AverageMetric(Callback):
    def __init__(self, func):
        self.func, = func, func.__name__

    def on_epoch_begin(self, **kwargs):
        self.val, self.count = 0.,0

    def on_batch_end(self, last_output, last_target, train, **kwargs):
        self.count += last_target.size(0)
        self.val += last_target.size(0) * self.func(last_output, last_target).detach().item()

    def on_epoch_end(self, **kwargs):
        self.metric = self.val/self.count

And here is another example that properly computes the precision for a given class.

class Precision(Callback):
    def on_epoch_begin(self, **kwargs):
        self.correct, = 0, 0
    def on_batch_end(self, last_output, last_target, **kwargs):
        preds = last_output.argmax(1)
        self.correct += ((preds==0) * (last_target==0)).float().sum() += (preds==0).float().sum()
    def on_epoch_end(self, **kwargs):
        self.metric = self.correct/

The following custom callback class example measures peak RAM usage during each epoch:

import tracemalloc
class TraceMallocMetric(Callback):
    def __init__(self):
        super().__init__() = "peak RAM"

    def on_epoch_begin(self, **kwargs):
    def on_epoch_end(self, **kwargs):
        current, peak =  tracemalloc.get_traced_memory()
        self.metric = torch.tensor(peak)

To deploy it, you need to pass an instance of this custom metric in the metrics argument:

learn = create_cnn(data, model, metrics=[accuracy, TraceMallocMetric()])
learn.fit_one_cycle(3, max_lr=1e-2)

And then the output changes to:

Total time: 00:54
epoch   train_loss  valid_loss  accuracy    peak RAM
   1    0.333352    0.084342    0.973800    2395541.000000
   2    0.096196    0.038386    0.988300    2342145.000000
   3    0.048722    0.029234    0.990200    2342680.000000

As mentioner earlier, using the metrics argument with a custom metrics class is limited in the number of phases of the callback system it can access, it can only return one numerical value and as you can see its output is hardcoded to have 6 points of precision in the output, even if the number is an int.

To overcome these limitations callback classes should be used instead.

For example, the following class:

  • uses phases not available for the metric classes
  • it reports 3 columns, instead of just one
  • its column report ints, instead of floats
import tracemalloc
class TraceMallocMultiColMetric(LearnerCallback):
    _order=-20 # Needs to run before the recorder
    def __init__(self, learn):
        self.train_max = 0

    def on_train_begin(self, **kwargs):
        self.learn.recorder.add_metric_names(['used', 'max_used', 'peak'])
    def on_batch_end(self, train, **kwargs):
        # track max memory usage during the train phase
        if train:
            current, peak =  tracemalloc.get_traced_memory()
            self.train_max = max(self.train_max, current)
    def on_epoch_begin(self, **kwargs):

    def on_epoch_end(self, **kwargs):
        current, peak =  tracemalloc.get_traced_memory()
        self.learn.recorder.add_metrics([current, self.train_max, peak])

Note, that it subclasses LearnerCallback and not Callback, since the former provides extra features not available in the latter.

Also _order=-20 is crucial - without it the custom columns will not be added - it tells the callback system to run this callback before the recorder system.

To deploy it, you need to pass the name of the class (not an instance!) of the class in the callback_fns argument. This is because the learn object doesn't exist yet, and it's required to instantiate TraceMallocMultiColMetric. The system will do it for us automatically as soon as the learn object has been created.

learn = create_cnn(data, model, metrics=[accuracy], callback_fns=TraceMallocMultiColMetric)
learn.fit_one_cycle(3, max_lr=1e-2)

And then the output changes to:

Total time: 00:53
epoch   train_loss valid_loss   accuracy     used   max_used   peak
    1   0.321233    0.068252    0.978600    156504  2408404   2419891 
    2   0.093551    0.032776    0.988500     79343  2408404   2348085
    3   0.047178    0.025307    0.992100     79568  2408404   2342754

Another way to do the same is by using learn.callbacks.append, and this time we need to instantiate TraceMallocMultiColMetric with learn object which we now have, as it is called after the latter was created:

learn = create_cnn(data, model, metrics=[accuracy])
learn.fit_one_cycle(3, max_lr=1e-2)

Configuring the custom metrics in the learn object sets them to run in all future fit-family calls. However, if you'd like to configure it for just one call, you can configure it directly inside fit or fit_one_cycle:

learn = create_cnn(data, model, metrics=[accuracy])
learn.fit_one_cycle(3, max_lr=1e-2, callbacks=TraceMallocMultiColMetric(learn))

And to stress the differences:

  • the callback_fns argument expects a classname or a list of those
  • the callbacks argument expects an instance of a class or a list of those
  • learn.callbacks.append expects a single instance of a class

For more examples, look inside fastai codebase and its test suite, search for classes that subclass either Callback, LearnerCallback and subclasses of those two.

Finally, while the above examples all add to the metrics, it's not a requirement. A callback can do anything it wants and it is not required to add its outcomes to the metrics printout.