How to override fit_one_cycle method?

keeplearning · April 13, 2020, 9:09pm

I want to override the fit_one_cycle method of the Learner class. When I go through the code in fastai/basic_train.py, I don’t find that method. I see that fit_one_cycle is implemented in fastai/train.py, but I don’t understand how it is linked to the Learner class when I create a Learner object, because it is not defined within the class itself. Thanks in advance.

morgan · April 14, 2020, 10:13am

What would you like to do instead of fit_one_cycle? Is it .fit you are looking for?

dougforrest · April 14, 2020, 10:30am

fit_one_cycle is defined in train.py and the function is assigned to the Learn class on line 67.

github.com

fastai/fastai/blob/master/fastai/train.py#L67


    for cb in learn.callbacks:
        if isinstance(cb, MixedPrecision): learn.callbacks.remove(cb)
    learn.model = learn.model.float()
    return learn


def mixup(learn:Learner, alpha:float=0.4, stack_x:bool=False, stack_y:bool=True) -> Learner:
    "Add mixup https://arxiv.org/abs/1710.09412 to `learn`."
    learn.callback_fns.append(partial(MixUpCallback, alpha=alpha, stack_x=stack_x, stack_y=stack_y))
    return learn


Learner.fit_one_cycle = fit_one_cycle
Learner.lr_find = lr_find
Learner.to_fp16 = to_fp16
Learner.to_fp32 = to_fp32
Learner.mixup = mixup
Learner.fit_fc = fit_fc


class ShowGraph(LearnerCallback):
    "Update a graph of learner stats and metrics after each epoch."
    def on_epoch_end(self, n_epochs:int, last_metrics:MetricsList, **kwargs)->bool:
        "If we have `last_metrics` plot them in our pbar graph"