Confusion about cycle_len

In lots of dl1 codes, I saw lots of cycle_len parameter in call.
For example:, 3, cycle_len=1, cycle_mult=2, wds=wd)

while in /fastai/, I don’t see it is defined in
def fit(self, epochs: int, lr: Union[Floats, slice]=default_lr,
wd: Floats=None, callbacks: Collection[Callback]=None) -> None:

Even in the whole, I don’t see cycle_len

Very interesting! I didn’t think of digging into the source; I only tried different combinations of cycle, cycle_len and cycle_mult to see the relation. Now that you inspired me, I did a grep and found in fastai-master/old/fastai/
def fit_gen(self, model, data, layer_opt, n_cycle, cycle_len=None, cycle_mult=1, cycle_save_name=None, best_save_name=None,
use_clr=None, use_clr_beta=None, metrics=None, callbacks=None, use_wd_sched=False, norm_wds=False,
wds_sched_mult=None, use_swa=False, swa_start=1, swa_eval_freq=5, **kwargs):

    """Method does some preparation before finally delegating to the 'fit' method for
    fitting the model. Namely, if cycle_len is defined, it adds a 'Cosine Annealing'
    scheduler for varying the learning rate across iterations.

    Method also computes the total number of epochs to fit based on provided 'cycle_len',
    'cycle_mult', and 'n_cycle' parameters.

        model (Learner):  Any neural architecture for solving a supported problem.
            Eg. ResNet-34, RNN_Learner etc.

        data (ModelData): An instance of ModelData.

        layer_opt (LayerOptimizer): An instance of the LayerOptimizer class

        n_cycle (int): number of cycles

        cycle_len (int):  number of epochs before lr is reset to the initial value.
            E.g if cycle_len = 3, then the lr is varied between a maximum
            and minimum value over 3 epochs.

        cycle_mult (int): additional parameter for influencing how the lr resets over
            the cycles. For an intuitive explanation, please see

under the old folder, means this is not been used anymore.
now the fit() goes to fastai/

cycle_len appears in


but it doesn’t appear in


If isn’t used anymore, the easiest test would be to move it away and see if your notebook runs…

I believe old isn’t obsolete. I believe fastai-master/old/fastai/*.py are in use in (Clouderizer) and 0.7.x (Paperspace) – as opposed to fastai-v3 (Clouderizer) and 1.0 / PyTorch 1.0 BETA (Paperspace)

thank you very much, now I know, it is symbol link:
./courses/dl1/fastai -> …/…/old/fastai
so that old/fastai/ is in use,
the not in use one is