Train 2 epochs head, unfreeze / learning rate finder, continue training (fit_one_cycle)

In a transfer learning setting, I want to freeze the body and only train the head for 2 epochs. Then I want to unfreeze the whole network and use the Learning Rate finder, before continue training again.

What I want to do is similar to FastAI’s fit_one_cycle.
To do the same with PyTorch Lightning, I tried the following:

Trainer(max_epochs=2, min_epochs=0, auto_lr_find=True), data_module)  # FastAI: learn.fit_one_cycle(2)

trainer.max_epochs = 5
# model.unfreeze()  # allow the whole body to be trained
# trainer.tune(model)  # LR finder, data_module)  # FastAI: learn.fit_one_cycle(3)

Unfortunately, this would invoke the training op epoch 1 twice.

Any ideas of how to approach what I want to do?

Edit 1: Link to Colab demonstrating 2x epoch 1:

Forwarded from slack:

The first fit doesn’t advance self.trainer.current_epoch at the end of training, try doing +=1 in between

@s-rog Thank you! This seems to indeed solve the epoch logging issue.

Now I’m trying to invoke the trainer.tune(model) after the training for the first 2 epochs.
However, this fails:

LR finder stopped early due to diverging loss.
Failed to compute suggesting for `lr`. There might not be enough points.
Traceback (most recent call last):
  File "/usr/local/lib/python3.6/dist-packages/pytorch_lightning/tuner/", line 340, in suggestion
    min_grad = np.gradient(loss).argmin()
  File "<__array_function__ internals>", line 6, in gradient
  File "/usr/local/lib/python3.6/dist-packages/numpy/lib/", line 1042, in gradient
    "Shape of array too small to calculate a numerical gradient, "
ValueError: Shape of array too small to calculate a numerical gradient, at least (edge_order + 1) elements are required.

Colab showing this issue:

cc @SkafteNicki.

Not sure we tried using the LR finder after training like this, so we likely need to make tweaks on our end.

But second, depending on your model freezing and unfreezing may not make a giant difference.

Take a look at our video that goes over this:

You can also see a finetuning colab here where we show no real difference in performance for this model (in fact, unfreezing first converges faster).

1 Like

I can confirm that the learning rate finder have never been tested in such a way, but I completely agree that we should probably support it.
I was planning on doing a bit of refactors of the tuning interface within the next month, that will hopefully solve this problem.

1 Like

I don’t think assigning max_epoch = 5 like this will work because Trainer state is not reset here for eg global_step, loggers, etc. Try creating the trainer instance again with max_epochs=5, reload the model weights if required after the first fit cycle and then continue with .tune

While the freezing was specific to my use-case, the lr-finder issue also appears without freezing involved.

I was also able to reproduce the issue by wrapping lr_finder in a Callback (which might be easier to test against @SkafteNicki ):

Since this seems to be an issue (or feature request), I made a bug report:

@williamfalcon Thank you for the video and Colab. I will take a look at it. That freezing for an epoch or 2 helps for Transfer Learning came from the FastAI 2 cours . I’ll check which approach works better in combination with lr-finder once I get it to work.

@goku Thank you for the suggestion. I haven’t tried it yet, but it sounds like a workaround that would work.