Feat: Add BackboneLambdaFinetunningCallback #5377
Conversation
tchaton
requested review from
Borda,
justusschock,
SeanNaren and
williamFalcon
as code owners
|
Hello @tchaton! Thanks for updating this PR. There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻 Comment last updated at 2021-01-08 17:42:50 UTC |
tchaton
added
design
Codecov Report
@@ Coverage Diff @@
## release/1.2-dev #5377 +/- ##
================================================
- Coverage 93% 93% -0%
================================================
Files 150 151 +1
Lines 10490 10592 +102
================================================
+ Hits 9719 9808 +89
- Misses 771 784 +13 |
|
Looks good to me, I just would also expose the |
Co-authored-by: Justus Schock <12886177+justusschock@users.noreply.github.com>
Co-authored-by: Justus Schock <12886177+justusschock@users.noreply.github.com>
| def finetunning_function(self, pl_module: pl.LightningModule, epoch: int, optimizer: Optimizer, opt_idx: int): | ||
| if epoch == self.milestones[0]: | ||
| unfreeze_and_add_param_group( | ||
| module=pl_module.feature_extractor[-5:], |
There was a problem hiding this comment.
Should we add some info of what is actually be unfrozen here?
…hLightning/pytorch-lightning into feat/finetunning_callback
| class FinetunningBoringModel(BoringModel): | ||
| def __init__(self): | ||
| super().__init__() | ||
| self.backbone = nn.Sequential(nn.Linear(32, 32, bias=False), nn.BatchNorm1d(32), nn.ReLU()) |
There was a problem hiding this comment.
does it work without a backbone, with pure BoringModel?
There was a problem hiding this comment.
No, it expects a module backbone.
| @@ -27,6 +27,10 @@ class Callback(abc.ABC): | |||
| Subclass this class and override any of the relevant hooks | |||
| """ | |||
|
|
|||
| def on_before_accelerator_backend_setup(self, trainer, pl_module): | |||
There was a problem hiding this comment.
do we need to add a new callback hook? can't it be done with the existing ones?
There was a problem hiding this comment.
Great question !
We have on_init_start, on_init_end in init function. We don't have access to the model.
Then we have on_fit_start in fit function, but configure_optimizers had already being called during accelerator_backend.setup().
Therefore, I needed to introduce a new callback, so I could freeze the model before configure_optimizers call and then filter parameters if they don't require gradients with
optimizer = optim.Adam(filter(lambda p: p.requires_grad, self.parameters()), lr=self.lr)
Hope it answer your question.
There was a problem hiding this comment.
In that case, I think on_before_configure_optimizers would be a better name. Easier for users to understand.
Can you also mention the added hook in the CHANGELOG?
Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com>
Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com>
…hLightning/pytorch-lightning into feat/finetunning_callback
What does this PR do?
This PR adds a Fine-tuning Callback + a new callback
on_before_accelerator_backend_setup+ update fine-tuning example to use the new Callback. The example was broken and didn't converge.Fixes # (issue) <- this links related issue to this PR
Before submitting
PR review
Anyone in the community is free to review the PR once the tests have passed.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:
Did you have fun?
Make sure you had fun coding 🙃