BYOL Single GPU implementation #1
Open
pranavsinghps1 wants to merge 38 commits intoiseessel:masterfrom
Open
Conversation
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
iseessel
added a commit
that referenced
this pull request
Jul 26, 2021
) Summary: Pull Request resolved: facebookresearch#343 Some basic changes to make this script work within FBinfra. 1. Register Manifold in PathManager. 1. In order to do #1, create fb/extra_scripts/convert_sharded_checkpoint.y and add necessary dependencies in TARGETS 1. Replace some torch.loads using PathManager. Reviewed By: prigoyal Differential Revision: D29166520 fbshipit-source-id: a61b4eb80d74526b0a7e2d38f973eb688b311a94
iseessel
requested changes
Jul 26, 2021
Author
|
As per the commit fc1217d addressed the following:
|
iseessel
reviewed
Aug 9, 2021
iseessel
reviewed
Aug 9, 2021
iseessel
reviewed
Aug 9, 2021
iseessel
reviewed
Aug 9, 2021
iseessel
reviewed
Aug 9, 2021
iseessel
reviewed
Aug 9, 2021
iseessel
reviewed
Aug 9, 2021
iseessel
reviewed
Aug 10, 2021
| def __init__(self, base_momentum: float, shuffle_batch: bool = True): | ||
| super().__init__() | ||
| self.base_momentum = base_momentum | ||
| self.is_distributed = False |
iseessel
reviewed
Aug 10, 2021
|
|
||
| class BYOLHook(ClassyHook): | ||
| """ | ||
| TODO: Update description |
iseessel
reviewed
Aug 10, 2021
|
|
||
| @staticmethod | ||
| def cosine_decay(training_iter, max_iters, initial_value): | ||
| # TODO: Why do we need this min statement? |
Owner
There was a problem hiding this comment.
Comment this method.
Put types in the method.
iseessel
reviewed
Aug 10, 2021
| return initial_value * cosine_decay_value | ||
|
|
||
| @staticmethod | ||
| def target_ema(training_iter, base_ema, max_iters): |
Owner
There was a problem hiding this comment.
Let's comment every method in the byol_hooks.py and byol_losses.py and make sure they all have type hints.
iseessel
reviewed
Aug 10, 2021
|
|
||
| def _build_byol_target_network(self, task: tasks.ClassyTask) -> None: | ||
| """ | ||
| Create the model replica called the target. This will slowly track |
Owner
There was a problem hiding this comment.
Improve comment. Something like: "Target network is exponential moving average of online network, ... "
iseessel
reviewed
Aug 10, 2021
| @torch.no_grad() | ||
| def on_forward(self, task: tasks.ClassyTask) -> None: | ||
| """ | ||
| - Update the target model. |
Owner
There was a problem hiding this comment.
Update comment for BYOL (this was copy/pasted from moco).
iseessel
reviewed
Aug 10, 2021
| @register_loss("byol_loss") | ||
| class BYOLLoss(ClassyLoss): | ||
| """ | ||
| TODO: change description |
iseessel
reviewed
Aug 10, 2021
| and https://github.com/facebookresearch/moco for a reference implementation, reused here | ||
|
|
||
| Config params: | ||
| embedding_dim (int): head output output dimension |
iseessel
reviewed
Aug 10, 2021
| "_BYOLLossConfig", ["embedding_dim", "momentum"] | ||
| ) | ||
|
|
||
| def regression_loss(x, y): |
Owner
There was a problem hiding this comment.
Type-hints + comments for all these functions.
iseessel
reviewed
Aug 10, 2021
| @classmethod | ||
| def from_config(cls, config: BYOLLossConfig): | ||
| """ | ||
| Instantiates BYOLLoss from configuration. |
Owner
There was a problem hiding this comment.
Put in the config options in the docstring here.
iseessel
reviewed
Aug 10, 2021
|
|
||
| def forward(self, online_network_prediction: torch.Tensor, *args, **kwargs) -> torch.Tensor: | ||
| """ | ||
| Given the encoder queries, the key and the queue of the previous queries, |
Owner
There was a problem hiding this comment.
This comment I think is copy/pasted.
iseessel
reviewed
Aug 10, 2021
| self.is_distributed = False | ||
|
|
||
| self.momentum = None | ||
| self.inv_momentum = None |
iseessel
reviewed
Aug 10, 2021
|
|
||
| self.momentum = None | ||
| self.inv_momentum = None | ||
| self.total_iters = None |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Implementation of BYOL: https://arxiv.org/abs/2006.07733 on Single GPU
Issue #190