Aioli Algorithm for Dynamic Mixtures #158
Open
Conversation
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
✅ Result of Pytest Coverage---------- coverage: platform linux, python 3.10.0-final-0 -----------
|
MaxiBoether
requested changes
Dec 18, 2024
Contributor
MaxiBoether
left a comment
MaxiBoether
left a comment
There was a problem hiding this comment.
Thank you, Beste. I checked the code on the level of understanding of the algorithm that I have. You are the expert on the actual logic. I just gave some small nitpicky remarks :) After fixing them and fixing CI, we can merge this and you can build on top of it. I think it would be good to test this with an actual training, to see how it behaves
Comment on lines
92
to
101
|
|
||
| Args: | ||
| losses: A numpy array of losses per domain. | ||
| counts: A numpy array of counts per domain. | ||
| """ | ||
| num_incoming_domains = len(losses) | ||
| num_internal_domains = len(self.losses) | ||
| num_domains = max(num_incoming_domains, num_internal_domains) | ||
|
|
||
| if num_internal_domains < num_domains: |
Contributor
There was a problem hiding this comment.
Is this doing anything more than the parent class? Similar to ADO, we definitely want a super() call here and only do additional work if necessary
mixtera/core/algo/aioli/aioli.py
Outdated
| self.aioli_diagonal = aioli_diagonal | ||
| self.domain_count = len(self.losses) | ||
| self.graph = np.zeros((self.domain_count, self.domain_count)) | ||
| self.weights: np.ndarray = None # The latest weights after update steps have been completed. |
Contributor
There was a problem hiding this comment.
if something can be none, the type is typically | None
mixtera/core/algo/aioli/aioli.py
Outdated
Comment on lines
82
to
84
| perturbed_domain = self.last_received_mixture % (self.domain_count + 1) | ||
| if perturbed_domain != self.domain_count: | ||
| self.graph[:, perturbed_domain] += self.losses - losses |
Contributor
There was a problem hiding this comment.
Do you want some warmup logic here? Do you immediately start perturbing? Because imagine the very first client feedback, that will have update_at_client = True. Not sure if some checking here is needed.
mixtera/core/algo/aioli/aioli.py
Outdated
Comment on lines
146
to
148
| if self.weights is None: | ||
| self.weights = np.multiply(weights_init, np.exp(self.eta * self.graph.sum(axis=0))) | ||
| else: |
Contributor
There was a problem hiding this comment.
weights = weights_init if self.weights is None else self.weights
self.weights = np.multiply(weights, np.exp(...))
mixtera/core/algo/aioli/aioli.py
Outdated
| self.weights = np.multiply(self.weights, np.exp(self.eta * self.graph.sum(axis=0))) | ||
|
|
||
| logger.info(f"The new mixture proportions={self.weights/sum(self.weights)}. ") | ||
| self.weights = self.weights / sum(self.weights) |
Contributor
There was a problem hiding this comment.
probably want np.sum for speed
mixtera/core/algo/aioli/aioli.py
Outdated
Comment on lines
57
to
60
| for i in range(self.domain_count): | ||
| weight_row = np.ones(self.domain_count) * (1 - self.one_hot_factor) / (self.domain_count - 1) | ||
| weight_row[i] = self.one_hot_factor | ||
| weight_matrix[i] = weight_row |
Contributor
There was a problem hiding this comment.
if you can vectorize this for the entire matrix, that would of course be even better, but I don't know if this currently would be a major bottleneck
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
In this pr, the initial Aioli algorithm is created for dynamic mixtures.