feat(ocl): add EWC

[tachyonicClock]

No description provided.

[Copilot]

Pull request overview

Adds an Elastic Weight Consolidation (EWC) continual-learning strategy and refactors replay/coreset utilities to support it, along with new tests.

Changes:

• Introduce EWC strategy implementation and export it from capymoa.ocl.strategy.
• Replace/extend coreset utilities with a unified replay buffer module (_replay.py) and add a BufferList helper for registering tensor buffers.
• Update strategy imports and extend OCL strategy test coverage (including new EWC cases and a new BufferList unit test).

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
tests/ocl/test_strategy.py	Adds EWC test cases and refactors learner construction logic.
tests/ocl/test_buffer_list.py	Adds unit test for BufferList ordering/identity behavior.
src/capymoa/ocl/util/_replay.py	Introduces ReplayBuffer base and adds SlidingWindow; migrates samplers here.
src/capymoa/ocl/util/_buffer_list.py	Adds BufferList module to manage tensors as registered buffers.
src/capymoa/ocl/strategy/_rar.py	Updates import to use replay module.
src/capymoa/ocl/strategy/_gdumb.py	Updates import to use replay module.
src/capymoa/ocl/strategy/_experience_replay.py	Updates import to use replay module.
src/capymoa/ocl/strategy/_ewc.py	Adds the new EWC strategy implementation.
src/capymoa/ocl/strategy/init.py	Exports EWC from the strategy package.

Comments suppressed due to low confidence (4)

src/capymoa/ocl/util/_replay.py:120

• GreedySampler.update forces examples to CPU via xi.cpu() when writing into _buffer_x. This breaks if the replay buffer is moved to a non-CPU device (and also introduces unnecessary device transfers). Store tensors on the buffer’s device instead (e.g., xi/xi.to(self.device)) in both places where xi.cpu() is used.
  src/capymoa/ocl/util/_replay.py:94
• ReservoirSampler.update uses index = torch.randint(..., (1,), ...) and then compares it in if index < self.capacity:. Because index is a 1-D tensor, this will raise when the buffer is full (a Tensor can’t be used as a Python boolean). Convert index to a Python int (e.g., use size () and cast, or call .item()) before the comparison and indexing.
  src/capymoa/ocl/util/_replay.py:27
• sample() ignores the per-buffer RNG (self._rng) and uses the global RNG. Since ReplayBuffer stores an RNG and subclasses already rely on it for determinism, sample() should also pass generator=self._rng to keep sampling reproducible across runs and independent of global state.
  src/capymoa/ocl/util/_replay.py:57
• Defaulting rng to torch.Generator() creates a single shared Generator instance at import time, so different buffers will unintentionally share RNG state. Prefer rng: torch.Generator | None = None and create a new generator inside __init__ when rng is None.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.