Add frames_in_cpu and responses_in_cpu to project_stimulus_batched

[Copilot]

Under GPU backends (torch_cuda/torch_mps), project_stimulus_batched moved the entire stimulus to VRAM and allocated the full (nimages, nfilters) response matrix on GPU — defeating the VRAM-saving purpose of batching.

Changes

moten/core.py

• Added frames_in_cpu=False: when True, keeps the full stimulus in CPU RAM; only the per-temporal-batch slice is transferred to the device via backend.asarray(stim_chunk) inside the loop.
• Added responses_in_cpu=False: when True, allocates the output with np.zeros(...) on CPU and copies each batch result back via backend.to_numpy(channel_response). Return value is always a NumPy ndarray.
• Both default to False — no behavior change for existing callers.

moten/pyramids.py

• MotionEnergyPyramid.project_stimulus_batched accepts and forwards both new kwargs.

moten/tests/test_batched.py

• New TestCpuMemoryFlags class (12 tests) covering: numpy/torch backends, flags individually and combined, with stimulus_batch_size, and passing a NumPy stimulus with frames_in_cpu=True under torch.

Usage

# Keep both stimulus and output off GPU; only batch slices touch VRAM
responses = pyramid.project_stimulus_batched(
    stimulus,
    stimulus_batch_size=32,
    frames_in_cpu=True,
    responses_in_cpu=True,
)
# responses is a NumPy ndarray regardless of active backend

[Copilot]

Pull request overview

This PR adds two optional flags to project_stimulus_batched to reduce GPU VRAM usage on torch GPU backends by (a) keeping the full stimulus in CPU RAM and transferring only per-temporal-batch chunks to the device, and/or (b) keeping the full (nimages, nfilters) response matrix on CPU and copying each batch result back.

Changes:

• Add frames_in_cpu and responses_in_cpu kwargs to moten.core.project_stimulus_batched.
• Expose and forward the same kwargs via MotionEnergyPyramid.project_stimulus_batched.
• Add a new TestCpuMemoryFlags test class covering numpy/torch backends and flag combinations.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
moten/core.py	Implements the new CPU-memory flags in the batched projection core loop and updates the docstring.
moten/pyramids.py	Extends MotionEnergyPyramid.project_stimulus_batched signature/docs and forwards the new kwargs to core.
moten/tests/test_batched.py	Adds coverage for frames_in_cpu / responses_in_cpu behavior across numpy and torch backends.

---

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