[feat] Channel-First Storage (C, H, W)?

[kuang-da]

When optimizing for GPU, (C, H, W) gives contigent memory if we transfer data between GPU and CPU. But we currently use (H, W, C) to save image.

The following is the potential plan, but we do not have to implement it right now.

Implementation: Change codex_patches.py to save/load as (C, H, W) instead of (H, W, C)

Current: patch_img[:, :, idx] → non-contiguous

Proposed: patch_img[idx, :, :] → contiguous!

Pros:

• Zero runtime overhead - slices are naturally contiguous
• Most efficient long-term solution

Cons:

• Breaking change - existing saved patches incompatible
• Requires re-running segmentation or migration code
• Many libraries expect (H, W, C): scikit-image, matplotlib, OpenCV
• Segmentation models (DeepCell) expect (H, W, C) input
• High risk of bugs throughout pipeline

[kuang-da]

Memory Layout Issue

Channel-last (H, W, C): arr[:, :, idx] → stride = C*dtype_size → NON-contiguous
Channel-first (C, H, W): arr[idx, :, :] → stride = dtype_size → contiguous