Feature request: Add `fenn.vision` module for image preprocessing

[blkdmr]

Introduce a new fenn.vision module that provides a small, composable toolbox for image dataset inspection and preprocessing, built on top of NumPy, Pillow and OpenCV. The focus should be ergonomic helpers for common early-stage image EDA (Exploratory Data Analysis) and preprocessing patterns, not re-implementing full computer vision libraries.

Goal

Create an initial version of fenn.vision with a clear, well-documented API to simplify early-stage image exploration and preprocessing for ML workflows. The module should make it easy to quickly inspect, summarize, sanity-check, and standardize image datasets represented as numpy.ndarray tensors or directory-based image collections.

Proposed features

Ideas for a first iteration (not all are mandatory for a single PR):

• image_summary(array): one-shot overview for a batch of images, combining shape distribution (H, W, C), dtype, value range, per-channel mean/std, and simple NaN/inf checks.
• image_dir_summary(path): scan a folder (optionally recursive) and summarize counts, common resolutions, formats, and a few representative examples; does not load all pixels into memory at once.
• check_image_batch(array): fast sanity checks for typical issues (mixed dtypes, unexpected channel counts, extreme outliers, non-finite values) returning a structured report rather than raising immediately.
• ensure_color_mode(array, mode="RGB"): convert grayscale / RGBA to the desired channel layout (e.g. expand gray to 3 channels, drop alpha) for NumPy images.
• resize_batch(array, size, interpolation="bilinear"): convenience wrapper to resize a batch of images to a target size, preserving channel order and dtype where possible.
• normalize_batch(array, mode="0_1" | "minus1_1" | "imagenet_stats" | "zscore"): simple, opinionated normalization helpers with explicit, documented behavior and no hidden global state.
• standard_preprocess(array, *, size=None, color_mode="RGB", normalize="0_1"): compose common operations (dtype cast, channel fix, optional resize, normalization) into a deterministic, testable pipeline; returns a new array without in-place modification.
• augment_debug_view(array, augment_fn, n=8): run a user-provided augmentation function on a small sample and return the original/augmented pairs as a tiny batch for quick visual sanity checks (no plotting, just data).

These functions should be pure utilities (no global state) and must not alter the inputs in place; all outputs are new arrays or small summary objects.

Tasks

• Create the fenn/vision/__init__.py module and basic package structure, mirroring the style of fenn.tabular but tailored to images.
• Implement a first subset of utilities (for example: image_summary, check_image_batch, standard_preprocess).
• Add type hints and docstrings with small usage examples showing typical NumPy batch shapes like (N, H, W, C) and (H, W, C).
• Add unit tests with small synthetic arrays: tiny RGB/gray images, mixed dtypes, arrays containing NaNs/infs, and non-uniform shapes for directory-based helpers.
• Integrate the new module into the public API and update the documentation with a short “Image preprocessing and EDA” section, including a minimal end‑to‑end example.

Contributing

If you want to work on this:

• Read the project’s CONTRIBUTING guide before starting.
• Comment on the issue to claim a subset and join the Discord server to coordinate scope and API choices for the first iteration.

[luke-cunningham-1]

I would like to pickup image_summary, resize_batch, and standard_preprocess

[luke-cunningham-1]

I am also picking up ensure_color_mode and normalize_batch

[ParallelVoid]

I am picking up check_image_batch

[devendrajadhav2470]

Hi @blkdmr, is there any update on this issue? Can I work on this?

[blkdmr]

Hi @devendrajadhav2470,

At this point, almost all features have been implemented and can be found in the experimental directory, but they are not yet integrated into the "fenn ecosystem". The codebase needs to be reorganized to improve usability, properly documented (see the docs repository for reference), and complemented with a small, ready-to-use template. For examples, refer to resnet-devs-only in the templates repo for a cnn setup or the mlp-* directories for a complete template.

[blkdmr]

@devendrajadhav2470 If you want to contribute, let me know which of these activities you want to take ownership of (you can also take all of them). Right now, the highest priority is creating a complete cnn template, noting that part of this work will involve the templates repository.

[devendrajadhav2470]

@blkdmr, Thanks for the detailed explanation. I’d be happy to contribute.
I can take ownership of creating the complete CNN template.
Let me know if this works.

[devendrajadhav2470]

Hi @blkdmr, any update? can I work on the CNN template?

[blkdmr]

Oh sorry, yes @devendrajadhav2470 you can work on it!

[blkdmr]

For everyone: there’s still some work to be done on this part of the project. Please leave a comment if you’d like to work on it!