[ashishsoni-ai]

Description

Resolves #47.

Currently, models in \devolearn\ only allow inference on images by specifying a file path on disk. If users have loaded image data in memory as a NumPy array (e.g. from OpenCV, video frames, preprocessing pipelines, or synthetic generators) or a PIL Image, they previously had to write the image to a temporary file on disk first before calling .predict().

This PR extends the .predict()\ method across all models (\cell_membrane_segmentor, \cell_nucleus_segmentor, and \lineage_population_model) to directly accept in-memory NumPy arrays and PIL Images in addition to file paths, while maintaining full backward compatibility.

Changes Made

1. **Added \InferenceEngine.load_image()**:

   • Centralized image loading and standardization in \devolearn.base_inference_engine.InferenceEngine.
   • Supports:
     • File paths (\str\ / \os.PathLike)
     • In-memory NumPy arrays (2D grayscale, 3D BGR/RGB, 4D BGRA/RGBA, and float [0.0, 1.0]\ normalized arrays)
     • PIL Images (\PIL.Image.Image)
   • Performs proper format validation and error handling (\FileNotFoundError, \ValueError, \TypeError).

2. Updated Model \predict()\ Signatures & Implementations:

   • Updated \cell_membrane_segmentor.predict(image_path=None, ..., image=None).
   • Updated \cell_nucleus_segmentor.predict(image_path=None, ..., image=None).
   • Updated \lineage_population_model.predict(image_path=None, image=None).
   • Users can now call either \model.predict('path/to/img.png'), \model.predict(img_array), or explicitly pass \model.predict(image=img_array).

3. **Fixed \InferenceEngine.repr**:

   • Fixed _repr_\ to return a string (\str(self.model)) instead of returning an
     n.Module\ object, avoiding \TypeError: repr returned non-string\ during debugging, logging, or test failure reporting.

4. Testing:

   • Added unit tests in \devolearn/tests/test.py\ validating NumPy array prediction (both positional and keyword arguments) across all three inference models.
   • Added a dedicated test case (\ est_predict_numpy_and_pil_inputs) verifying:
     • 2D grayscale arrays
     • 3D BGR arrays
     • Float [0, 1]\ arrays
     • PIL Images
     • Centroid mode with NumPy array inputs
     • Exception handling on invalid inputs
   • Verified that all tests pass (\pytest devolearn/tests/test.py).

5. Documentation:

   • Added an example in \README.md\ demonstrating how to predict directly on in-memory NumPy arrays.