Fix Sliding Window Logic and Add Error Handling in Inference Scripts

[stephen-netu]

Greetings,

This pull request addresses a critical bug in the sliding window implementation that prevented proper image coverage for non-square inputs, and adds comprehensive error handling to improve script reliability and user experience.

Key Changes

1. Sliding Window Bug Fix (reward_model_example_inference.py)

• Problem: Original logic only slid windows in one direction based on aspect ratio, missing significant image regions for tall/wide images
• Solution: Implemented bidirectional sliding with 50% overlap for complete coverage regardless of aspect ratio
• Impact: Ensures accurate UI quality scoring across all image types

2. Error Handling Improvements

• generator_model_client_inference.py:

  • Added network request error handling (timeouts, connection failures)
  • Added JSON parsing validation with KeyError handling
  • Improved error messages for debugging

• reward_model_example_inference.py:

  • Added file loading error handling for images
  • Added model/processor loading error handling
  • Added CUDA availability check with fallback to CPU
  • Added descriptive error messages for all failure modes

3. Code Quality Enhancements

• Added function docstrings for better documentation
• Consistent code formatting and style improvements
• Removed redundant logic in sliding window implementation

Testing Performed

• Verified sliding window now covers entire image with proper overlap
• Tested error conditions (missing files, invalid models, network failures)
• Confirmed backward compatibility with existing usage
• Validated performance impact is minimal

Files Modified

• generator_model_client_inference.py
• reward_model_example_inference.py