[Good First Issue] Add ONNX Runtime inference path with PyTorch fallback #12

Overview

scripts/export_model.py exports PyTorch models to ONNX, but the API still loads .pth checkpoints directly. ONNX Runtime would provide faster inference and lower memory usage.

Scope

• Update _load_model() in src/climatevision/inference/pipeline.py to prefer .onnx files when present
• Fall back to PyTorch .pth if ONNX file is missing
• Verify ONNX input/output tensor signatures match per analysis type:
  • UNet: (N, n_channels, 256, 256) → (N, n_classes, 256, 256)
  • Siamese: (N, C, 256, 256) × 2 → (N, 2, 256, 256)
• Add latency benchmark logging (compare PyTorch vs ONNX inference time)
• Update scripts/export_model.py to export all analysis-specific models

Acceptance Criteria

• When .onnx checkpoint exists, API loads it via ONNX Runtime
• Latency difference is logged per request
• Fallback to PyTorch works seamlessly when ONNX export is absent
• All 3 analysis types supported

Resources

• src/climatevision/inference/pipeline.py — _load_model() function
• scripts/export_model.py — existing ONNX export script
• references/model-architectures.md — tensor shape specs

Difficulty: Intermediate
Labels: help wanted, backend, mlops, performance