galthran-wq / face-recognition-service

74%

Build Type

push

github

Committed by web-flow

Commit Message

TensorRT FP16, GIL-free alignment, and threaded batch processing (#4)

- Enable TensorRT EP with FP16 via FACE_USE_TENSORRT env var (1.5-2.3x speedup)
- Replace skimage SimilarityTransform with pure numpy lstsq (1.5x faster, releases GIL)
- Parallelize image decoding and face alignment in batch endpoints with ThreadPoolExecutor
- Add CUDA EP provider_options (arena strategy, cuDNN exhaustive search)
- Inject ORT SessionOptions via PickableInferenceSession monkey-patch
- Increase max batch size from 20 to 64
- Add TensorRT runtime libs to Dockerfile.gpu
- Add GPU performance section to README with benchmark results
- Add quality comparison, batch profiling, and GIL test benchmarks

Batch=32 embed pipeline: 531ms → 359ms (1.5x). Single-image embed: 78ms → ~14ms.
FP16 embedding quality: cosine similarity 0.9998+ vs FP32.

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Run Details

45 of 91 new or added lines in 2 files covered. (49.45%)

3 existing lines in 1 file now uncovered.

439 of 590 relevant lines covered (74.41%)

0.74 hits per line