The models provide 16 x float32 or 8 x float32 output vectors.
Data distribution allows float32 packing to int16 values,
providing 32-byte or 16-byte respective int8 vectors.
Face Detection - finding face boundaries on the image.
Face Recognition (Identification) - finding whether two face photos belong to the same person.
- Graphcore IPUs x1/x4/x16
- Nvidia GPUs x4/x16/x32 (torch: DataParallel, DistributedDataParallel)
- Google TPUs (no meaningful results)
- Google Colaboratory (fast networking)
- Kaggle (datasets)
- spell.ml (workflows, Graphcore hardware)
- Paperspace (workflows, Graphcore hardware)
- immers.cloud (dozens of GPUs, unlimited RAM, fast SSDs)
- FaceId_gc.ipynb - train
- FaceId_infer.ipynb - validation
- FaceId_gc_ov8.ipynb - train
- FaceId_infer_ov8.ipynb - validation