CIANNA usage examples

Classical computer vision examples

Image classification - MNIST

Top-1 accuracy ~99.3% / LeNet-5 like backbone / 630000 ips@(28x28)

Image classification - Imagenet 1000 classes

Top-1 accuracy ~74.7% / Top-5 accuracy ~91.7% / Darknet19 backbone / 740 ips@(448x448)

Object detection - COCO 1000 classes

mAP@50 ~40.1% / COCO-mAP ~21.9% / Darknet19 backbone / 690 ips@(416x416)

Astronomical data examples

Source detection - SKA SDC1: 2D continuum images

560 MHz - 1000h / SDC score 479372 pts / 17 conv. layers backbone / 500 ips@(512x512)

Link to the associated paper => A&A, V690, A211

Source detection - SKA SDC2: 3D HI emission cube

950-1150 MHz - 2000h / SDC score 25453 pts / 23 3D-conv. layers backbone / 300 ips@(64x64x256)

Link to the associated paper => A&A, forthcoming

  1. Images (or Inputs) per second (ips) are provided for an RTX 4090 GPU in inference using FP16C_FP32A mixed precision at the specified resolution and with the maximum batch size to saturate performance.