Product Performance Deepdive
World-class AI model compression performance, with the highest Accuracy Retention
| Type | Object Detection | Classification | Super Resolution | Transformer | ||
|---|---|---|---|---|---|---|
| Name | YoloV7 | RetinaFace-ResNet50 | ResNet18 | MobilenetV3 (L) | IMDN | ViT B16 |
| Original | 53.2 | 54.7 | 69.7 | 75.3 | 27.9 | 81.07 |
| CLIKA | No Loss | No Loss | No Loss | -0.6 | -0.1 | -1.45 |
| Meta PyTorch | -3.2 | -1.8 | -0.7 | -2.0 | -1.2 | FAILED |
| Nvidia TensorRT | -21.8 | -1.6 | -0.6 | FAILED | -9.0 | N/A |
| Google TFLite | N/A | N/A | FAILED | FAILED | N/A | N/A |
| Intel NNCF | No Loss | -1.25 | -0.5 | -4.9 | -1.94 | -6.4 |