Image Classification on ImageNet V2 (Std, Disc, Recov Metrics)
82.3Std AccuracyStandard Winograd
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Standard WinogradArchitecture=EfficientNet-B0, Winograd Eligibility=0%, Precision=FP16, Tile Size=F(6,3)2025.12 | 82.3 | — | — | |
| Standard WinogradArchitecture=EfficientNet-B0, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=0%2025.12 | 82.3 | — | — | |
| NOVAArchitecture=EfficientNet-B0, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=0%2025.12 | 82.3 | — | 0 | |
| NOVAArchitecture=ResNet-50, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=25%2025.12 | 80.6 | — | 42 | |
| NOVAArchitecture=ResNet-18, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=65%2025.12 | 77.8 | — | 67 | |
| NOVAArchitecture=VGG-16-BN, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=100%2025.12 | 77.5 | — | 73 | |
| Standard WinogradArchitecture=MobileNet-V2, Winograd Eligibility=0%, Precision=FP16, Tile Size=F(6,3)2025.12 | 77 | — | — | |
| Standard WinogradArchitecture=MobileNet-V2, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=0%2025.12 | 77 | — | — | |
| NOVAArchitecture=MobileNet-V2, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=0%2025.12 | 77 | — | 0 | |
| NOVAArchitecture=VGG-16, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=100%2025.12 | 75.3 | — | 71 | |
| Standard WinogradArchitecture=ResNet-50, Winograd Eligibility=25%, Precision=FP16, Tile Size=F(6,3)2025.12 | 38.3 | — | — | |
| Standard WinogradArchitecture=ResNet-50, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=25%2025.12 | 38.3 | — | — | |
| Standard WinogradArchitecture=ResNet-18, Winograd Eligibility=65%, Precision=FP16, Tile Size=F(6,3)2025.12 | 10.8 | — | — | |
| Standard WinogradArchitecture=ResNet-18, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=65%2025.12 | 10.8 | — | — | |
| Standard WinogradArchitecture=VGG-16, Winograd Eligibility=100%, Precision=FP16, Tile Size=F(6,3)2025.12 | 4.7 | — | — | |
| Standard WinogradArchitecture=VGG-16-BN, Winograd Eligibility=100%, Precision=FP16, Tile Size=F(6,3)2025.12 | 4.7 | — | — | |
| Standard WinogradArchitecture=VGG-16, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=100%2025.12 | 4.7 | — | — | |
| Standard WinogradArchitecture=VGG-16-BN, Winograd Tile=F(6,3), Precision=FP16, Layer Eligibility=100%2025.12 | 4.7 | — | — | |
| NOVAArchitecture=ResNet-18, Winograd Eligibility=65%, Precision=FP16, Tile Size=F(6,3)2025.12 | — | 77.8 | 67 | |
| NOVAArchitecture=ResNet-50, Winograd Eligibility=25%, Precision=FP16, Tile Size=F(6,3)2025.12 | — | 80.6 | 42 | |
| NOVAArchitecture=VGG-16, Winograd Eligibility=100%, Precision=FP16, Tile Size=F(6,3)2025.12 | — | 75.3 | 71 | |
| NOVAArchitecture=VGG-16-BN, Winograd Eligibility=100%, Precision=FP16, Tile Size=F(6,3)2025.12 | — | 77.5 | 73 | |
| NOVAArchitecture=MobileNet-V2, Winograd Eligibility=0%, Precision=FP16, Tile Size=F(6,3)2025.12 | — | 77 | 0 | |
| NOVAArchitecture=EfficientNet-B0, Winograd Eligibility=0%, Precision=FP16, Tile Size=F(6,3)2025.12 | — | 82.3 | 0 |