FQ-ViT: Post-Training Quantization for Fully Quantized Vision Transformer

Open paper
Benchmark Model Rank Results
quantization-on-imagenetFQ-ViT (ViT-L)#1Top-1 Accuracy (%): 85.03Weight bits: 8Activation bits: 8
quantization-on-imagenetFQ-ViT (ViT-B)#2Top-1 Accuracy (%): 83.31Weight bits: 8Activation bits: 8
quantization-on-imagenetFQ-ViT (Swin-B)#3Top-1 Accuracy (%): 82.97Weight bits: 8Activation bits: 8
quantization-on-imagenetFQ-ViT (Swin-S)#4Top-1 Accuracy (%): 82.71Weight bits: 8Activation bits: 8
quantization-on-imagenetFQ-ViT (DeiT-B)#5Top-1 Accuracy (%): 81.20Weight bits: 8Activation bits: 8
quantization-on-imagenetFQ-ViT (Swin-T)#6Top-1 Accuracy (%): 80.51Weight bits: 8Activation bits: 8
quantization-on-imagenetFQ-ViT (DeiT-S)#7Top-1 Accuracy (%): 79.17Weight bits: 8Activation bits: 8
quantization-on-imagenetFQ-ViT (DeiT-T)#21Top-1 Accuracy (%): 71.61Weight bits: 8Activation bits: 8

Rank counts only results with a code link.