Image Classification on ImageNet 1k (val) (Top-1/Top-5 Accuracy)
85Top-1 AccCastling-MViTv2-B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Castling-MViTv2-BMACs Ranges=>10G, Params (M)=51.9, MACs (G)=9.822022.11 | 85 | 97.2 | |
| Castling-MViTv2-SMACs Ranges=3~10G, Params (M)=34.7, MACs (G)=6.952022.11 | 84.6 | 97 | |
| MViTv2-BMACs Ranges=>10G, Params (M)=51.2, MACs (G)=10.072022.11 | 84.4 | — | |
| CSWin-BMACs Ranges=>10G, Params (M)=78.0, MACs (G)=15.002022.11 | 84.2 | — | |
| Castling-DeiT-BMACs Ranges=>10G, Params (M)=87.22, MACs (G)=17.282022.11 | 84.2 | — | |
| Castling-MViTv2-TMACs Ranges=3~10G, Params (M)=24.1, MACs (G)=4.502022.11 | 84.1 | 96.8 | |
| Castling-LeViT-384MACs Ranges=1~3G, Params (M)=45.8, MACs (G)=2.902022.11 | 83.7 | 96.7 | |
| EfficientNet-B5MACs Ranges=3~10G, Params (M)=30.0, MACs (G)=9.902022.11 | 83.6 | 96.7 | |
| CSWin-SMACs Ranges=3~10G, Params (M)=35.0, MACs (G)=6.902022.11 | 83.6 | — | |
| MViTv2-SMACs Ranges=3~10G, Params (M)=34.7, MACs (G)=6.952022.11 | 83.6 | — | |
| Swin-BMACs Ranges=>10G, Params (M)=88.0, MACs (G)=15.402022.11 | 83.5 | 96.5 | |
| DeiT-BMACs Ranges=>10G, Params (M)=86.3, MACs (G)=17.562022.11 | 83.4 | 96.5 | |
| CaiT-S36MACs Ranges=>10G, Params (M)=68.0, MACs (G)=13.902022.11 | 83.3 | — | |
| Swin-SMACs Ranges=3~10G, Params (M)=50.0, MACs (G)=8.702022.11 | 83.2 | 96.2 | |
| EfficientNet-B4MACs Ranges=3~10G, Params (M)=19.0, MACs (G)=4.202022.11 | 82.9 | 96.4 | |
| CSWin-TMACs Ranges=3~10G, Params (M)=23.0, MACs (G)=4.302022.11 | 82.7 | — | |
| LeViT-384MACs Ranges=1~3G, Params (M)=39.1, MACs (G)=2.352022.11 | 82.6 | — | |
| Castling-LeViT-256MACs Ranges=1~3G, Params (M)=22.0, MACs (G)=1.402022.11 | 82.6 | 96.1 | |
| AutoFormer-BMACs Ranges=>10G, Params (M)=54.0, MACs (G)=11.002022.11 | 82.4 | 95.7 | |
| MViTv2-TMACs Ranges=3~10G, Params (M)=24.0, MACs (G)=4.682022.11 | 82.3 | — | |
| PVTv2-V2MACs Ranges=3~10G, Params (M)=25.0, MACs (G)=4.002022.11 | 82 | — | |
| RegNetY-8GMACs Ranges=3~10G, Params (M)=39.0, MACs (G)=8.002022.11 | 81.7 | — | |
| AutoFormer-SMACs Ranges=3~10G, Params (M)=22.9, MACs (G)=5.102022.11 | 81.7 | 95.7 | |
| EfficientNet-B3MACs Ranges=1~3G, Params (M)=12.0, MACs (G)=1.802022.11 | 81.6 | 95.7 | |
| LeViT-256MACs Ranges=1~3G, Params (M)=18.9, MACs (G)=1.122022.11 | 81.6 | — | |
| Castling-LeViT-192MACs Ranges=<1G, Params (M)=12.7, MACs (G)=0.822022.11 | 81.3 | 95.5 | |
| Swin-TMACs Ranges=3~10G, Params (M)=29.0, MACs (G)=4.502022.11 | 81.3 | 95.5 | |
| DeiT-SMACs Ranges=3~10G, Params (M)=21.9, MACs (G)=4.602022.11 | 81.2 | 95.4 | |
| GhostNetV3Scale=1.6x, Params (M)=12.3, FLOPs (M)=399, Latency(ms) CPU=18.87, Latency(ms) Mobile=38.462024.04 | 80.4 | 95.2 | |
| LeViT-192MACs Ranges=<1G, Params (M)=10.9, MACs (G)=0.662022.11 | 80 | — | |
| RegNetY-4GMACs Ranges=3~10G, Params (M)=21.0, MACs (G)=4.002022.11 | 80 | — | |
| Castling-LeViT-128MACs Ranges=<1G, Params (M)=10.5, MACs (G)=0.492022.11 | 79.6 | 94.6 | |
| MobileOne-S4Params (M)=14.8, FLOPs (M)=2978, Latency(ms) CPU=52.86, Latency(ms) Mobile=92.062024.04 | 79.4 | — | |
| Mobileformer-508Params (M)=14, FLOPs (M)=508, Latency(ms) CPU=129.58, Latency(ms) Mobile=163.392024.04 | 79.3 | — | |
| SAQNetwork=ViT-B/16, Bitwidth (W/A)=4/42021.11 | 79.2 | 94.2 | |
| GhostNetV3Scale=1.3x, Params (M)=8.9, FLOPs (M)=269, Latency(ms) CPU=14.46, Latency(ms) Mobile=28.692024.04 | 79.1 | 94.5 | |
| EfficientNet-B1MACs Ranges=<1G, Params (M)=7.8, MACs (G)=0.702022.11 | 79.1 | 94.4 | |
| MixNet-LParams (M)=7.3, FLOPs (M)=565, Latency(ms) CPU=34.01, Latency(ms) Mobile=67.932024.04 | 78.9 | 94.2 | |
| LeViT-128MACs Ranges=<1G, Params (M)=9.2, MACs (G)=0.412022.11 | 78.6 | — | |
| HRFormer-TMACs Ranges=1~3G, Params (M)=8.0, MACs (G)=1.802022.11 | 78.5 | — | |
| MobileViT-SParams (M)=5.6, FLOPs (M)=1792, Latency(ms) CPU=35.93, Latency(ms) Mobile=92.282024.04 | 78.4 | — | |
| MobileOne-S3Params (M)=10.1, FLOPs (M)=1896, Latency(ms) CPU=33.31, Latency(ms) Mobile=58.332024.04 | 78.1 | — | |
| LSQ+Network=ViT-B/16, Bitwidth (W/A)=4/42021.11 | 78 | 93.4 | |
| GhostNetV2Scale=1.6x, Params (M)=12.3, FLOPs (M)=399, Latency(ms) CPU=18.87, Latency(ms) Mobile=38.462024.04 | 77.8 | 93.8 | |
| MobileOne-S2Params (M)=7.8, FLOPs (M)=1299, Latency(ms) CPU=23.04, Latency(ms) Mobile=40.282024.04 | 77.4 | — | |
| GhostNetV1Scale=1.7x, Params (M)=11, FLOPs (M)=378, Latency(ms) CPU=18.45, Latency(ms) Mobile=37.242024.04 | 77.2 | 93.4 | |
| GhostNetV3Scale=1.0x, Params (M)=6.1, FLOPs (M)=167, Latency(ms) CPU=7.81, Latency(ms) Mobile=14.692024.04 | 77.1 | 93.3 | |
| EfficientNet-B0MACs Ranges=<1G, Params (M)=5.3, MACs (G)=0.392022.11 | 77.1 | 93.3 | |
| MixNet-MParams (M)=5, FLOPs (M)=360, Latency(ms) CPU=21.44, Latency(ms) Mobile=43.122024.04 | 77 | 93.3 | |
| SAQNetwork=ViT-S/16, Bitwidth (W/A)=4/42021.11 | 76.9 | 93.5 | |
| GhostNetV2Scale=1.3x, Params (M)=8.9, FLOPs (M)=269, Latency(ms) CPU=14.46, Latency(ms) Mobile=28.692024.04 | 76.9 | 93.4 | |
| TeacherTeacher backbone=DenseNet201, Student backbone=MobileNet2022.10 | 76.89 | 93.37 | |
| EfficientNet-B0Params (M)=5.3, FLOPs (M)=390, Latency(ms) CPU=12.5, Latency(ms) Mobile=37.672024.04 | 76.3 | 93.2 | |
| LSQ+Network=ViT-S/16, Bitwidth (W/A)=4/42021.11 | 76.1 | 93 | |
| MobileNeXtScale=1.4x, Params (M)=6.1, FLOPs (M)=590, Latency(ms) CPU=11.46, Latency(ms) Mobile=21.652024.04 | 76.1 | — | |
| Castling-DeiT-TMACs Ranges=1~3G, Params (M)=5.64, MACs (G)=1.182022.11 | 76 | 92.5 | |
| MobileOne-S1Params (M)=4.8, FLOPs (M)=825, Latency(ms) CPU=14.86, Latency(ms) Mobile=27.212024.04 | 75.9 | — | |
| MixNet-SParams (M)=4.1, FLOPs (M)=256, Latency(ms) CPU=15.75, Latency(ms) Mobile=30.792024.04 | 75.8 | 92.8 | |
| GhostNetV1Scale=1.3x, Params (M)=7.3, FLOPs (M)=226, Latency(ms) CPU=12.69, Latency(ms) Mobile=24.312024.04 | 75.7 | 93.4 | |
| GhostNetV2Scale=1.0x, Params (M)=6.1, FLOPs (M)=167, Latency(ms) CPU=7.81, Latency(ms) Mobile=14.692024.04 | 75.3 | 92.4 | |
| MobileNetV3-LParams (M)=5.4, FLOPs (M)=219, Latency(ms) CPU=7.25, Latency(ms) Mobile=14.242024.04 | 75.2 | — | |
| MNASNet-A1Params (M)=3.9, FLOPs (M)=312, Latency(ms) CPU=7.81, Latency(ms) Mobile=13.572024.04 | 75.2 | 92.5 | |
| MobileNetV2Scale=1.0x, Training Strategy=GhostNetV3 strategy, Params (M)=3.4, FLOPs (M)=300, Latency(ms) CPU=7.96, Latency(ms) Mobile=14.972024.04 | 75 | 92.1 | |
| ShuffleNetV2Scale=2.0x, Params (M)=7.4, FLOPs (M)=591, Latency(ms) CPU=15.62, Latency(ms) Mobile=25.362024.04 | 74.9 | — | |
| FBNet-CParams (M)=5.5, FLOPs (M)=375, Latency(ms) CPU=9.37, Latency(ms) Mobile=15.672024.04 | 74.9 | — | |
| MobileViT-XSParams (M)=2.3, FLOPs (M)=724, Latency(ms) CPU=23.43, Latency(ms) Mobile=64.342024.04 | 74.8 | — | |
| MobileViT-XSMACs Ranges=<1G, Params (M)=2.3, MACs (G)=0.702022.11 | 74.8 | 92.3 | |
| MobileNetV2Scale=1.4x, Params (M)=6.9, FLOPs (M)=585, Latency(ms) CPU=13.07, Latency(ms) Mobile=22.852024.04 | 74.7 | — | |
| MobileNetV2-1.4MACs Ranges=<1G, Params (M)=6.9, MACs (G)=0.582022.11 | 74.7 | — | |
| DeiT-TMACs Ranges=1~3G, Params (M)=5.6, MACs (G)=1.252022.11 | 74.5 | 91.9 | |
| MobileNeXtScale=1.0x, Params (M)=3.4, FLOPs (M)=311, Latency(ms) CPU=8.63, Latency(ms) Mobile=16.172024.04 | 74 | — | |
| GhostNetV1Scale=1.0x, Params (M)=5.2, FLOPs (M)=141, Latency(ms) CPU=6.25, Latency(ms) Mobile=12.652024.04 | 73.9 | 91.4 | |
| PEFDTeacher backbone=DenseNet201, Student backbone=MobileNet2022.10 | 73.24 | 91.47 | |
| ConViT-tinyParams (M)=5.7, FLOPs (M)=1000, Latency(ms) CPU=28.75, Latency(ms) Mobile=58.732024.04 | 73.1 | 91.7 | |
| Mobileformer-96Params (M)=4.6, FLOPs (M)=96, Latency(ms) CPU=42.18, Latency(ms) Mobile=101.372024.04 | 72.8 | — | |
| SAQNetwork=ViT-B/32, Bitwidth (W/A)=4/42021.11 | 72.7 | 90.7 | |
| PRACTISELatency=34.9 ms, Training samples=500, Backbone=ResNet-342022.02 | 72.5 | 90.9 | |
| PRACTISELatency=34.9 ms, Training samples=1000, Backbone=ResNet-342022.02 | 72.5 | 91 | |
| PiT-xsParams (M)=10.6, FLOPs (M)=1400, Latency(ms) CPU=32.92, Latency(ms) Mobile=69.72024.04 | 72.4 | — | |
| SRRLTeacher backbone=DenseNet201, Student backbone=MobileNet2022.10 | 72.28 | 90.9 | |
| DC-ViTBackbone=ResNet-34, Latency (ms)=28.8, Training samples=502024.03 | 72.2 | 90.7 | |
| DeiT-tinyParams (M)=5.9, FLOPs (M)=1300, Latency(ms) CPU=28.01, Latency(ms) Mobile=56.42024.04 | 72.2 | — | |
| LSQ+Network=ViT-B/32, Bitwidth (W/A)=4/42021.11 | 72.1 | 90.4 | |
| SAQNetwork=MobileNetV2, Bitwidth (W/A)=4/42021.11 | 72 | 90.4 | |
| MobileNetV2Scale=1.0x, Params (M)=3.4, FLOPs (M)=300, Latency(ms) CPU=7.96, Latency(ms) Mobile=14.972024.04 | 72 | 90.8 | |
| CIDTeacher backbone=DenseNet201, Student backbone=MobileNet2022.10 | 71.9 | 90.97 | |
| OriginalLatency (ms)=37.6, Backbone=MobileNetV22022.02 | 71.9 | 90.3 | |
| PROFITNetwork=MobileNetV2, Bitwidth (W/A)=4/42021.11 | 71.6 | 90.4 | |
| ShuffleNetV2Scale=1.0x, Training Strategy=GhostNetV3 strategy, Params (M)=2.3, FLOPs (M)=146, Latency(ms) CPU=4.69, Latency(ms) Mobile=11.072024.04 | 71.6 | 89.8 | |
| PRACTISELatency=34.9 ms, Training samples=100, Backbone=ResNet-342022.02 | 71.5 | 90.3 | |
| MobileOne-S0Params (M)=2.1, FLOPs (M)=275, Latency(ms) CPU=5.48, Latency(ms) Mobile=12.342024.04 | 71.4 | — | |
| PiT-tiParams (M)=4.9, FLOPs (M)=710, Latency(ms) CPU=21.87, Latency(ms) Mobile=45.342024.04 | 71.3 | — | |
| CRDTeacher backbone=DenseNet201, Student backbone=MobileNet2022.10 | 70.82 | 90.09 | |
| OOQNetwork=MobileNetV2, Bitwidth (W/A)=4/42021.11 | 70.6 | — | |
| MobileNetV1Params (M)=4.2, FLOPs (M)=575, Latency(ms) CPU=7.02, Latency(ms) Mobile=12.852024.04 | 70.6 | — | |
| BRNetwork=MobileNetV2, Bitwidth (W/A)=4/42021.11 | 70.4 | 89.4 | |
| SPTeacher backbone=DenseNet201, Student backbone=MobileNet2022.10 | 70.34 | 89.63 | |
| PRACTISELatency=34.9 ms, Training samples=50, Backbone=ResNet-342022.02 | 70.3 | 89.6 | |
| PRACTISELatency (ms)=31.9, Latency reduction (%)=15.2%, Training samples=500, Backbone=MobileNetV22022.02 | 70.3 | 89.5 | |
| EWGSNetwork=MobileNetV2, Bitwidth (W/A)=4/42021.11 | 70.3 | — |