JieShibo / PETL-ViT

[ICCV 2023] Binary Adapters, [AAAI 2023] FacT, [Tech report] Convpass
MIT License
162 stars 7 forks source link

Few-shot results #11

Closed heekhero closed 1 year ago

heekhero commented 1 year ago

Hi, great work! Could you please provide me with the exact numbers of results in Fig.6 in terms of FacT-TT (61K), so I can use it to check the reproducibility and compare with my research. Looking forward to hearing from you.

JieShibo commented 1 year ago

Hi @heekhero ,

-- Food101 StanfordCars Flowers102 FGVCAircraft OxfordPets
NOAH-1 shot 31.8 7.3 88.1 8.2 69.7
NOAH-2 shot 51.6 13.2 96.9 13.5 78.1
NOAH-4 shot 63.8 25.5 98.1 23.1 84.9
NOAH-8 shot 71.5 46.2 99.4 34.0 88.0
NOAH-16 shot 76.3 68.6 99.5 49.1 89.0
VPT-1 shot 23.5 5.5 58.8 6.4 56.8
VPT-2 shot 46.7 9.6 95.4 11.2 66.1
VPT-4 shot 60.6 20.7 98.0 19.0 85.3
VPT-8 shot 67.8 33.5 99.2 28.9 88.6
VPT-16 shot 72.6 56.0 99.4 42.5 89.6
Lora-1 shot 31.8 6.7 88.5 9.0 68.6
Lora-2 shot 48.7 13.5 96.4 13.5 76.1
Lora-4 shot 59.0 26.4 98.1 22.1 85.7
Lora-8 shot 66.4 45.9 99.1 34.6 87.3
Lora-16 shot 72.5 68.2 99.6 47.6 88.7
Adapter-1 shot 31.4 7.5 88.6 9.1 68.0
Adapter-2 shot 49.0 13.4 96.3 13.3 76.3
Adapter-4 shot 59.8 23.7 98.1 21.1 85.8
Adapter-8 shot 66.8 39.9 99.2 32.0 88.2
Adapter-16 shot 71.7 60.4 99.5 45.2 89.1
AdapterFormer-1 shot 31.6 6.7 90.1 9.6 62.6
AdapterFormer-2 shot 48.2 13.2 96.4 13.2 74.7
AdapterFormer-4 shot 59.3 26.5 97.7 21.7 83.6
AdapterFormer-8 shot 67 46.6 99.0 33.9 87.2
AdapterFormer-16 shot 72.4 66.1 99.4 48.5 88.8
FacT-1 shot 34.17 8.33 91.61 11.06 73.97
FacT-2 shot 54.00 15.44 96.95 15.50 81.54
FacT-4 shot 64.34 29.49 98.44 24.14 87.14
FacT-8 shot 71.29 52.18 99.4 36.76 89.22
FacT-16 shot 75.85 72.06 99.69 50.98 89.84

Note that the results of Flowers102 in the arxiv paper are out of date. Please refer to the table above.

heekhero commented 1 year ago

Thank you for sharing the results!

zhuyuedlut commented 3 months ago

@JieShibo Hello, it's a great work. I have two question about FGVC dataset.

  1. As we known, NOAH split FGVC by different seed. So the result that in above table is average result ?
  2. I test the TT-16 in FGVC dataset, I find the result when scale is 100 be very terrible, I am not sure is that correct ?