search

Sha-Lab / FEAT

The code repository for "Few-Shot Learning via Embedding Adaptation with Set-to-Set Functions"
MIT License
418 stars 84 forks source link

About reproducing result on CUB dataset #22

Closed daiszh closed 4 years ago

daiszh commented 4 years ago

I tried to evaluate the feat on CUB dataset using your provided checkpoint:

python eval_feat.py --model_type ConvNet --dataset CUB --model_path ./saves/FEAT-Models/CUB-Conv-1-Shot-5-Way.pth --shot 1 --way 5 --gpu 0

But only get Test Acc 0.5310 + 0.0022. Is there anything need to be modified to reproduce the result in the paper (68.87%)?

Han-Jia commented 4 years ago

Hi, daiszh,

Sorry for the late reply. I re-run the command on my machine, and it could get similar results as re-produced in the paper.

  1. The saved models are for FEAT_STAR, you can also try eval_feat_star with the model. (I find the weights for feat_star can also apply for FEAT)
  2. I crop the CUB images based on the bounding box as usual. Do you use the original images?
  3. Maybe you can try re-train the model.