Domain-specific-Fewshot-Learning

Pytorch implementation of the Trans paper "Few-Shot Learning for Domain-Specific Fine-Grained Image Classification"

Centerloss and CNloss on FashionMNIST

Space distribution

For example, the 2-th category Pullover (green color) and 6-th category Shirt (pink color). Centerloss is difficult to form their own cluster fastly and robustly.

Test on FashionMNIST

Performance in general classification tasks.

python3 test_loss.py

Accuracy	Softmax loss	Center loss	CN loss
（%）	89.5± 0.2	90.0±0.2	91.42±0.3

The three use a unified network and parameter settings.

CUB-200-2011

Usage

Prepare dataset

• Download CUB_200_2011 Dataset. Unzip and locate it in this directory.

  /home/***/datasets/CUB_200_2011/

• If you want the focus-areas generated by us. Download CUB_Attention_Area.

  /home/***/datasets/CUB_Attention/

• If you want the focus-areas generated by yourself. ( Grad-cam and crop-resize operation are about 9 images per second. )

  • Please re-training model.

  • Run

    python3 generate_cub_focus.py
    

Testing

Default testing, only on novel classes and without focus-areas.

cd CUB
python3 test_cub.py 

You can change the switch of "withAtt", "novelonly","SUPPORT_NUM" of test_cub.py.

Re-training

python3 train_cub.py

miniPPlankton

Dataset

• Download miniPPlankton. Unzip and locate it in this directory.

  /home/***/datasets/phytoplankton/fewshot/

  | SmallTrain

  | SmallTest

• Download miniPPlanktonFocusAreas.Unzip and locate it in this directory. (You can also generate them during the training and testing process.)

  /home/***/datasets/phytoplankton/fewshot/

  | SmallTrainAtt

  | SmallTestAtt

Testing

python3 test_plank.py

The experimental results here use a pre-trained ( ImageNet ) model.

Acc(%)	5way-1shot	5way-5shot	10way-1shot	10way-5shot
Imprint	81.55	93.27	71.93	89.65
+CNloss	83.24	93.79	73.76	90.31
+Att	83.35	94.10	73.9	90.45

The model may have reached the data bottleneck, using "Att" only brings a slight boost.

Others

• Please refer to the citations in this paper for miniDogs datasets.
• The center point dimension of loss functions can be set by itself.