/TransMatch_code

Pytorch code for our paper TransMatch: a transfer learning scheme for semi-supervised few-shot learning

Primary LanguagePython

TransMatch

PyTorch implementation of our paper 'TransMatch: a transfer learning scheme for semi-supervised few-shot learning'. The following code is for miniImagenet.

Summary

Requirements

  • Python 3.5
  • Pytorch 0.4+

Download data and pre-trained model

  1. Download miniImageNet: download link

  2. Download pretrained model on base-class data: download link

Then Unzip the MY_mini_data.zip.

The project folder should look like:

main.py
mini_loader.py
wide_models.py
wideresnet_legacy.py
pretrained_model_on_base_class.pth.tar
utils
MY_mini_data
|   base_data
|   novel_data
└─── miniImagenet_base_novel
    └───base
    └───novel

Run experiments on miniImagenet

Please check main.py for the details.

Notice:

  1. The corresponding results are stored in final_result.csv (totally 600 test results)
  2. Imprinting will be stored in imprinting_result.csv (totally 600 test results)
  3. logXXX.png is the plot for the test of 0 to XXX; each line stands for the change of test accuracy over epochs
  4. Fine-tuning for each test experiment may last for 10-20 minutes, so it may take 4-6 days to finish all 600 test experiments. You could also just let the code run for 100 test experiments. The results are similar.

5-way 5-shot 100-unlabel

TransMatch

python main.py --gpu 0 --num-way 5 --num-sample 5 --unlabelnumber 100 --epoch 25 --checkpoint TransMatch_5_5_100

MixMatch

python main.py --gpu 0 --random --num-way 5 --num-sample 5 --unlabelnumber 100 --epoch 25 --checkpoint MixMatch_5_5_100

TransMatch with distractor class

python main.py --gpu 0 --num-way 5 --num-sample 5 --unlabelnumber 100 --distractor --distractor_class 2 --epoch 25 --checkpoint TransMatch_5_5_100_distractor_2

each log file will be like in this format:

Learning_Rate	Train_Loss	Train_Loss_X	Train_Loss_U	Valid_Loss	Valid_Acc
0.001000	0.908974	0.776229	0.026549	1.061434	56.000000	
0.001000	0.768066	0.660366	0.021540	0.905195	66.666667	
0.001000	0.813665	0.708609	0.021011	0.946537	65.333333	
0.001000	0.765839	0.673886	0.018390	0.896166	66.666667	
0.001000	0.705718	0.611745	0.018795	0.878929	69.333333	
...
0.001000	0.639709	0.563627	0.015217	0.803816	72.000000

The change of test accuracy over 25 epochs for 600 experiments will be like (each line represents one experiment)

5-way 1-shot 100-unlabel

TransMatch

python main.py --gpu 0 --num-way 5 --num-sample 1 --unlabelnumber 100 --epoch 20 --checkpoint TransMatch_5_1_100

MixMatch

python main.py --gpu 0 --random --num-way 5 --num-sample 1 --unlabelnumber 100 --epoch 20 --checkpoint MixMatch_5_1_100

TransMatch with distractor class

python main.py --gpu 0 --num-way 5 --num-sample 1 --unlabelnumber 100 --distractor --distractor_class 2 --epoch 20 --checkpoint TransMatch_5_1_100_distractor_2

Performance

No distractor class for unlabeled images

(%) Imprinting MixMatch TranMatch (Ours)
1-shot 58.68 ± 0.81 52.00 ± 1.20 63.02 ± 1.07
5-shot 76.06 ± 0.59 79.97 ± 0.62 81.19 ± 0.59

With distractor classes for unlabeled images

(%) Imprinting MixMatch TranMatch (Ours)
1-shot 58.68 ± 0.81 50.68 ± 1.15 60.41 ± 1.02
5-shot 76.06 ± 0.59 78.07 ± 0.69 79.48 ± 0.64