chenjiehu / MGCN_ODA

A re-implementation of MGCN

Environment

• python 3.9.7
• pytorch 1.10.1
• pycharm

Instructions

1 Before running you program, the path of the dataset may need to be modified on your computer, and you may need to modify some values of some variables according to your data set in utils.py

2 Split multi-pair shot dataset and the query dataset: run data_split_csv.py

3 For open set dataset, please follow these steps to perform the program:

(1) run train_step1.py (pretrain Conv4)

(2) run train_step2.py (train MGCN)

(3) run test_noDA.py (test MGCN without DA)

(4) run test_DA.py (test MGCN with DA)

You can test the classification accuracy in different situations by modifying the relevant parameters train-way,test-way,shot

The data folder should be structured as follows: (when ths number of common categories is 6)

├── Data/
│   ├── Auxiliary data/
|   |   ├── 1_common category1/
|   |   ├── 2_common category2/
|   |   ├── ....
|   |   ├── 7_private category1/
|   |   ├── 7_private category2/
|   |   ├── ....
│   ├── Task data/
|   |   ├── 1_common category1/
|   |   ├── 2_common category2/
|   |   ├── ....
|   |   ├── 7_private category1/
|   |   ├── 8_private category2/
|   |   ├── ....
│   ├── Val data/
|   |   ├── category1/
|   |   ├── category2/

Please note that the common category folders should be arranged at the front of the private category folders to ensure that the label number of the common category is less than the label number of the private category. And the common category labels of the auxiliary dataset and the task dataset should correspond one by one.

Contact

[e-mail]