This is the repository for the Units-ML code

• Python3
• Pytorch

Datasets:
Omniglot Dataset is included with the code
For MiniImageNet and CifarFS, Download the datasets, extract the datasets and place them in the Folder: datasets/
Dataset structure for the three datasets

Cifarfs

datasets/cifarfs
              ├── test
              ├── train
              └── val

mini-ImageNet

datasets/mini_imagenet_full_size
                ├── test
                ├── train
                └── val

Omniglot

datasets/omniglot_dataset
├── images_background
└── images_evaluation

To run the experiments, follow python3 train_vac_inc_system.py --json_file ABC where ABC denotes the json file for experiment (for_eg. config_units/cifarfs_1_5_0.json) The hyperparameter details are present in config file and can be set accordingly

By default, task selection is done on the basis of vacuous belief and conflicting belief. To do task selection using uncertainty score (lambda * vacuous_belief + (1 - lambda) * conflicting_belief) In the json config file, set "task_sel_with_inc" to True

Alternatively, this can be done directly by python3 train_vac_inc_system.py --json_file json_file --vac_inc_balance vac_bal_value --use_bash "True" --experiment_name name_of_experiment --task_sel_with_inc "True" where json_file is config_units_ts/...

python3 train_vac_inc_system.py --json_file config_basic_units/omniglot_K_N.json

python3 train_vac_inc_system.py --json_file config_basic_units/omniglot_1_5.json The seed values can be specified from the json file.

For other custom experiments (say 7way 3 shot), the json file can be edited

python3 train_vac_inc_system.py --json_file config_basic_units/mini-imagenet_K_N.json

run python3 train_vac_inc_system.py --json_file config_basic_units/cifarfs_K_N.json

python3 train_vac_inc_system.py --json_file config_ts_units/omniglot_K_N.json

The json files can be edited and customized to run Units-ML for any specific setting. For e.g., num_samples_per_class in the json files can be changed to change the number of shots of the task.

This code is built on top of open source MAML code provided at https://github.com/AntreasAntoniou/HowToTrainYourMAMLPytorch