TapNet: Neural Network Augmented with Task-Adaptive Projection for Few-Shot Learning

An unofficial Code for ICML 2019 paper TapNet: Neural Network Augmented with Task-Adaptive Projection for Few-Shot Learning in PyTorch.

Original Code (Chainer) : https://github.com/istarjun/TapNet

Dependencies

python == 3.6.13
torch == 1.8.1
torchmeta == 1.7.0
cupy == 9.2.0

Conda enviornment

conda env create -f TapNet-torch.yaml
source activate TapNet

Fix the prefix in the yaml file and use it.

pip dependency

pip install -r requirements.txt

How to run

python run_mini.py		# For mini-ImageNet experiment
python run_tiered.py	# For tiered-ImageNet experiment

Usage

usage: run_{DATASET}.py [-h] [--n_gpu N_GPU] [--use_parallel USE_PARALLEL]
               [--data_root DATA_ROOT] [--dataset DATASET] [--n_shot N_SHOT]
               [--n_class_train N_CLASS_TRAIN] [--n_class_test N_CLASS_TEST]
               [--n_query_train N_QUERY_TRAIN] [--n_query_test N_QUERY_TEST]
               [--dim DIM] [--n_train_episodes N_TRAIN_EPISODES]
               [--n_val_episodes N_VAL_EPISODES]
               [--n_test_episodes N_TEST_EPISODES] [--n_test_iter N_TEST_ITER]
               [--meta_batch META_BATCH] [--wd_rate WD_RATE] [--lr LR]
               [--lr_decay LR_DECAY] [--lr_step LR_STEP]
               [--save_root SAVE_ROOT] [--info INFO]

optional arguments:
  -h, --help            show this help message and exit
  --n_gpu N_GPU         GPU number to use
  --use_parallel USE_PARALLEL
                        Whether to use all GPU
  --data_root DATA_ROOT
  --dataset DATASET     Dataset
  --n_shot N_SHOT       Number of training samples per class
  --n_class_train N_CLASS_TRAIN
                        Number of training classes
  --n_class_test N_CLASS_TEST
                        Number of test classes
  --n_query_train N_QUERY_TRAIN
                        Number of queries per class in training
  --n_query_test N_QUERY_TEST
                        Number of queries per class in test
  --dim DIM             Dimension of features
  --n_train_episodes N_TRAIN_EPISODES
                        Number of train episodes
  --n_val_episodes N_VAL_EPISODES
                        Number of validation episodes
  --n_test_episodes N_TEST_EPISODES
                        Number of test episodes
  --n_test_iter N_TEST_ITER
                        Iteration number for test
  --meta_batch META_BATCH
                        Meta-batch size (number of episodes, but not used in
                        here)
  --wd_rate WD_RATE     Weight decay rate in Adam optimizer
  --lr LR               Learning rate
  --lr_decay LR_DECAY   Use learning rate decay
  --lr_step LR_STEP		Learning rate decaying steps
  --drop_out DROP_OUT	Probability of drop out layer
  --save_root SAVE_ROOT
                        save path for best ckpt
  --info INFO           Additional notes in the experimental name

See function parser_args() in run_mini.py or run_tiered.py file for detailed optional arguments.

Result

*Performance not reproduced*

Omniglot

	20-way 1-shot	20-way 5-shot
Original Paper	98.07%	99.49%
Our

MiniImageNet

	5-way 1-shot	5-way 5-shot
Original Paper	61.65 $\pm$ 0.15%	76.36 $\pm$ 0.10%
Our

TieredImageNet

	5-way 1-shot	5-way 5-shot
Original Paper	63.08 $\pm$ 0.15%	80.26 $\pm$ 0.12%
Our