CosML: Combining Domain-Specific Meta-Learners

Environment Setup

Requirements:

Python 3.6
PyTorch 1.2

cd cosml/
virtualenv --python=/usr/bin/python3.6 dgml_venv_py36
source dgml_venv_py36/bin/activate
pip install -r requirements.txt

Datasets

To download and preprocess the data for the Mini-Imagenet, Cars, CUB, Places, and Plantae datasets, please follow the instructions from Tseng el al. (2020)'s GitHub repo here.

Pre-trained 5w5s CosML Models

Please download and unzip all the models in the cosml_project/cosml/output directory.

Download Models

Please use the links below to download trained models for cross-domain few-shot classification (5w5s) on the specified unseen dataset:

Testing Pre-trained Models:

First, make sure that you are inside the cosml directory:

cd cosml/

Unseen Domain	Accuracy	Command
Cars	60.17 +/- 0.63%	`bash test_pretrained_cosml_cars.sh`
CUB	66.15 +/- 0.63%	`bash test_pretrained_cosml_cub.sh`
Places	88.08 +/- 0.46%	`bash test_pretrained_cosml_places.sh`
Plantae	42.96 +/- 0.57%	`bash test_pretrained_cosml_plantae.sh`

Training CosML Models

First, make sure that you are inside the cosml directory:

cd cosml/

Pre-training

The command below non-episodically trains a model on using the Mini-Imagenet dataset. Note that [PATH TO DATASETS] should be the path to the parent directly where all 5 datasets are stored (it should end in CrossDomainFewShot-master/filelists/)

python pretrain.py --testset cub,cars,places,plantae --data_dir [PATH TO DATASETS]

# e.g.
python pretrain.py --testset cub,cars,places,plantae --data_dir ../../CrossDomainFewShot-master/filelists/

Meta-training

Inside train_cosml.sh, you may wish to update TESTSET and DATADIR depending on which model you want to train and where your downloaded data is stored.

To train a model to perform cross-domain few-shot classification on the unseen dataset Cars, TESTSET should be set to cars. The resulting model is named cosml_miniImagenet_cub_places_plantae_5w5s_conv4_conv2+linear_euclidean, which is located in output/checkpoints/. This model name will be the name you specific in test_cosml.sh if you wish to test this model.

After modifying TESTSET and DATADIR, you can train the model using the following command:

bash train_cosml.sh

Meta-testing

bash test_cosml.sh

Note

This code is built upon the implementation from CloserLookFewShot and CrossDomainFewShot. We would like to thank the authors of CloserLookFewShot and CrossDomainFewShot for kindly making their implementations publicly available.

shumanpng/CosML