kejuzhang / dapn

A pytorch implementation of "Domain-Adaptive Few-Shot Learning"

Introduction

The framework is implemented and tested with Ubuntu 16.04, CUDA 8.0/9.0, Python 3, Pytorch 0.4/1.0/1.1, NVIDIA TITANX GPU.

Requirements

• Cuda & Cudnn & Python & Pytorch

  This project is tested with CUDA 8.0/9.0, Python 3, Pytorch 0.4/1.0, NVIDIA TITANX GPUs.

  Please install proper CUDA and CUDNN version, and then install Anaconda3 and Pytorch. Almost all the packages we use are covered by Anaconda.

• My settings

  source ~/anaconda3/bin/activate (python 3.6.5)
    (base)  pip list
    torch                              0.4.1
    torchvision                        0.2.2.post3
    numpy                              1.18.1
    numpydoc                           0.8.0
    numba                              0.42.0
    opencv-python                      4.0.0.21

Data preparation

Download and unzip the datasets: MiniImageNet, TieredImageNet, DomainNet.

Here we provide the datasets of target domain in Google Drive, miniImageNet, tieredImageNet.

Format: (E.g. mini-imagenet)

MINI_DIR/
  --  train/
      --  n01532829/
      --  n01558993/
      ...
  --  train_new_domain/
  --  val/
  --  val_new_domain/
  --  test/
  --  test_new_domain/

Training

First set the dataset path MINI_DIR/, TIERED_DIR/, DOMAIN_DIR/ for the three datasets.

For each dataset, we use its training set to train a pre-trained model (e.g. tiered-imagenet).

cd pretrain/
python -u main_resnet.py --epochs 50 --batch_size 1024  --dir_path TIERED_DIR 2>&1 | tee log.txt &

We then use the corresponding pre-trained model to train on each dataset. (e.g. mini-imagenet)

python -u train_cross.py --gpu_id 0 --net ResNet50 --dset mini-imagenet --s_dset_path MINI_DIR --fsl_test_path MINI_DIR --shot 5 --train-way 16 --pretrained 'mini_checkpoint.pth.tar' --output_dir mini_way_16

Testing

python -u test.py --load MODEL_PATH --root MINI_DIR