Collaborative Learning with Disentangled Features for Zero-shot Domain Adaptation

[ICCV 2021] This repository is the PyTorch implementation of DF-ZSDA

Collaborative Learning with Disentangled Features for Zero-shot Domain Adaptation
Won Young Jhoo, and Jae-Pil Heo
Sungkyunkwan University

This code is an experimental version for MNIST/EMNIST/Fashion-MNIST experiments and did not be fully optimized and refactored yet. You can further improve performance by modifying some parts of the training code I guess.

Requirements

• torch == 1.10.0
• torchvision == 0.11.1
• Pillow == 7.2.0
• python-opencv
• tqdm

Preparing Datasets

1. Download BSDS500 Dataset

We need BSDS500 dataset for synthesizing the color domain (C).

Download the dataset and unzip it to the 'data/BSDS500' folder.

2. Generate 4 Synthetic Domains

We used Gray(G), Color (C), Edge (E), and Negative (N) domain for our experiments.

sh dataset_gen.sh

Run this script will generate each domain dataset in the 'data' folder.

Training

1. Write Config File

We need config file to start training. The example config file consists option like below.

train:
    name: EFGC  # Experiment result folder name
    batch_size: 64  # Batch size for a each task
    checkpoints_dir: ./checkpoints  #  Root result directory
    gpu_ids: [0]    # GPU ids: e.g. [0] [0,1,2]. use [] for CPU
    epoch_count: 1
    n_epochs: 100   # Number of epochs
    save_weight: true   # Do not save model parameter when set to false

dataset:
    data_root: ./data   # Path to root data directory
    rt_data: EMNIST # ToI dataset name [MNIST | FashionMNIST | EMNIST]
    irt_data: FashionMNIST  # IrT dataset name [MNIST | FashionMNIST | EMNIST]
    s_domain: G # Source domain [G | C | E | N]
    t_domain: C # Target domain [G | C | E | N]
    rt_classes: 27  # Number of classes in ToI
    irt_classes: 10 # Number of classes in IrT
    img_size: 28    # Input image size
    workers: 8  # Number of workers for dataloader
    match_sampling: true    # Enable class-matching batch sampling method 

optimizer:
    lr: 0.0002  # Initial learning rate for adam optimizer
    lr_policy: step
    lr_decay_iters: 30  # Multiply by a gamma every lr_decay_iters iterations
    n_epochs_decay: 30  # Number of epochs to linearly decay learning rate to zero
    beta1: 0.5  # Momentum term of adam

logger:
    print_loss: false   # Whether show loss
    print_freq: 500 # Frequency of showing loss

models:
    FS: SA
    FS_n_layers: 2

2. Start Training

Train the model:

python train.py --config CONFIG_FILE

Citation

@inproceedings{Jhoo_2021_ICCV,
    title     = {Collaborative Learning With Disentangled Features for Zero-Shot Domain Adaptation},
    author    = {Jhoo, Won Young and Heo, Jae-Pil},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year      = {2021}
}