Information-theoretic Regularization for Multi-source Domain Adaptation

Pytorch implementation of MIAN: Multi-source Information-regularized Adaptation Network. Provided as a supplementary code for ICCV 2021. Pytorch version: 1.4.0

Dataset

• We support three multi-domain adaptation datasets:
  • Digits-Five (Peng et al., 2019): MNIST-M, MNIST, SVHN, SYNTH, USPS
  • Office-31 (Saenko et al., 2010): Amazon, Webcam, DSLR
    • Link
  • Office-Home (Venkateswara et al., 2017): Art, Clipart, Realworld, Product
    • Link
• Every datasets should be downloaded in data/ directory following the order below:
  • data/{office/office_home}/{domain}/{class}/{images}
  • Digits-Five require .pkl files.
    • data/digits/{domain}/{train/val}.pkl
  • every words in path should be written in lower-case.
  • Due to memory issue, data and pretrained weights are not provided at this moment.

Train examples

python3 main.py --gpu 0 --task office --target Amazon --exp_name Amazon_test \ 
                        --advcoeff 0.1 --SVD_ld 0.0001 --no_MCD

• advcoeff: $\lambda_0$ in paper
• SVD_ld: $\mu_0$ in paper
  • Set to 0 for Vanila MIAN.
• no_MCD: Run main.py without Maximum Classifier Discrepancy (MCD, Saito et al., 2018)

python3 main.py --gpu 0 --task office --target Amazon --partial_domain Amazon DSLR --exp_name Amazon_test_partial_DSLR \ 
                        --advcoeff 0.1 --SVD_ld 0.0001 --no_MCD

• partial_domain: Specify domains to be utilized. (Includes target domain)