hossen-code / mt-mtda Goto Github PK

This is the official implementation for WACV 2021 paper of Unsupervised Multi-Target Domain Adaptation ThroughKnowledge Distillation,

• pytorch <= 1.4.0
• sacred (https://github.com/IDSIA/sacred)

• You can download Office31 and OfficeHome from:

    Office-31: https://people.eecs.berkeley.edu/~jhoffman/domainadapt/
    Office Home: http://hemanthdv.org/OfficeHome-Dataset/
  

• In order to easily run our code, we expect the datasets for Office31 and OfficeHome to found be at: '~/datasets/', here are how we get the path to our datasets in our code:

    a = os.path.expanduser('~/datasets/amazon/images')
    w = os.path.expanduser('~/datasets/webcam/images')
    d = os.path.expanduser('~/datasets/dslr/images')
  
    Ar = os.path.expanduser('~/datasets/OfficeHome/Art')
    Cl = os.path.expanduser('~/datasets/OfficeHome/Clipart')
    Pr = os.path.expanduser('~/datasets/OfficeHome/Product')
    Rw = os.path.expanduser('~/datasets/OfficeHome/RealWorld')
  

• Originally, this project used a mongoObserver from sacred in combination with Omniboard (https://github.com/vivekratnavel/omniboard) to visualize experiments. You can either use this or integrate your own visualizer (e.g. visdom)

• All the code is run using sacred, you can see examples in the folder experiments/

• Example, to run all of our results on the OfficeHome with AlexNet backbone:

    python experiments/exp_kd_multi_target_da_grl_cst_fac.py
  

• For other experiments you can edit exp_kd_multi_target_da_grl_cst_fac.py to change our hyper-parameters

• Coming soon.

We used some of the code from:

• https://github.com/thuml/CDAN: to load ImageClef dataset
• https://github.com/jindongwang/transferlearning/tree/master/code/deep/TCP: to load Office31 dataset