Conventional unsupervised domain adaptation (UDA) assumes that training data are sampled from a single domain. This neglects the more practical scenario where training data are collected from multiple sources, requiring multisource domain adaptation. To address this problem, a paper was published Moment Matching for Multi-Source Domain Adaptation in which a dataset called Domain Net, which contain six domains and 0.6 million images distributed among 345 categories, was created. In the same paper they proposed a new deep learning approach, Moment Matching for Multi-Source Domain Adaptation (M3SDA). As part of this project, I have extended work of this paper. I have written code for image’s dataset and changed the distance function used in original paper to calculate loss. Two different distance functions are used to train our model:
- Dynamic Partial Distance Function
- Mahalanobis distance
To run the code
$ ./experiment.sh <parameter1> <parameter2> <parameter3> <parameter4>where parameter1 =Target Domain, parameter2 = max_epoch, parameter3 = GPUID, parameter4 = record_folder
All model weights can be found here
For detail understanding refer documentation