yhldhit / WMMD-Caffe

WMMD-Caffe

This is an implementation of CVPR17 paper "Mind the Class Weight Bias: Weighted Maximum Mean Discrepancy for Unsupervised Domain Adaptation". We fork the repository from mmd-caffe and make our modifications. We show them below:

• mmd layer: the backward function in mmd_layer.cu is adjusted to replace the conventional mmd with weighted mmd specified in the paper
• softmax loss layer: Instead of ignoring the empirical loss on target domain, we modified the function so that logistic loss is added based on pseudo label.
• data layer: we add an parameter to the data label so that the number of classes is conveyed to mmd layer.

The machines configuration that run the experiments are specificied below:

• OS: UNIX
• GPU: TiTan X
• CUDA: version 7.5
• CUDNN: version 3.0

Slight changes may not results instabilities.

Usage

• prepare model: The bvlc_reference_caffenet and bvlc_googlenet are used as initialzation for Alexnet and GoogleNet, respectively. They can be download from here. The model structure is specified in the relative path model/task_name/train_val.prototxt, e.g., when we transfer from amazon to caltech in office-10+caltech-10 dataset, replace the task_name with amazon_to_caltech.
• prepare data: Since we reach the raw images on the disk when training, all the images file path need to be written into the .txt file, kept in data/task_name/ directory. For example, data/amazon_to_caltech/train(value).txt. To constrcuct such txtfile, a python script is offered in file images/code/data_constructor.py. Following are the main steps:
  • change into the directory image/code/: cd image/code
  • python data_constructor.py. You will be asked to input the source and target domain name, e.g., amazon and caltech, respectively.
  • the generated file traina2c.txt and texta2c.txt could be found in the parent directory, copy them into ../data/amazon2caltech/: cp *a2c.txt ../amazon2caltech.
• fine tune a model: To fine tune the model paramenters, run the shell ./model/amazon2caltech/train.sh You will see the accuracy results once the script is finished. The detailed results could be found in files that are stored in the ./log. And the tuned model will be stored in model/amazon2caltech/

Fine tune the model parameters

Three model parameters are tunned in our experiments, i.e., $\lambda$, $\beta$, $lr$ and please refer to the paper for the meaning of these parameters. Manually tuning them in net defination file, i.e., 'trainval.protxt', could be rather tedious. We provide python scripts to help tune these parameters which can be found in directory script_fine_tune/. To correctly run these scripts, please follow the instruction to install the Python runtime library of Protobuf. In addition, the Caffe python's interface should be compiled: make pycaffe.

Citation