This is the implementation of ICML paper "Learning Transferable Features with Deep Adaptation Networks". We fork the repository with version ID c6414ea from Caffe and make our modifications. The main modifications are listed as follow:
- Change the label from a single integer to a tuple, which contains the origin label and an indicator to distinguish source and target
- Add mmd layer described in the paper to neuron layers
In models/icml/amazon_to_webcam_common_setting, we give an example model based on Alexnet. In this model, we insert mmd layers after inception fc7 and fc8 individually.
The bvlc_reference_caffenet is used as the pre-trained model.
In data/office/amazon_to_webcam_common_setting/*.txt, we give an example to show how to prepare the train and test data file. In this task, amazon dataset is the source domain and webcam dataset is the target domain. For training data in source domain, it's label looks like (-1, origin_label). For training data in target domain, it's label looks like (origin_label, -1). For testing data, it's label looks like (origin_label, -1). Integer -1 is used to distinguish source and target during training and testing, and origin_label is the acutal label of the image.
For semi-supervised tasks, the label of labeled target data looks like (I, origin_label). I is any positive integer (larger than the number of classes) and should be added to ignore_label of entropy loss layer.
In mmd-layer, parameter iter_of_epoch should be set to tell mmd layer when to update data. Practically, iter_of_epoch = (source_num + target_num) / batch_size. Users only need to tune mmd_lambda parameter.
We use CGAL to solve Quadratic Programming problem when updating beta. Please install it before compiling this code. We add CGAL as LIBRARIES in Makefile. However, if you prefer to compile caffe with cmake, maybe you should config CGAL yourself.