This is a reimplementation of D2AE in Chainer. The paper title is Exploring Disentangled Feature Representation Beyond Face Identification, which is presented at CVPR 2018 by Liu et al.

The step of training D2AE model is separated into two steps.

We provide a pretrained model on this link and you can skip the pretrain step using it. A main code for training of the second step is main.py script. You can train it by:

python main.py rgb 

To train a new model, use the main.py script.

The command to reproduce the original TSN experiments of RGB modality can be

python main.py rgb 

For flow models:

python main.py flow 

No test code because epick kitchen dataset doesn't have true labels for test data.