This is a document for the program of this project. Most files have been well commented, described and orgainized.
Dependencies:
- Python 3.7.3
- PyTorch 1.1.0
- matplotlib 3.1.0
- numpy 1.16.4
- scipy 1.2.1
- scikit-learn 0.21.2
The file train_dct.py and train_seq2seq.py can be run directly. They are the main programs for training and you can modify hyperparameters inside to test different configurations. I tried to add GPU support but failed, so only add some basic GPU-related code and uncomment them.
For other files like data_process.py, token_process.py, eval_dct.py, eval_seq2seq.py,it is recommended to run them using ipython/jupyter kernel in PyCharm or VS Code. So you can do more operations on data. The files dataset.py and models.py contains the core code for the two models. And utils.py contains some useful tools.
The other necessary data files are also included in this folder. With these, you can reproduce most of the experiments in the thesis. Because the time was limited when writing thesis, some visualization was done in jupyter notebook and that code was a messey, hard to organized. So I can only restore some code by memory. Also, I write the code in the end of eval_xxx.py file that saves all outputs of test set, so you can make use of that.
Besides, there are two folders glove and Recordings_October_2014 is not included because they are too large and is public online and on afs, you can make a link to them.
The synthesised animation and survey webpage are in folder subj_eval_case. The webpage is made by very simple HTML. By swap <video> labels, you can shuffle the order of test cases, which must not been seen by participants.
You can contact me by email: j652350094@gmail.com Good luck!
ZihangJiao May 2020