- Python (>=3.6)
- Chainer (>=6.3.0)
- cupy
- numpy
Basically, all training settings are handled in the config file.
The config file has a dictionary hierarchy and is parsed according to core/utils/config.py.
Typical properties are described below.
| Property | Description |
|---|---|
| class_equal | Whether or not to include transformation between the same style-class. |
| n_gesture | The number of gesture classes. |
| generator.top / descriminator.top | The number of output channels of the first convolutional layer in the networks. |
| generator.use_sigmoid | Whether or not to apply sigmoid to the output of the generator. |
| display_interval | Interval iterations of print logs on display. |
| preview_interval | Interval iterations of saving preview data. |
| save_interval | Interval iterations of saving models. |
| Property | Description |
|---|---|
| ges_class | The gesture class of source data. |
| target_style | Target style ID (If the style is gesture, use the gesture ID; if it is user, use the user ID.) |
Before training, please edit core/dataset/dataset.py to fit your data.
To train the networks:
python train.py {PARH_TO_CONFIG_FILE}
exsample:
python train.py configs/StarGAN_config.py
Before testing, please edit data_load method in test.py to fit your data.
To transform data with the trained network:
python test.py {PARH_TO_CONFIG_FILE}
exsample:
python test.py configs/StarGAN_config.py
If you find this work useful for your research, please cite our paper:
@article{10.1145/3432199,
author = {Suzuki, Noeru and Watanabe, Yuki and Nakazawa, Atsushi},
title = {GAN-Based Style Transformation to Improve Gesture-Recognition Accuracy},
year = {2020},
issue_date = {December 2020},
volume = {4},
number = {4},
journal = {Proc. ACM Interact. Mob. Wearable Ubiquitous Technol.},
month = dec,
articleno = {154},
numpages = {20},
}