Use infoGAN to generate multi-dimentional temporal sequences.
In MNIST, continuous noise can represent character width and rotation angle. But in here, their meanings are not quite intuitive.
In MNIST, modifying continuous noise would not change the classtype of generated samples. But in here, it happens.
See codes for MNIST at https://github.com/SongDark/GAN_collections.
Name | Link | Class | Dimension | Train Size | Test Size | Truncated |
---|---|---|---|---|---|---|
CharacterTrajectories | Download | 20 | 3 | 1422 | 1436 | 182 |
Unzip CharacterTrajectories.zip at data/CharacterTrajectories
, then run dataprocess.py
.
python dataprocess.py
Epoch 0 | Epoch 200 | Epoch 500 |
---|---|---|
Epoch 0 | Epoch 200 | Epoch 500 |
---|---|---|
It seems that the generated sequences are not corresponding to their one-hot labels.
label=14(s) | label=16(v) | label=16(y) |
---|---|---|
https://github.com/hwalsuklee/tensorflow-generative-model-collections/blob/master/infoGAN.py
https://github.com/buriburisuri/timeseries_gan/blob/master/train.py