1

Currently, I am working on Human Activity Recognition using wearable sensor data (i.e., accelerometer, gyroscope, etc). Now, I am trying to generate some synthetic sensor data from accelerometer (xyz) data.

I used GAN to generate synthetic data set from 3D accelerometer data. However, the result is not good (the generated/fake data not similar to the real data). I have used several sequence models (i.e., LSTM, bi-directional LSTM), but the result is the same. I got a repeating pattern on my fake data. GAN result

Is there any suggestion for this? Some explanation would be much appreciated. Thank you :)

furanzup
  • 91
  • 1
  • 8

1 Answers1

0

Designing and training GANs -specifically for temporal tasks- is a bit tricky. It's generally a good idea to use state-of-the art architectures instead of writing and training your own model.

If you must use your own model, something you might want to try is using depth-wise convolutions (i.e. along the time dimension) instead of LSTMs.

The training of GANs is also a tricky process so it might help to look into some of the implemented codes to get some tips on avoiding mode collapses etc.

Bahman Rouhani
  • 1,139
  • 2
  • 14
  • 33