We present enhanced Gaussian Process Dynamical Model (EGPDM), a Bayesian method used for capturing high-dimension data's dynamics and transfer learning ability. Our code is based on PyTorch with CUDA supported.

We also provide an original GPDM and its tutorial in IceLab's repo.

• add EGPDM model
• update: calculate the original RMSE instead of normalized
• fix bugs: 3D-trajectory plot; kernel noises
• fix bugs: cycle limitation in LBFGS

• NumPy
• Pandas
• Matplotlib
• SciPy
• scikit-learn
• PyTorch

We provide several processed NASA batteries' data (B. Saha and K. Goebel, 2007) for demonstration. The original and whole datasets we used in the paper can be downloaded from: NASA dataset and Oxford dataset.

The script train_tran.py is a ready-to-run demo. We demonstrate the GPDM training and testing process along with the transfer learning ability. Please refer to the detailed comments in the code.

Try to setup different hyperparameters to evaluate the model:

• D: observation space dimension (determined by observation data)
• Q: desired latent space dimension, empirically $Q << D$
• dyn_target: full or delta, delta models higher order feature by defining latent points as $x_t - x_{t-1}$

We offer linear, RBF, Matern3 and Matern5 kernel functions in the code. You can custom your desired kernels in self.observationGP_kernel() and self.dynamicGP_kernel() using linear combination of kernels.

Note that our model initializes most of the learnable parameters to $1$. If you want a more random initialization, set the parameters to torch.randn() and use a random seed to control.

This project is licensed under the GNU General Public License v3.0.

This work partly uses the code from CIGP and CGPDM.