Deep Gaussian Processes with Convolutional Kernels

Important files and scripts

• gpflowrc is a setting file. It helps in fixing float type(extremely useful in case of cholesky error). To fix cholesky error, add jitter or increase it.
• wconv_dgp_train.py file is the one which is needed to be run.
• As file runs, credentials such as nlpp,accuracy and time/epochs will be saved in separate folder(at step size of 5 epochs, user can change that in optimizer function).
• loading_model.py is an example of loading a pretrained model.
• ‘data’ directory contains the data. It contains MNIST dataset and example files for convex non convex dataset.

Saving Model

• Get parameters of a model in a dictionary:

  param_dict = m_dgp2.read_trainables()

• Saving parameters as dictionary into a file using pickle :

  import pickle
  f=open(“file.pkl”,”wb”)
  pickle.dump(param_dict,f)
  f.close()

Loading Model

• Loading dictionary back from pickle file:

  param_dict = pickle.load(open(“file.pkl”,”rb”))

• Assigning parameters to the model using loaded dictionary:

  make a similar model, say, my_model, then do
  my_model.assign(param_dict)

Dependencies

• Python 3
• Tensorflow v1.10
• GPflow v1.1