aboustati/dgplib

Library for Deep Gaussian Processes based on GPflow

dgplib's Issues

Port to GPflow 2.0

Bugs in DSDGP tests

Multikernel layer object

Predicting function values for all the layers does not work for DSDGP

When running prediction for all the function values;

model.predict_all_layers(Xs, 1)

The following error occurs:

TypeError: Using a `tf.Tensor` as a Python `bool` is not allowed. Use `if t is not None:` instead of `if t:` to test if a tensor is defined, and use TensorFlow ops such as tf.cond to execute subgraphs conditioned on the value of a tensor.

Non-whitened representation of layers is redundant

Sampling from posterior does not work for DSDGP

When predicting samples from the posterior:

model.predict_f_samples(Xs, 1)

The following error occurs:

AttributeError: 'DSDGP' object has no attribute 'num_latent'

Clean up docstrings

Docker image

Tests for DSDGP

Need more comprehensive tests for DSDGP

How to use different input X?

Hi Ayman,

Thank you for the algorithm.

I am wondering if we can use different x as input. The idea is we might get samples from different locations, and we are hoping to do inference on all of them. The example in your notebook seems that it's restricted to use same x. Please let me know if you have solutions for this issue.

Chih-Yu

Prediction with full covariance does not work for DSDGP

When running prediction with with full covariance, I get an InvalidArgumentError .

mu, sigma = model.predict_f_full_cov(Xs)

The error is as follows:

InvalidArgumentError: Cholesky decomposition was not successful. The input might not be valid.
	 [[Node: autoflow/DSDGP/predict_f_full_cov/Cholesky = Cholesky[T=DT_DOUBLE, _device="/job:localhost/replica:0/task:0/device:CPU:0"](autoflow/DSDGP/predict_f_full_cov/transpose_1)]]

This also occurs for predict_all_layers_full_cov method.

How to turn off task-specific component in initial layer.

Hi Ayman,

I am trying to implement one of the architectures you mention in your paper relating to this repository. Specifically I am interested in MTL-DGP* where the task-specific component is turned off.

I am not sure how to do this from reading the code. I understand that in order to switch off the shared components I set the MTL variable to false. Is there a similar way for this to be done for the task specific components?

Thanks

Continous integration with Travis

Do simple_example_multitask.ipynb implement cMDGP from paper?

Hi,
I hope you are doing well. I just wanted to ask, do the multitask examples implement cMDGP from the paper for the asymmetric multi-task learning case? I am a bit confused because, in the paper it says to do cMDGP, "the top layer is set to a single coregionalised GP that outputs all the tasks". But, I couldn't find such definition in the code. But, the plot seems like they don't have data at the same points for each task.
Thank you for your time.

multitask DSDGP model