aboleth
aboleth copied to clipboard
gradientinstitute
A bare-bones TensorFlow framework for Bayesian deep learning and Gaussian process approximation
https://arxiv.org/abs/1803.04386 Eqn. (4) could be implemented as an alternative to `_sample_W` in `layers.py`, caveats: - This maybe be hard getting to work with full-covariance weights - This may require the...
At the moment we assume the Dense layers can only work with float32 inputs/distributions. Do we want to support other types?
Add utility for checking compatibility of layer input / output shapes in stack operation
This *may* not be the best framework for it, but it would be a good exercise to see how it looks compared to competing frameworks, e.g. Keras, Edward.
right now we are just testing shapes in many cases, like in `test_dense_embeddings`.
this is because sometimes `None` or `?` dimensions are propagated in unexpected ways, especially when we have reshapes (with -1)!
For high dimensional data, and input data with different types - Low dimensional embedding - Categorical embedding - Joint embedding
Metadata
Owner
Metadata
A bare-bones TensorFlow framework for Bayesian deep learning and Gaussian process approximation