Mike Dusenberry

@niketanpansare Thanks for running this experiment. A 1.7x speedup certainly shows that our current methods are leaving a large amount of performance on the table.

I think it’s quite reasonable to use the same PyPI package and mark it as version 2.0. If we choose to do that, we might want to reconsider whether the...

Here's a runnable snippet using the reparameterized LSTM cell. As Dustin said, `ed.layers.LSTMCellReparameterization` is a drop-in replacement for `tf.keras.layers.LSTMCell`, provided that `model.losses` is used as part of the loss. ```python...

Hi! This should be fixed now as of https://github.com/google/uncertainty-baselines/commit/e1d8f01682074842319f2bb4a329f16edacfe777. Can you try it again? If you still run into issues, please let us know!

Thanks for raising this, @xiaoxiong007! The issue here is that the `baselines` directory is not part of the `uncertainty_baselines` library API. Instead, the `baselines` directory is meant to contain standalone...

+1 I think having some type of `Rank1Tiling` Keras layer to insert at the beginning of the model paired with, say, a `Rank1Marginalize` layer at the end of the model...

Hi @rteja1113. That note is saying that the implementation of `sample_mode` is not (or at least, at the time was not) differentiable w.r.t. the inputs based on how it was...