benchmark Using FP32 for DRQ network.

Moving the temperature numpy log to FP32.

Feb 24 '23 22:02 kulinseth

I am wondering what is the motivation of this change?

If we are changing this, I hope we can update the upstream code at the same time: https://github.com/denisyarats/drq/blob/master/drq.py#L188

Feb 24 '23 23:02 xuzhao9

I am wondering what is the motivation of this change?

If we are changing this, I hope we can update the upstream code at the same time: https://github.com/denisyarats/drq/blob/master/drq.py#L188

Hi @xuzhao9 , the motivation is that Metal doesn't support FP64. And as a result this benchmark fails on MPS device. From the change it seems like by default np.log returns FP64 , if it won't have any impact on the networks performance. It will be nice to just use FP32.

Feb 24 '23 23:02 kulinseth

@kulinseth Since we don't do any end-to-end accuracy checks in the core models set, we need to be very careful when modifying the model code.

Therefore, I believe we can only accept this PR if upstream model code repo uses FP32, otherwise we will keep it as-is.

cc @adnanaziz for ideas and suggestions.

Feb 25 '23 04:02 xuzhao9

Hi @xuzhao9 , that makes sense. I will create an upstream PR to propose similar fix.

Mar 09 '23 18:03 kulinseth