Perhaps I’m missing something, but It looks like there’s a dimensionality disagreement:

Both products of X_t

*W_x and H_t-1*W_h have shapes nxh, yet the biases have a shape 1xh.

Is there an implicit broadcasting being made for the bias terms to enable the summation?

Yes. We also mentioned it in linear regression:

Nonetheless, I’ve just added such explanations: https://github.com/d2l-ai/d2l-en/commit/99b92a706b543cfee03b5f9cd874d4771c97cd37

For optimizing the hyperparams on question 2, do we need to perform k-fold validation (and thus augment train_ch8), or just try out different hypers strait into the train_ch8 function itself?

I think in theory at least it would be correct to optimizer our hypers via the use of hold-out right?

Possible typo at http://d2l.ai/chapter_recurrent-modern/index.html

Furthermore, we will expand the RNN architecture with a single

undirectionalhidden layer that has been discussed so far.

Should it be “unidirectional”?

Yup indeed a typo. Thanks!

A good article on Convex Combinations (mentioned in Section 9.1.1.3 [Hidden State])

This looks lovely, is this wandb?

I think there is a small problem with GRU visualization, -1 should be (x) to h_{t-1} and not the output of tanh gate.