Patrick Kidger

Possibly related: #523

Signatory's `exp` map only takes in a path increment; not a generic logsignature. I'm afraid Signatory does not implement a full logsig-to-sig conversion. As another alternate solution, you could do...

Hey there. Yep, the push to PyPI failed on the last release for some reason. I've not gotten around (and will realistically never get around) to figuring out why that...

Yep, that looks fine. I'd probably neaten it slightly to: ``` sum(x.size for x in jtu.tree_leaves(eqx.filter(pytree, eqx.is_array))) ``` Btw, `equinox.custom_types.PyTree` is private API. The good news is that the [jaxtyping](https://github.com/google/jaxtyping)...

The justification is really just "that is what PyTorch does". (I copied the initialisers over when making Equinox.) If you want a zero bias then this can be done by...

Thanks, I'm glad you like it! Regarding adding this to the docs: yep, this is on the horizon (#185). Regarding making this an additional argument: so far I've been resisting...

So I can think of two non-hacky methods for setting *all* biases to zero. Option 1: ```python def linear(*args, **kwargs): out = eqx.nn.Linear(*args, **kwargs) out = eqx.tree_at(lambda l: l.bias, out,...

Approach 1: the above "Option 2" should work out-of-the-box: it will detect each linear layer as these all have a `bias` attribute. Approach 2: if you wanted, you could also...

Hmm. This works fine on my machine. (CPU, Equinox version 0.7.1; JAX version 0.3.17, jaxlib version 0.3.15) The error itself isn't familiar to me either. This looks like some deeper...

Oh, interesting. Yup, I can reproduce this on my GPU machine too. The bad news is that I don't think this is something that can be fixed quickly. This looks...