Lists and tuples not considered array_like -> ravel doesn't work on them?

[nikihowe]

In pure numpy, np.append(np.array([1, 2]), (3, 4)) gives array([1 2 3 4]).

In jax numpy, jnp.append(jnp.array([1, 2]), (3, 4)) gives TypeError: reshape requires ndarray or scalar arguments, got <class 'tuple'> at position 0.

This is because jnp.append calls jnp.ravel on the (3, 4), which checks (and fails the check) whether the tuple is array_like. Using a list in place of the tuple also fails the test.

My (possibly incorrect) understanding is that array_like in numpy is anything that can be put in a call to np.array. It seems that in jax, array_like is a property held only by ndarrays, scalars, and objects with the '__jax_array__' attribute.

I'm wondering if this is intended behaviour, or if there is a bug here? If the former, then perhaps there should be an explanation in the jax documentation that, unlike in numpy, array_like doesn't include tuples or lists.

[jakevdp]

Thanks for the question! This is an intended departure from numpy's API: in most places JAX does not silently convert iterables to arrays, because this can hide performance issues.

If you'd like to append a tuple to an array, you can call jnp.asarray explicitly; e.g.

tup = (3, 4)
jnp.append(jnp.array([1, 2]), jnp.array(tup))

I think we should add a note about this to the Sharp Bits documentation.

[jakevdp]

Update (with some discussion) in #6835

Edit: update is now live at https://jax.readthedocs.io/en/latest/notebooks/Common_Gotchas_in_JAX.html#non-array-inputs-numpy-vs-jax

[nikihowe]

Thank you!