Torchish.__init__ take np.ndarray?

[mavenlin]

As jax is transparent with numpy array, I suppose this shouldn't have a problem?

[samuela]

when would this be necessary? ie shouldn't Torchish.__init__ only be called by the torch2jax harness code when tracing pytorch code? there's no material effect to the public api iiuc

[mavenlin]

What I actually ran into is that sometimes torch requires int64 but Jax automatically casts to int32, which makes the test cases fail. So I'm thinking of passing them numpy int64 arrays so that each can do their own casting.

But putting that aside, a Jax function can always consume numpy arrays, therefore there could be expectations that the functions generated by t2j can do the same.

[samuela]

But putting that aside, a Jax function can always consume numpy arrays, therefore there could be expectations that the functions generated by t2j can do the same.

ah good point, we should try to match what jax supports in terms of function args. ok i'm supportive

[samuela]

fixed in #28