(sorry, rain delay till ~11:30am)

@rjnw In the reply that Huang just sent, here’s one bit that might help you debug AugurV2 NB: “Another thing that could be going wrong is that the dtype in the Python is incorrect. For example, if you said doc : Vec Int in the model but in Python what you passed in as doc had dtype=float64, then there would be issues.”

hmm, we are not even specifying the type for doc, let me try with explicit type

no difference after giving dtype

@rjnw I’m not sure what you mean by giving dtype — can you explain by giving the diff for the change you tried just now?

https://github.com/rjnw/hakaru-benchmarks/commit/41897d3121ff12670b44fbe90320bf0ee2e1dab3

I added dtype argument to each np.array function

I see. It’s int32, not int64, right?

I used int32 as the experiments in augur use int32

Can you use gdb or printf in the generated code to make sure it’s accessing the bytes as the intended numbers?