Three `RuntimeWarning`s when running the tutorial experiment

Monty Code
1

when running

% python benchmarks/run.py -e randrot_noise_10distinctobj_surf_agent
# or
% python benchmarks/run_parallel.py -e randrot_noise_10distinctobj_surf_agent

the following run-time errors can be observed:

.../tbp/monty/frameworks/models/motor_policies.py:1259: RuntimeWarning: divide by zero encountered in scalar divide
  return -np.degrees(np.arctan(x / z))
.../numpy/core/fromnumeric.py:3464: RuntimeWarning: Mean of empty slice.
  return _methods._mean(a, axis=axis, dtype=dtype,
.../numpy/core/_methods.py:192: RuntimeWarning: invalid value encountered in scalar divide
  ret = ret.dtype.type(ret / rcount)

(running on an M2 Macbook Air / 16gb)

as these might affect numeric results, have you observed these on other machines?

2

Hi @DLed We have been observing these warnings as well but so far haven’t gotten around to fixing them. They haven’t been our top priority since they are just warnings, not errors. But if you see the cause & solution for them, please feel free to open a PR :slight_smile:

1 Like
3

the division by zero is gone for the scenario

Pull request: Fix division by zero when running the tutorial experiment by d-led · Pull Request #109 · thousandbrainsproject/tbp.monty · GitHub

as the overall/percent_correct 100 seems to be better than in my previous runs, I’d expect an improvement in your benchmarks with this code. If the orientation angle is important, many previous benchmarks might have had a built-in error.

Interestingly, at first, I had a np.sign(x) instead of np.sign(y) for the vertical case, and the percent_correct was 100. This probably hints at more subtle suboptimal results. I’d check whether the formula makes sense.

Upd: I ran a couple of larger benchmarks. The improvements are in 1-2% percent correct range. This is probably understandable as the error is triggered only in a corner case. One could try a mini-benchmark (dependency-free) to see if inverting the conditional wins any speed improvement

3 Likes