[variational.estimators.tests] automatized benchmarking

[franknoe]

Automatized benchmarking. Please run python variational/estimators/tests/benchmark_moments.py to check efficiency on different platforms. The output looks more or less like below. The important criterion is that the "speed-up" should not be significantly smaller than 1 in any case. The speed-up is the improvement of the "clever" over the trivial implementation. When merged, please run under Linux and Windoof and write the output to a file such that we can compare the performance on different systems.

moments_XX  remove_mean = False sym = False const = False
L, data points  100000  100000  100000
N, dimensions   100 100 100
S, nonzeros 10  90  100
time trivial    0.041   0.041   0.041
time moments_XX 0.041   0.044   0.043
speed-up    1.002   0.936   0.949

moments_XX  remove_mean = False sym = False const = True
L, data points  100000  100000  100000
N, dimensions   100 100 100
S, nonzeros 10  90  100
time trivial    0.045   0.045   0.045
time moments_XX 0.044   0.043   0.041
speed-up    1.028   1.032   1.095

moments_XX  remove_mean = True  sym = False const = False
L, data points  100000  100000  100000
N, dimensions   100 100 100
S, nonzeros 10  90  100
time trivial    0.077   0.077   0.077
time moments_XX 0.040   0.084   0.084
speed-up    1.899   0.910   0.916

moments_XX  remove_mean = True  sym = False const = True
L, data points  100000  100000  100000
N, dimensions   100 100 100
S, nonzeros 10  90  100
time trivial    0.075   0.075   0.075
time moments_XX 0.041   0.087   0.086
speed-up    1.839   0.860   0.868

[franknoe]

Benchmarking ready. @marscher and @fnueske , please try. Careful: this might run an hour or so and for best results the machine is not otherwise used, so a good thing to run before leaving in the evening or for the lunch break.

[marscher]

https://gist.github.com/marscher/b0a07de65bf93d374d73