MNT speed-up example plot_scalable_poly_kernels.py#21731
Conversation
Corrected a typo in the original file. It read ``featrues`` instead of ``features``
| # much faster if the number of training samples increases. | ||
|
|
||
| N_COMPONENTS = [250, 500, 1000, 2000] | ||
| N_COMPONENTS = [250, 450, 850, 1750] |
There was a problem hiding this comment.
Actually you should not have to redefine this variable. It could be reused from the previous cells.
| @@ -99,7 +99,7 @@ | |||
| # compensate for the stochastic nature of :class:`PolynomialCountSketch`. | |||
There was a problem hiding this comment.
The previous line mentions x5 runs. Indeed we do only x3
| @@ -99,7 +99,7 @@ | |||
| # compensate for the stochastic nature of :class:`PolynomialCountSketch`. | |||
|
|
|||
| n_runs = 3 | |||
There was a problem hiding this comment.
I am thinking that we could maybe avoid several runs and just mention that one should repeat it to have a more stable estimate in practice. I think that this is the only way that we can speed up, even more, the experiment.
@ogrisel WDYT?
There was a problem hiding this comment.
I agree, it sounds unnecessary to me to run this 3 times. We should either do it at least 5 times, or only once, and just say it can be run a few times to get a more reliable result there.
Co-authored-by: Guillaume Lemaitre <g.lemaitre58@gmail.com>
|
Removing multiple run of Linearsvc with PolynomialCountSketch. results are similar but runtime improves 23 +/- 1.56 seconds. with 3 iterations with single iteration |
|
@sveneschlbeck would you have a chance to apply the suggested changes? |
|
Putting this available for any contributor who wants to finish the work. |
|
I am happy to complete the work. |
|
Finished in #22903. Thanks @sveneschlbeck |
#21598 @adrinjalali
Adapted the number of components (
N_COMPONENTS) to make the example quicker while maintaining the same characteristic output graph and message of the example.Before:
After
: