[MRG] Accelerate example plot_sparse_logistic_regression_mnist.py #21862
Conversation
2ce42ca to
993327c
Compare
993327c to
2b8610e
Compare
|
Using SAGA on a toy dataset defeat the bit the purpose of the example. I don't think that we should do that the Ben if the example is taking 25 secs. We could down sample a bit the original dataset depending of the number of sample. |
|
Thanks for having a quick look, @glemaitre! The example already used only 5000 from the original 70k MNIST samples for training. As the execution time is mainly limited by Sorry, @webbdays, I overlooked you mentioned in #21598 that you are already looking into how to speed up this example. |
|
I agree with @glemaitre that we want to using MNIST in this example. We will have to find a way to accelerate the MNIST dataset loading one way or another but changing the example itself is probably not the solution. Let me close this PR. Thanks anyways for the contribution @steinfurt. |
Reference Issues/PRs
References #21598
What does this implement/fix? Explain your changes.
This replaces the MNIST dataset with the internal toy dataset for hand-written digits, roughly keeping the weight vector sparsity and test score, but reducing the test execution time from ~20 seconds to < 1 second. Visualization of the weight vector may be slightly worse (after all it is only 8x8 digits), but the main features of the digits 0-9 are still recognizable.
Any other comments?
Mainly, the previous implementation was slow because of the
fetch_openmlfunction.I kept the reference to MNIST to comment in/out.