197 algorithm profiling

[DeepBlockDeepak]

Preliminary profiler notebook within /profiling, which displays top 10 slowest function calls as they occur in either cp_als() or cp_apr(), depending on which algorithm the user wants to test. The ranks list is looped over to test the functions with tensors of ranks 2,3 and finally 4. The data for the tests is read from /profiling/data/*.tns.

To test

The bottom code block allows you to choose which algorithm to test. At the moment, cp_apr takes ~5m30s while cp_als takes about 10s.

profile(test_files, ranks, 'cp_apr')
profile(test_files, ranks, "cp_als")

---

:books: Documentation preview :books:: https://pyttb--199.org.readthedocs.build/en/199/

[dmdunla]

@DeepBlockDeepak This is a great start!

Here are some questions/suggestions:

• Why do you check against ktensor for the import?
• Change the variable name tensor to something that is not a pyttb class.
• Make the current notebook the cp_algorithms_profiling.ipynb file.
• Create an analogous notebook for Tucker algorithms: tucker_algorithms_profiling.ipynb.
• The inputs for cp_apr should include only the *integer* files. See data/README.md regarding this. If cp_apr tries to run on the *continuous* files without error, this should be submitted as an Issue. Only nonnegative integer data should be modeled via cp_apr.

[DeepBlockDeepak]

Thanks!

• I originally was testing on the /tests/data/*.tns files (without any input validation), but forgot to remove the ktensor check once you supplied the /profiling/data/ directory.
• Will do on the rest. Thanks for the feedback!

[dmdunla]

Looking good!

Here are some suggestions:

• Instead of commenting out all but one call in the last cell, put each call in a separate cell
• To get more runtime/computation out of each run, your can change stoptol, maxiters, maxinneriters (where available)
• Also, test the different algorithm choices in cp_apr: mu (default), pqnr, pdnr

[DeepBlockDeepak]

Latest commit includes a new file cp_and_tucker_algorithms_profiling_with_gprof2dot.ipynb where the profiling stats are created in a generated sub-directory /profiling/pstats_files.

Example:

• When profile(test_files, ranks, "hosvd", tol=1e-4) is executed in the notebook, the data-input file tensor_dense_integer_size_1000_3_3_rng_1.tns will be used during computation at some point. The resulting profile stats will be created and saved as ./pstats_files/tensor_dense_integer_size_1000_3_3_rng_1_hosvd.pstats relative to our notebook's location. Notice that the last part of the basename of that filepath includes a reference to the algorithm that was used.
  • Similarly, when cp_als() uses the same input tensor file, the resulting profile stats are saved like: ./pstats_files/tensor_dense_integer_size_1000_3_3_rng_1_cp_als.pstats.

To create corresponding .png files for the pstats data:

• Right now, the unsophisticated way I'm running this is to first navigate to the /profiling directory, then running the following command in Terminal (swap out appropriate algo name for the var algorithm below):

algorithm=<specific_algorithm_name>
mkdir -p gprof2dot_images/${algorithm}
for file in pstats_files/${algorithm}/*.pstats; do
    gprof2dot -f pstats $file \
        | dot -Tpng -o gprof2dot_images/${algorithm}/$(basename $file .pstats).png
done

This generates the /profiling/gprof2dot_images directory, where the created .png profile visuals are stored. I believe I had to install both graphviz to my system and package install gprof2dot beforehand, a la this introduction.

[DeepBlockDeepak]

The current visual profiling output is the result the pstats from testing with cell 11 : profile(test_files, ranks, "tucker_als").

[dmdunla]

@DeepBlockDeepak Can you add the profiling output with images for all 4 algorithms?

[DeepBlockDeepak]

@dmdunla The images exist inside of /profiling/gprof2dot_images/<algo_name> per each algorithm.