Hi, First of, I would like to apologize for submitting this report late. I forgot the deadline was 10h and I assume full responsibility for this mistake. I promise other reports will be submitted on time.
Here is the progress in the project: Alternative Smart executors https://github.com/gablabc/hpxML/tree/submodules 0- The hpx repository within hpxml has been built with clang6.0.0 and boost1.67 on rostam. Some modifications had to be made concerning missing headers. The loop-convert executable used to extract features of given loops has been built. the path to hpx headers must me set manually when calling the executable but this will be changed further on in the project. 1- A python machine learning repository as been added to hpxml. This repository contains a python script that uses scikit-learn's algorithms on data files. Those algorithms are Support Vector Regression, Neural-Network regression and k-Nearest-Neighbors regression. The current data files used to train the algorithms have been previously generated by Zahra but soon I should be able to train the algorithms using my own data files. To compare the different algorithms, the kfold-cross-validation technique is used. the error chosen to compare the regressions with the test set is the mean absolute error. The algorithm with the lowest error will be chosen. Also, since right now, the target values for chunk_size and prefetching distance are on different scales, there is an option to fit the log(Y) instead of fitting Y dues ensuring that the targets values are on the same scale. 2- A training data repository has been added. The goal of this repository is to generate a framework that anyone can use to automatically generate data. The algorithm folder contains various functions that apply a for_each loop on a given lambda function with different chunk_size values and output the execution time for all the chunk-size candidates. The number of functions will continue to grow as the project moves one. To make data, the user can use the training.txt file which contains a list of the functions and the number of iterations they want to run. Then using the sbatch command sbatch train.sbatch training.txt the user can automatically run all the functions and the results will be written in a data file. What is left to be done in the near future. 1-Currently the data generated in the training repository doesn't extract features using loop-convert. This will have to be added to ensure that the data files contain all the information necessary to train the pythons machine learning algorithms. 2-Once an optimal regression has been found using kfold-cross-validation, the algorithm will be fully implemented in python as a way for me to get familiar with the algorithm. Thank you very much, Once again I would like to apologize for missing the deadline. Gabriel Laberge _______________________________________________ hpx-users mailing list hpx-users@stellar.cct.lsu.edu https://mail.cct.lsu.edu/mailman/listinfo/hpx-users
