This project is part of a research project conducted by the University of Bologna about energy efficiency in high precision computing. It's been developed as a master degree thesis project.
- a compatible Linux-based distro (Ubuntu 18.04 has been proven to be compatible)
- Python 3.6 or later, latest version here
- Cplex and DOcplex installed
- numpy
- pandas
- tensorflow
- keras
- sklearn
- scipy
- networkx
- matplotlib
- colorama
All of them can be downloaded and installed using pip:
pip install <library_name>
EML is another required library, but since it wouldn't work from PYTHONPATH, it's been encapsulated in this project. To get the latest update of the library refer to the official EMLlib Git repo
The main module is al.py and it can be launched with the command
python3 al.py <arg1> <arg2> ...
There are two mandatory arguments, -bmand-exp`, respectively specifying the benchmark and the target error exponent:
python3 al.py -bm <benchmark_name> -exp <integer>
For a complete list of all parameters names refer to the following list or execute python3 al.py -help.
| parameter name | description | mandatory | default |
|---|---|---|---|
-bm |
Specifies the name of the target benchmark. It has to be one present inside the folder flexfloat-benchmarks |
yes | |
-exp |
Specifies the target error: -exp n is considered as the error |
yes | |
-B |
Specifies the maximum number of bits for a variable | 53 | |
-b |
Specifies the minimum number of bits for a variable | 4 | |
-cfr |
Specifies what type of classifier to use | DT | |
-ds |
Specifies the dataset used for training | 0 | |
-dump |
Specifies the directory where to dump the result of the computation as a json file | None |
|
-et |
Specifies the threshold over which errors are considered too large | 0.9 | |
-limit |
Specifies the orders of magnitude withing which to find the solution, starting from -exp |
0 | |
-manual |
If specified, it enables step-by-step iterations, forcing the user to press Enter between one and the next iteration |
||
-p |
Specifies the probability of changing a single length inside the configuration when generating neighbours | 0.3 | |
-pg |
If specified, the program will print the variable graph after the computation | ||
-reg |
Specifies what type of regressor to use | NN | |
-setsize |
Specifies the number fo entries to sample from the real dataset to use as training set and, in part, as test set (see -setsplit) |
1000 | |
-setsplit |
Specifies the ratio between test set and training set | 0.1 | |
-steps |
Specifies how many more steps steps to run after the first non-generated feasible solution has been found | 5 |
An example run is the following:
python3 al.py -bm convolution -exp 5 -limit 5 -steps 8 -manual -dump ..