This repository contains code for training neural networks to predict x and y if ln(x+y) is given.
The code here is almost the same as xozero/nn-multiply.
To run the experiments yourself, use main.py:
usage: main.py [-h] [--gendata] [--train] [-c CFG_ID] [-n [NRUNS]]
optional arguments:
-h, --help show this help message and exit
--gendata
--train
-c CFG_ID, --cfg_id CFG_ID
cfg_id
-n [NRUNS], --nruns [NRUNS]
To configure experiments, edit simulations.json. The CFG_ID in the above command usage is the key in the config. To find all keys in the config, run python3 config.py.
Example commands for generating data for configuration 7_high2_2000 and for training a model afterwards:
python3 main.py --gendata -c 10_nodes
python3 main.py --train -c 10_nodes
The generated data for all experiments configured in the checked in simulations.json can be found in datasets, it is all checked in. Thus, the --gendata step can be omitted for those.
The results are stored in the runs directory and the some plots are generated inside the plots directory.
A blog post with details can be found here.