Installing this package (by running pip install . in the source directory) makes it
possible to run the experiments from the command line with the searchnets command, like so:
$ searchnets train config.iniThe first argument to searchnets is a command, and the second is a config.ini file.
A single config.ini file corresponds to a single "experiment", so it will be used more than once,
each time with a different command, as described in the sections below.
For details of the .ini files, see this page
Typically commands are run in the order they are presented here.
Generates training, validation, and test data sets, and save them in a .gz file to be used
by the train and test commands.
The files used to create the data sets are produced by first running the searchstims package, using
the config.ini files supplied for that package in this repository.
Train neural net architecture, as specified in config.ini file.
Trained checkpoints will be saved in subdirectories created in the directory
specified by the MODEL_SAVE_PATH. There will be one sub-directory created for each number of
epochs specified by the EPOCHS option, e.g. if EPOCHS = 10 there will be a
trained_10_epochs directory, and if EPOCHS = [10, 40], there would be
an additional trained_40_epochs directory.
Within each trained_x_epochs directory there will be one sub-directory for each "replicate"
that contains the actual checkpoint files used to save and load train models.
Test accuracy of trained neural net architecture, as specified in config.ini file
Run all three commands in order: data, then train, then test.