sal: A Python repository from TristanThrush

The following instructions were tested with OSX and Ubuntu with and without CUDA.

Note that I use Git LFS, so you should clone this repository with git lfs clone.

First, add all of the directories (sal/sal, sal/trained_models, sal/domains) to your Python path. You may also need to install the relevant Python package dependencies for this project. The following commands, entered in a Python 3 terminal, show how to use SAL. Note that everywhere you see <desired model directory>, an absolute path must be typed.

To initialize SAL with self referential operators:

import sal
import manipulation
s = sal.SAL()
s.atoms = manipulation.ManipulationAtoms()
s.problems = manipulation.ManipulationProblems()

To initialize SAL without the self-referential operators:

import sal
import manipulation
s = sal.SAL()
s.atoms = manipulation.ManipulationAtoms()
s.problems = manipulation.ManipulationProblems()
del s.atoms.atoms['then']
del s.atoms.atoms['remember']
del s.atoms.atoms['forget']
del s.atoms.atoms['internalize']
del s.atoms.atoms['externalize']

Note that after SAL's problems are initialized, a simulator visual file will be printed, (such as '.49456.txt'). You will get a FileNotFoundError if you try to access this visualization too soon. But after SAL starts solving problems, you can open up another terminal and enter the following command to see a real time visualization of the problem that SAL is solving:

import manipulation
manipulation.SimulatorVisualizer.visualize('.<number given>.txt')

Trained SAL instances are provided in sal/trained_models/manipulation_self_referential and sal/trained_models/manipulation_not_self_referential. To load a saved SAL instance:

s.load(<desired model directory>)

To evaluate SAL's performance (first making |G(.,.)| and |G(.,.)_e| something reasonable):

s.learner.G = 1000  # Change to 150 if evaluating SAL without self-referential operators
s.learner.GE = 1000  # Change to 150 if evaluating SAL without self-referential operators
s.perform(1)

To train SAL yourself for 40 epochs, followed by saving your model:

s.learner.G = 500  # Change to 150 if training SAL without self-referential operators
s.learner.GE = 2
s.learn(40)
s.save(<desired model directory>)

To run the generalization tests from my thesis, when SAL has the self-referential operators:

import analysis
analysis.test_generalizability(s)

To print a latex-friendly readout of the entire training history of a saved SAL instance:

import analysis
analysis.latexable_innerese_operator_traces(<desired model directory>)

TristanThrush/sal