To turn off forces, force_coefficient=None or force_coefficient=0.?

[Alireza Khorshidi]

In the documentation here, lets make sure that "force_coefficient=None" is the way to turn off force training and not "force_coefficient=0".

[Alireza Khorshidi]

If you run this code, calculator number 3 and 4 will crash

#! /usr/bin/env python

from ase.calculators.emt import EMT
from ase import Atoms, Atom
from amp import Amp
from amp.descriptor.gaussian import Gaussian
from amp.model.neuralnetwork import NeuralNetwork
from amp.model import LossFunction


def generate_data():
    """Generates test or training data."""

    images = [Atoms([Atom('Au', (0, 0., 0.)),
                     Atom('Au', (1.9, 0., 0.))]),
                Atoms([Atom('Au', (0, 0., 0.)),
                     Atom('Au', (6., 0., 0.))]),]

    for atoms in images:
        atoms.set_calculator(EMT())
        atoms.get_potential_energy()
#        atoms.get_forces(apply_constraint=False) 

    return images

if True:
    images = generate_data()

    # force_coefficient is None on a single core
    calc1 = Amp(descriptor=Gaussian(),
           model=NeuralNetwork(hiddenlayers=(2, 2),),
              label='1',
           cores=1)
    calc1.model.lossfunction = LossFunction(force_coefficient=None,
                                            convergence={'energy_rmse': 0.02,
                                                         'force_rmse': None})
    calc1.train(images=images)

    # force_coefficient is 0. on a single core
    calc2 = Amp(descriptor=Gaussian(),
           model=NeuralNetwork(hiddenlayers=(2, 2),),
              label='2',
           cores=1)
    calc2.model.lossfunction = LossFunction(force_coefficient=0.,
                                            convergence={'energy_rmse': 0.02,
                                                         'force_rmse': None})
    calc2.train(images=images)

    # force_coefficient is None on two cores
    calc3 = Amp(descriptor=Gaussian(),
           model=NeuralNetwork(hiddenlayers=(2, 2),),
              label='3',
           cores=2)
    calc3.model.lossfunction = LossFunction(force_coefficient=None,
                                            convergence={'energy_rmse': 0.02,
                                                         'force_rmse': None})
    calc3.train(images=images)

    # force_coefficient is 0. on two cores
    calc4 = Amp(descriptor=Gaussian(),
           model=NeuralNetwork(hiddenlayers=(2, 2),),
              label='4',
           cores=2)
    calc4.model.lossfunction = LossFunction(force_coefficient=0.,
                                            convergence={'energy_rmse': 0.02,
                                                         'force_rmse': None})
    calc4.train(images=images)

[andrew_peterson]

Good catch that it is only a bug on multiple cores. That should make it easier to troubleshoot.

[andrew_peterson]

This bug is in your commit lines @akhorshi. See below; I think line 119 is the culprit, where it sends data to fortran.

Traceback (most recent call last):
  File "/usr/lib/python2.7/runpy.py", line 174, in _run_module_as_main
    "__main__", fname, loader, pkg_name)
  File "/usr/lib/python2.7/runpy.py", line 72, in _run_code
    exec code in run_globals
  File "/home/aap/Dropbox/repositories/Amp/amp/amp/model/__main__.py", line 119, in <module>
    fingerprintprimes,)
  File "/home/aap/Dropbox/repositories/Amp/amp/amp/model/__init__.py", line 996, in ravel_data
    fingerprintprimes)
  File "/home/aap/Dropbox/repositories/Amp/amp/amp/model/__init__.py", line 966, in ravel_neighborlists_and_fingerprintprimes
    for key, derafp in fingerprintprimes[hash].iteritems():
TypeError: 'NoneType' object has no attribute '__getitem__'

But I think we should have a larger conversation about this. I think we have some duplicate code that should be removed. I think (or at least suspect) that we can move the whole if model.fortran clauses from model/__main__.py, and that can be handled in model.LossFunction.get_loss as it is now for fortran-mode on a single core. That is, from the worker's perspective, it is running in serial (cores=1) mode, even though from the master's perspective is is running in parallel. Does that make sense? Probably easier to look at it in person.

What I'm not clear about is why/if this changed when we made it so you can turn off force training by the None keyword.

[andrew_peterson]

I think I figured this out in commit b44ccbe. There are still a couple of lingering issues that I marked with FIXME's though.