def FindMinimumInversePower(foldname):
""" Finds the true minimum corresponding:
Note Gradient Descent step should be adequately small for this to work
"""
foldpath = BASE_DIRECTORY + "/" + foldname
sysparams = load_params(foldpath)
(hs_radii, initial_coords, box_length) = load_secondary_params(foldpath)
box_length = float(box_length)
boxv = [box_length] * sysparams.ndim.value
potential = InversePower(
sysparams.power.value,
sysparams.eps.value,
use_cell_lists=False,
ndim=sysparams.ndim.value,
radii=hs_radii * 1.0,
boxvec=boxv,
)
print(box_length, "box_length")
print(len(initial_coords))
# initial_coords = np.loadtxt(foldpath + '/coords_of_minimum.txt',
# delimiter=',')
print(initial_coords, "initial_coords")
# ret = steepest_descent(initial_coords, potential, dx=1e-4)
# ret = modifiedfire_cpp(initial_coords, potential, iprint=1, tol=1e-4)
# E, V = print(potential.getEnergyGradient(initial_coords))
# ret = quench_mixed_optimizer(potential, initial_coords, conv_tol=1e-100)
# ret = quench_mixed_optimizer(potential, initial_coords, conv_tol=1e-100, nsteps=3000)
# ret = quench_steepest(
# potential,
# initial_coords, # make sure right coords are being passed
# stepsize=0.0001,
# nsteps=1000,
# tol=1e-7)
# ret = quench_cvode_opt(initial_coords, potential, )
ret = quench_cvode_opt(initial_coords,
potential,
tol=1e-9,
rtol=1e-10,
atol=1e-10)
finalcoords = ret.coords
print(ret.coords)
# np.savetxt(foldpath + '/coords_of_minimum.txt', finalcoords, delimiter=',')
print(ret)
# print(ret.coords)
# print(ret2.coords)
# print(ret2.coords-ret.coords)
print(finalcoords)
E, V, H = potential.getEnergyGradientHessian(finalcoords)
print(np.linalg.eigvals(H), "all eigenvalues")
def setup_inverse_power(SysParams, radii, box_length):
""" sets up the inverse power potential
Parameters
----------
SysParams : Enum
Parameters for the system we wish to study
radii : np.ndarray
radii of particles in the system
box_length : float
length of the box
"""
# check that the setup is for a dict
if type(SysParams) is not dict:
raise TypeError("system parameters need to be in dictionary form")
box_vec = np.array([box_length] * int(SysParams["ndim"]))
print(SysParams["use_cell_lists"])
potential = InversePower(
SysParams["power"],
SysParams["eps"],
use_cell_lists=SysParams["use_cell_lists"],
ndim=SysParams["ndim"],
radii=radii * 1.0,
boxvec=box_vec,
)
return potential
def test_same_minima_InversePower(self):
nparticles = 32
hs_radii = 0.05 * np.random.randn(nparticles) + 1
volpart = np.sum(4. / 3. * np.pi * hs_radii**3)
phi = 0.7
boxl = (volpart / phi)**(1 / 3.)
boxv = [boxl, boxl, boxl]
coords = np.random.rand(nparticles * 3) * boxl
ncellx_scale = get_ncellsx_scale(np.ones(nparticles), boxv)
bdim = 3
pot_cellists = InversePower(2.5,
1,
hs_radii,
ndim=3,
boxvec=boxv,
use_cell_lists=True,
ncellx_scale=ncellx_scale)
pot_no_cellists = InversePower(2.5,
1,
hs_radii,
ndim=3,
boxvec=boxv,
use_cell_lists=False,
ncellx_scale=ncellx_scale)
nsteps = 1000
tol = 1e-5
res_cell_lists = lbfgs_cpp(coords,
pot_cellists,
nsteps=nsteps,
tol=tol)
res_no_cell_lists = lbfgs_cpp(coords,
pot_no_cellists,
nsteps=nsteps,
tol=tol)
fcoords_cell_lists = res_cell_lists.coords
fcoords_no_cell_lists = res_no_cell_lists.coords
# self.assertEqual(fcoords_no_cell_lists,fcoords_cell_lists)
self.assertTrue(np.all(fcoords_no_cell_lists == fcoords_cell_lists))
def quench_single_inverse_power(coord_file_name, foldpath, sub_fold_name,
optimizer, opt_param_dict):
"""
figures out the minimum correspoding to a set of particle coords
Parameters
----------
coord_file_name: string
name of the path to the coordinates
foldername: str
folder definining the run
sub_fold_name:
name of subfolder where the run data is stored
optimizer: optimizer
quench
opt_param_dict: dict
dictionary of parameters for the optimizer
"""
sysparams = load_params(foldpath)
# path to quench coords
quench_coords_path = (foldpath + "/" + sub_fold_name + "/" + "ensemble/" +
coord_file_name)
quench_coords = np.loadtxt(quench_coords_path)
radii = get_hs_radii(foldpath, sub_fold_name)
box_length = get_box_length(radii, sysparams.ndim.value,
sysparams.phi.value)
boxv = [box_length] * sysparams.ndim.value
ncellx_scale = get_ncellsx_scale(radii, boxv)
potential = InversePower(
sysparams.power.value,
sysparams.eps.value,
use_cell_lists=False,
ndim=sysparams.ndim.value,
radii=radii * 1.0,
boxvec=boxv,
)
boxv = [box_length] * sysparams.ndim.value
# ncellx_scale = get_ncellsx_scale(radii, boxv)
print(potential.getEnergy(quench_coords))
ret = optimizer(quench_coords, potential, **opt_param_dict)
try:
ret = optimizer(quench_coords, potential, **opt_param_dict)
except:
print("exception occured")
# if exception occurs, treat as failure. this is for rattlers
# not enough failures occur that it would make a difference to not just assume this never happens
# but we shoudl switch this out
# but jic
return (quench_coords, False, 0, 0, 0, 0, 0)
# This exists because some runs don't have hessian evaluations
try:
ret["nhev"]
except:
ret["nhev"] = 0
# mixed optimizer statistics
try:
ret["n_phase_1"]
except:
ret["n_phase_1"] = 0
# mixed optimizer statistics
try:
ret["n_phase_2"]
except:
ret["n_phase_2"] = 0
print(ret.coords - quench_coords)
print(opt_param_dict)
results = (
ret.coords,
ret.success,
ret.nfev,
ret.nsteps,
ret.nhev,
ret.n_phase_1,
ret.n_phase_2,
)
print(quench_coords_path)
return results
0.13492935,
5.55656566,
5.19310115,
5.80610666,
6.53090015,
5.3332587,
3.07972527,
5.20893375,
]
test_radii = np.array(test_radii)
test_coords = np.array(test_coords)
box_length = 6.673592625078725
potential = InversePower(
2.5,
1.0,
use_cell_lists=False,
ndim=2.0,
radii=test_radii,
boxvec=[box_length, box_length],
)
test_gradient = np.copy(test_coords)
potential.getEnergyGradientInPlace(test_coords, test_gradient)
grad = potential.getEnergyGradient(test_coords)
print(test_gradient)
print(grad)
# np.savetxt('energy.csv', np.array([potential.getEnergy(test_coords)]), delimiter=',')
# np.savetxt('gradient.csv', potential.getEnergyGradient(test_coords)[1], delimiter=',')
# np.savetxt('hessian.csv', potential.getEnergyGradientHessian(test_coords)[2], delimiter=',')
def get_potential(self):
return InversePower(self.power, self.eps, self.radii, boxvec=self.boxvec, **self.potential_kwargs)
3.16224709493039,
]
ctol = 1e-2
ndim = 2
base_dir = "/home/praharsh/Dropbox/research/bv-libraries/basinerror/datainv"
foldnameInversePower = "ndim=2phi=0.9seed=0n_part=16r1=1.0r2=1.4rstd1=0.05rstd2=0.06999999999999999use_cell_lists=0power=2.5eps=1.0"
data_loc = base_dir + "/" + foldnameInversePower
minima_database_path = data_loc + "minima_database.npy"
(hs_radii, initial_coords, box_length) = load_secondary_params(data_loc)
sysparams = load_params(data_loc)
potential = InversePower(
sysparams.power.value,
sysparams.eps.value,
use_cell_lists=False,
ndim=sysparams.ndim.value,
radii=hs_radii * 1.0,
boxvec=[box_length, box_length],
)
minima_container_debug = CheckSameMinimum(
ctol,
ndim,
boxl=box_length,
minimalist_max_len=2000,
minima_database_location=minima_database_path,
update_database=True,
rattler_check=True,
potential=potential,
hs_radii=hs_radii,
)
def compare_runs_2d(fnames, foldpath, subfoldname, run_a, run_b, ctol):
""" compares runs on points with names fnames between
runs done with 2 minima find routines run_a and run_b
Args:
fnames filenames
run_a run a
run_b run b
Returns:
percentage of minima that are the same
"""
# TODO: replace this
sysparams = load_params(foldpath)
radii = get_hs_radii(foldpath, subfoldname)
box_length = get_box_length(radii, sysparams.ndim.value,
sysparams.phi.value)
subfoldpath = foldpath + "/" + subfoldname
data_path_a = subfoldpath + "/" + run_a
data_path_b = subfoldpath + "/" + run_b
potential = InversePower(
sysparams.power.value,
sysparams.eps.value,
use_cell_lists=False,
ndim=sysparams.ndim.value,
radii=radii * 1.0,
boxvec=[box_length, box_length],
)
minima_checker = CheckSameMinimum(ctol,
dim=2,
boxl=box_length,
hs_radii=radii,
potential=potential)
same_minimum_check_l = []
for fname in fnames:
print(fname)
minimum_a = np.loadtxt(data_path_a + "/" + fname, delimiter=",")
minimum_b = np.loadtxt(data_path_b + "/" + fname, delimiter=",")
boxed_minimum_a = minima_checker.box_reshape_coords(minimum_a)
boxed_minimum_b = minima_checker.box_reshape_coords(minimum_b)
# get first index that is not a rattler
rattlers_exist, rattlers = minima_checker._find_rattlers(minimum_a)
print(rattlers)
# number of non rattlers to ensure we're not in a fluid state
n_non_rattlers = np.count_nonzero(rattlers)
# only do calculations if all particles aren't rattlers
if n_non_rattlers != 0:
first_non_rattler = (np.argwhere(rattlers != 0).T)[0, 0]
# TODO: rewrite the CheckSameMinimum function
# load and make sure the particle being aligned is not a rattler later
# we're choosing -1 because that's always going to be of radius 1.4
# particle
aligned_minimum_b = minima_checker.align_structures(
boxed_minimum_a, boxed_minimum_b, part_ind=first_non_rattler)
same_minimum_check = minima_checker.check_same_structure(
aligned_minimum_b, boxed_minimum_a, rattlers)
same_minimum_check_l.append(same_minimum_check)
fraction_same_minimum = np.mean(same_minimum_check_l)
print(fraction_same_minimum)
return fraction_same_minimum
def map_binary_inversepower(
foldername,
particle_coords,
optimizer,
parameter_dict,
random_coord_0=0,
random_coord_1=-1,
z=0,
):
"""
Finds whether a point defined by particle_coord
on the meshgrid correspond to a minimum or not for a 2d
case.
"""
foldpath = BASE_DIRECTORY + "/" + foldername
# import params
sysparams = load_params(foldpath)
(hs_radii, initial_coords, box_length) = load_secondary_params(foldpath)
assert sysparams.ndim.value == 2
minimum_coords = np.loadtxt(foldpath + "/coords_of_minimum.txt", delimiter=",")
quench_coords = initial_coords.copy()
if len(quench_coords) == 16:
quench_coords = (
quench_coords
+ particle_coords[0] * VEC_8_0
+ particle_coords[1] * VEC_8_1
+ z * VEC_8_2
)
elif len(quench_coords) == 32:
quench_coords = (
quench_coords
+ particle_coords[0] * VEC_16_0
+ particle_coords[1] * VEC_16_1
+ z * VEC_16_2
)
elif len(quench_coords) == 64:
quench_coords = (
quench_coords
+ particle_coords[0] * VEC_16_0
+ particle_coords[1] * VEC_16_1
+ z * VEC_16_2
)
else:
raise Exception("error other random coords have not been generated")
print(quench_coords, "quench")
# print(quench_coords, 'quench coords')
# box length
box_length = float(box_length)
boxv = [box_length] * sysparams.ndim.value
ncellx_scale = get_ncellsx_scale(hs_radii, boxv)
potential = InversePower(
sysparams.power.value,
sysparams.eps.value,
use_cell_lists=False,
ndim=sysparams.ndim.value,
radii=hs_radii * 1.0,
boxvec=boxv,
)
# ret = quench_mixed_optimizer(potential,
# quench_coords, # make sure right coords are being passed
# T=10,
# step=1,
# nsteps=100000,
# conv_tol=1e-8,
# tol=1e-6, rtol=1e-4, atol=1e-4)
# ret = quench_steepest(
# potential,
# quench_coords, # make sure right coords are being passed
# nsteps=2000000,
# stepsize=5e-3, # for steepest descent step size should be small
# tol=1e-4)
# ret = quench_cvode_opt(potential, quench_coords, tol=1e-6, rtol=1e-4, atol=1e-4)
try:
ret = optimizer(quench_coords, potential, **parameter_dict)
except:
print(quench_coords, "failed here")
print(initial_coords, "coords")
print(len(quench_coords))
raise Exception("failure")
# ret = lbfgs_cpp(quench_coords, potential, tol=1e-8, M=1)
# This exists because some runs don't have hessian evaluations
try:
ret["nhev"]
except:
ret["nhev"] = 0
coordarg = 0
results = (ret.coords, ret.success, coordarg, ret.nfev, ret.nsteps, ret.nhev)
# the reason the potential is being passed is because quench coords needs the potential to figure out what to do
return results
def map_pointset_loop_xy(
foldname,
pointset,
optimizer,
parameter_dict,
ctol=1e-2,
ndim=2,
use_minima_database=True,
minima_database_path=None,
coord_arg_0=0,
coord_arg_1=1,
z=0,
):
""" Checks a bunch of points if they match to a minimum by using a for loop
"""
is_same_minimum_list = []
resultlist = []
foldpath = BASE_DIRECTORY + "/" + foldname
sysparams = load_params(foldpath)
(hs_radii, initial_coords, box_length) = load_secondary_params(foldpath)
minimum_coords = np.loadtxt(foldpath + "/coords_of_minimum.txt", delimiter=",")
# Initialize CheckSameMinimum
potential = InversePower(
sysparams.power.value,
sysparams.eps.value,
use_cell_lists=False,
ndim=sysparams.ndim.value,
radii=hs_radii * 1.0,
boxvec=[box_length, box_length],
)
minima_container = CheckSameMinimum(
ctol,
ndim,
boxl=box_length,
minimalist_max_len=200000,
minima_database_location=minima_database_path,
update_database=True,
rattler_check=True,
potential=potential,
hs_radii=hs_radii,
)
if use_minima_database == True:
try:
minima_container.minimalist = [
minima_container.box_reshape_coords(x)
for x in np.load(minima_database_path)
]
except:
print("warning no minima data found. generating")
minima_container.minimalist = [
# minima_container.box_reshape_coords(minimum_coords)
]
nfevlist = []
nstepslist = []
nhevlist = []
for index, point in enumerate(pointset):
res = map_binary_inversepower(
foldname,
point,
optimizer,
parameter_dict,
random_coord_0=coord_arg_0,
random_coord_1=coord_arg_1,
z=z,
)
minima_container.add_minimum(res[0], point, res[2])
# print(index)
# print(minima_container.nrattlermin, 'nrattlermin')
# print(minima_container.nfluidstates, 'nfluidstates')
nfevlist.append(res[3])
nstepslist.append(res[4])
nhevlist.append(res[5])
# print(np.average(nfevlist), 'number of function evaluations')
# print(np.average(nstepslist), 'number of steps')
# print(np.average(nstepslist), 'number of steps')
# print(np.average(nhevlist), "number of hessian evaluations")
# print(minima_container.orderparamlist)
foldpathdata = foldpath + "/" + QUENCH_FOLDER_NAME + "/z_data_30_l6/" + str(z)
os.makedirs(foldpathdata, exist_ok=True)
minima_container.dump_map(foldpathdata)
run_diagnostics = {}
run_diagnostics["nfev"] = float(np.average(nfevlist))
run_diagnostics["nhev"] = float(np.average(nhevlist))
run_diagnostics["nsteps"] = float(np.average(nstepslist))
# print(minima_container.initial_coords_list)
# print(minima_container.orderparamlist)
# print(minima_container.orderparamlist)
return run_diagnostics, is_same_minimum_list, resultlist
def map_binary_inversepower(
foldername,
particle_coords,
optimizer,
parameter_dict,
random_coord_0=0,
random_coord_1=-1,
z=0,
index=None,
mesh_length=None,
):
"""
Finds whether a point defined by particle_coord
on the meshgrid correspond to a minimum or not for a 2d
case.
"""
foldpath = BASE_DIRECTORY + "/" + foldername
# import params
sysparams = load_params(foldpath)
(hs_radii, initial_coords, box_length) = load_secondary_params(foldpath)
assert sysparams.ndim.value == 2
quench_coords = initial_coords.copy()
print("initial_coords", initial_coords)
print(particle_coords[0], particle_coords[1],
"vector coefficients as passed")
print("z value")
quench_coords = (quench_coords + particle_coords[0] * VEC_16_0 +
particle_coords[1] * VEC_16_1 + z * VEC_16_2)
# TODO: save this as a unit meshgrid
# print(quench_coords, 'quench coords')
# box length
box_length = float(box_length)
boxv = [box_length] * sysparams.ndim.value
ncellx_scale = get_ncellsx_scale(hs_radii, boxv)
print(hs_radii)
potential = InversePower(
sysparams.power.value,
sysparams.eps.value,
use_cell_lists=False,
ndim=sysparams.ndim.value,
radii=hs_radii * 1.0,
boxvec=boxv,
)
# save potential parameters
# writing this out since Enum to dict conversion does not give the mapping we want
potential_params_fname = "potential_params.yaml"
potential_param_dict = {
"ndim": sysparams.ndim.value,
"phi": sysparams.phi.value,
"seed": 0,
"n_part": sysparams.n_part.value,
"r1": sysparams.r1.value,
"r2": sysparams.r2.value,
"rstd1": sysparams.rstd1.value,
"rstd2": sysparams.rstd2.value,
"use_cell_lists": int(False),
"power": sysparams.power.value,
"eps": sysparams.eps.value,
}
param_str = generate_param_str_from_dict(potential_param_dict)
# make directory to emulate structure by refactored code
emulation_folder = COMPARE_FOLDER_NAME + "/" + param_str
# add README to emulation folder
os.makedirs(emulation_folder, exist_ok=True)
with open(COMPARE_FOLDER_NAME + "/README.org", "w") as f:
f.write(" \#+AUTHOR Praharsh Suryadevara .\n")
f.write("* Read First.")
f.write(
"This dirctory has been auto-generated by old code before refactoring. Do not change.\n"
)
# emulates path of the refactored code
ensemble_folder_path = emulation_folder + "/ensemble/random_plane"
os.makedirs(ensemble_folder_path, exist_ok=True)
# 0 to emulate the addition of seed
sec_param_folder_path = emulation_folder + "/sec_params/0"
os.makedirs(sec_param_folder_path, exist_ok=True)
old_code_results_folder_path = emulation_folder + "/old_data"
os.makedirs(old_code_results_folder_path, exist_ok=True)
with open(emulation_folder + "/params.yaml", "w") as param_file:
yaml.dump(potential_param_dict, param_file)
opt_param_fname = "minima_finder_params.yaml"
with open(emulation_folder + "/" + opt_param_fname, "w") as param_file:
yaml.dump(parameter_dict, param_file)
initial_coords_fname = str(index) + ".txt"
np.savetxt(ensemble_folder_path + "/" + initial_coords_fname,
quench_coords,
delimiter=",")
mesh_coords_fname = str(index) + "_mesh.txt"
np.savetxt(ensemble_folder_path + "/" + mesh_coords_fname,
particle_coords,
delimiter=",")
if mesh_length is not None:
# 2 *[particle_coords[0], particle_coords[1]]/mesh_length -0.5 form our unit mesh_grid
# centered around [0, 0]
mesh_coords = (2 * np.array([particle_coords[0], particle_coords[1]]) /
mesh_length - 0.5)
mesh_coords_fname = str(index) + "_mesh.txt"
np.savetxt(ensemble_folder_path + "/" + mesh_coords_fname,
mesh_coords,
delimiter=",")
np.savetxt(
sec_param_folder_path + "/" + "initial_coords.txt",
initial_coords,
delimiter=",",
)
np.savetxt(
sec_param_folder_path + "/" + "box_length.txt",
np.array([box_length]),
delimiter=",",
)
np.savetxt(sec_param_folder_path + "/" + "hs_radii.txt",
hs_radii,
delimiter=",")
# ret = quench_mixed_optimizer(potential,
# quench_coords, # make sure right coords are being passed
# T=10,
# step=1,
# nsteps=100000,
# conv_tol=1e-8,
# tol=1e-6, rtol=1e-4, atol=1e-4)
# ret = quench_steepest(
# potential,
# quench_coords, # make sure right coords are being passed
# nsteps=2000000,
# stepsize=5e-3, # for steepest descent step size should be small
# tol=1e-4)
# ret = quench_cvode_opt(potential, quench_coords, tol=1e-6, rtol=1e-4, atol=1e-4)
try:
ret = optimizer(quench_coords, potential, **parameter_dict)
except:
raise Exception("failure")
# ret = lbfgs_cpp(quench_coords, potential, tol=1e-8, M=1)
# This exists because some runs don't have hessian evaluations
try:
ret["nhev"]
except:
ret["nhev"] = 0
coordarg = 0
final_coords = ret["coords"]
final_coords_fname = str(index) + "_coords.txt"
np.savetxt(
old_code_results_folder_path + "/" + final_coords_fname,
final_coords,
delimiter=",",
)
energy, grad, hess = potential.getEnergyGradientHessian(final_coords)
hessian_fname = str(index) + "_hessian.txt"
np.savetxt(old_code_results_folder_path + "/" + hessian_fname,
hess,
delimiter=",")
grad_fname = str(index) + "_grad.txt"
np.savetxt(old_code_results_folder_path + "/" + grad_fname,
grad,
delimiter=",")
last_step_fname = str(index) + "_step.txt"
# np.savetxt(old_code_results_folder_path + '/' +
# last_step_fname, ret['step'], delimiter=',')
# check that the minimum hessian eigenvalue is positive
# print out the eigenvector corresponding to it
eigvals, eigvecs = np.linalg.eigh(hess)
print(eigvals[1])
print(eigvecs[:, 1])
# step_in_eigvec_basis = np.matmul(eigvecs.T, ret['step'])
# step_in_eigvec_basis_normalized = step_in_eigvec_basis/np.linalg.norm(step_in_eigvec_basis)
# -1e-15 is to avoid numerical issues
if np.min(eigvals) < -1e-3:
print("minimum Eigenvalue: ", np.min(eigvals))
print("Eigenvalues: ", eigvals)
raise Exception("negative eigenvalue")
# save heuristics as a dict
results = (
ret.coords,
ret.success,
coordarg,
ret.nfev,
ret.nsteps,
ret.nhev,
eigvecs,
)
# simplified dictionary since we're only using the success variable
res_dict = {
"success": bool(ret.success),
"nsteps": ret.nsteps,
"energy": ret.energy,
"nfev": ret.nfev,
}
# print("-------------------------------------------------steppp", step_in_eigvec_basis_normalized)
# print("eigenvalues", eigvals)
yaml_fname = str(index) + ".yaml"
# save heuristics as yaml
with open(old_code_results_folder_path + "/" + yaml_fname, "w") as f:
yaml.dump(res_dict, f)
print(ret.nsteps, "nsteps")
# the reason the potential is being passed is because quench coords needs the potential to figure out what to do
return results