def forceDistrRecord(self, coord_x, coord_y, updated_Fsys, d): if self.ndims == 1: self.colvars_force[getIndices(coord_x, self.bins)] += updated_Fsys if self.ndims == 2: self.colvars_force[d][getIndices(coord_x, self.bins)][getIndices(coord_y, self.bins)] += updated_Fsys
def histDistrRecord(self, coord_x, coord_y, d): if self.ndims == 1: self.colvars_count[getIndices(coord_x, self.bins)] += 1 if self.ndims == 2: self.colvars_count[d][getIndices(coord_x, self.bins)][getIndices(coord_y, self.bins)] += 1
def histDistrRecord(self,
coord_x,
coord_y,
d=None): # TODO better strucuture
if self.ndims == 1:
self.colvars_count[getIndices(coord_x, self.bins)] += 1
if self.ndims == 2:
self.colvars_count[d][getIndices(coord_x, self.bins)][getIndices(
coord_y, self.bins)] += 1
def getLocalForce(self, coord_x, coord_y, vel, d=None):
coord_x = truncateFloat(coord_x)
coord_y = truncateFloat(coord_y)
# Regular MD
if self.abfCheckFlag == "no" and self.nnCheckFlag == "no":
Fu = self.PotentialForce(coord_x, coord_y, d)
Fsys = self.PotentialForce(
coord_x, coord_y, d) + self.visForce(vel) + self.randForce()
self.forceDistrRecord(coord_x, coord_y, Fsys, d)
return Fu / self.mass
# Regular MD with ABF
if self.abfCheckFlag == "yes" and self.nnCheckFlag == "no":
Fu = self.PotentialForce(coord_x, coord_y, d)
Fsys = self.PotentialForce(
coord_x, coord_y, d) + self.visForce(vel) + self.randForce()
self.forceDistrRecord(coord_x, coord_y, Fsys, d)
Fabf = self.appliedBiasForce(coord_x, coord_y, d)
return (Fu + Fabf) / self.mass
# Regular MD with ABF and ANN (I)
if self.abfCheckFlag == "yes" and self.nnCheckFlag == "yes":
Fu = self.PotentialForce(coord_x, coord_y, d)
Fsys = self.PotentialForce(
coord_x, coord_y, d) + self.visForce(vel) + self.randForce()
self.forceDistrRecord(coord_x, coord_y, Fsys, d)
if self.frame % self.Frequency == 0 and self.frame != 0:
output = trainingNN("loss.dat", "hyperparam.dat", self.ndims,
len(self.bins))
self.colvars_force = (self.colvars_force / self.colvars_count)
self.colvars_force[np.isnan(
self.colvars_force)] = 0 # 0/0 = nan n/0 = inf
self.colvars_force_NN = \
output.training(self.colvars_coord, self.colvars_force, self.learning_rate, self.regularCoeff, self.epoch, self.NNoutputFreq)
self.colvars_force = (self.colvars_force * self.colvars_count)
self.forceDistrRecord(coord_x, coord_y, self.colvars_force_NN,
d)
Fabf = -self.colvars_force_NN[getIndices(coord_x, self.bins)]
return (Fu + Fabf) / self.mass
else:
if self.frame < self.Frequency:
Fabf = self.appliedBiasForce(coord_x, coord_y, d)
return (Fu + Fabf) / self.mass
else:
Fabf = -self.colvars_force_NN[getIndices(
coord_x, self.bins)]
return (Fu + Fabf) / self.mass
def forceDistrRecord(self,
coord_x,
coord_y,
updated_Fsys,
d=None): # TODO better strucuture
if self.ndims == 1:
self.colvars_force[getIndices(coord_x, self.bins)] += updated_Fsys
if self.ndims == 2:
self.colvars_force[d][getIndices(coord_x, self.bins)][getIndices(
coord_y, self.bins)] += updated_Fsys
def getLocalForce(self, coord_x, coord_y, vel, d=None): #TODO 2D
coord_x = truncateFloat(coord_x)
coord_y = truncateFloat(coord_y)
Fu = self.PotentialForce(coord_x, coord_y, d)
Fsys = self.PotentialForce(coord_x, coord_y,
d) + self.visForce(vel) + self.randForce()
# Regular MD
if self.abfCheckFlag == "no" and self.nnCheckFlag == "no":
self.forceDistrRecord(coord_x, coord_y, Fsys, d)
self.histDistrRecord(coord_x, coord_y, d)
return Fu / self.mass
# Regular MD with ABF and ANN (I)
if self.abfCheckFlag == "yes" and self.nnCheckFlag == "yes":
if self.frame < self.Frequency:
Fabf = self.appliedBiasForce(coord_x, coord_y, d)
self.forceDistrRecord(coord_x, coord_y, Fsys, d)
self.histDistrRecord(coord_x, coord_y, d)
return (Fu + Fabf) / self.mass
else: # NN takes over here
self.forceDistrRecord(coord_x, coord_y, Fsys, d)
self.histDistrRecord(coord_x, coord_y, d)
if self.ndims == 1:
Fabf = self.gradient[getIndices(coord_x, self.bins)]
if self.ndims == 2:
pass
return (Fu + Fabf) / self.mass
def checkBoltzDistr(self, fileIn, fileOut): with open(fileIn, "r") as fin: if self.ndims == 1: # use xmgrace instead prob_x = np.zeros((self.binNum), dtype = np.int32) nsamples = 0 for line in fin: line = line.split() if line[0] != "#": nsamples += 1 prob_x[getIndices(truncateFloat(float(line[2])), self.x_axis)] += 1 prob_x = np.array(prob_x) prob_x = (prob_x / nsamples) # final probability distribution with open(fileOut, "w") as fout: for i in range(len(prob_x)): # in xmgrace, one can discard the point on pi (boundary error) fout.write(str(self.x_axis[i]) + " " + str(prob_x[i]) + "\n") if self.ndims == 2: prob_xy = np.zeros((self.binNum, self.binNum), dtype = np.float64) nsamples = 0 for line in fin: line = line.split() if line[0] != "#": nsamples += 1 prob_xy[getIndices(truncateFloat(float(line[2])), self.x_axis)][getIndices(truncateFloat(float(line[4])), self.y_axis)] += 1 # 2 for cartcoord_1D, 2 4 for cartcoord_2D prob_xy = (prob_xy / nsamples) # final probability distribution with open(fileOut, "w") as fout: for i in range(self.binNum): # discard the point on the positive boundary (PBC issues) for j in range(self.binNum): # discard the point on the positive boundary (PBC issues) fout.write(str(self.x_axis[i]) + " ") fout.write(str(self.y_axis[j]) + " " + str(prob_xy[i][j]) + "\n") prob_xy = np.delete(prob_xy, -1, 0) # rendering; prevent boundary error prob_xy = np.delete(prob_xy, -1, 1) r = rendering(self.ndims, self.half_boxboundary, self.binNum, self.temperature) r.render(prob_xy, name=str(self.abfCheckFlag + "_" + self.nnCheckFlag + "_" + "boltz2D"))
def appliedBiasPotential(self, coord_x, coord_y, d=None): #TODO 2D
if self.abfCheckFlag == "yes" and self.nnCheckFlag == "yes":
if self.ndims == 1:
if self.frame <= self.Frequency: # initial sweep
return 0 # np.sin(coord_x)
else:
return self.biasingPotentialFromNN[getIndices(
coord_x, self.bins)]
if self.ndims == 2:
pass
def forceDistrRecord(self, coord_x, coord_y, updated_Fsys, d):
if self.ndims == 1:
if isinstance(updated_Fsys, float) or isinstance(
updated_Fsys,
int): # conventional ABF collection, which is a float
self.colvars_force[getIndices(coord_x,
self.bins)] += updated_Fsys
self.colvars_count[getIndices(coord_x, self.bins)] += 1
else: # refined force from NN, which is a np.ndarray
self.colvars_force += updated_Fsys
self.colvars_count += 1
if self.ndims == 2:
if isinstance(updated_Fsys, float) or isinstance(
updated_Fsys, int):
self.colvars_force[d][getIndices(
coord_x, self.bins)][getIndices(coord_y,
self.bins)] += updated_Fsys
self.colvars_count[d][getIndices(coord_x,
self.bins)][getIndices(
coord_y, self.bins)] += 1
else:
self.colvars_force += updated_Fsys
self.colvars_count += 1
def appliedBiasForce(self, coord_x, coord_y, d): if self.ndims == 1: if self.colvars_count[getIndices(coord_x, self.bins)] == 0: return 0 else: return -(self.colvars_force[getIndices(coord_x, self.bins)] / self.colvars_count[getIndices(coord_x, self.bins)]) if self.ndims == 2: if self.colvars_count[d][getIndices(coord_x, self.bins)][getIndices(coord_y, self.bins)] == 0: return 0 else: return -(self.colvars_force[d][getIndices(coord_x, self.bins)][getIndices(coord_y, self.bins)] / self.colvars_count[d][getIndices(coord_x, self.bins)][getIndices(coord_y, self.bins)])