def run(step=50000,
convergence=0.99,
loss_conver=20.0,
writelib=1000,
opt=None,
cons=['val', 'vale', 'lp3', 'cutoff'],
print_step=10,
lr=1.0e-4,
maxcycle=20,
batch=50):
accu, loss = 0.0, 100
c = 0
while accu < convergence and c < maxcycle:
loss, accu, accMax, p, zpe, i = train_reax(step=step,
opt=opt,
cons=cons,
convergence=convergence,
lr=lr,
writelib=writelib,
print_step=print_step)
if loss > loss_conver and accu > convergence:
convergence = accu + 0.0003
# system('cp ffield.json ffield%dt%de%df%d.json' %(int(loss),messages,ef,fm))
c += 1
if loss > 0.0:
send_msg('- Convergence reached, the loss %7.4f and accuracy %7.4f.' %
(loss, accu))
def train_reax(writelib=1000,
print_step=10,
step=500,
opt=None,
cons=None,
lr=1.0e-4,
convergence=0.97,
lossConvergence=100.0,
batch=50):
rn = ReaxFF(libfile='ffield.json',
direcs=direcs,
dft='siesta',
optword='nocoul',
opt=None,
cons=None,
batch_size=batch,
losFunc='n2',
lambda_bd=10000.0,
lambda_me=0.01,
weight={
'al4': 2.0,
'others': 2.0
},
convergence=convergence,
lossConvergence=lossConvergence
) # Loss Functon can be n2,abs,mse,huber
# GradientDescentOptimizer AdamOptimizer AdagradOptimizer RMSPropOptimizer
loss, accu, accMax, i, zpe = rn.run(learning_rate=lr,
step=step,
print_step=print_step,
writelib=writelib)
libstep = int(i - i % writelib)
if i == libstep:
libstep = libstep - writelib
if libstep <= 0:
ffd = 'ffield.json'
else:
ffd = 'ffield_' + str(libstep) + '.json'
if loss == 0.0 and accu == 0.0:
send_msg(
'- Warning: the loss is NaN, parameters from %s changed auomatically ...'
% ffd)
return 0.0, 1.0, 1.0, None, None, i
# with open(ffd,'r') as fj:
# j = js.load(fj)
# ic = Init_Check(nanv=nanv)
# j['p'] = ic.auto(j['p'])
# ic.close()
#with open('ffield.json','w') as fj:
# js.dump(j,fj,sort_keys=True,indent=2)
p = rn.p_
ME = rn.MolEnergy_
rn.close()
return loss, accu, accMax, p, ME, i