def adiabatic_opt(
flim: Tuple, b1lim: Tuple, nf: int, nb: int, b1max: Number,
device: torch.device = torch.device('cpu'),
dtype: torch.dtype = torch.float32
):
target, cube, pIni, b1map = init(flim, b1lim, nf, nb, b1max, device, dtype)
fn_err = metrics.err_l2z
fn_pen = penalties.pen_null
pIni0 = pIni
for _ in range(1):
pIni1, _ = optimizers.arctanLBFGS(target, cube, pIni0,
fn_err=fn_err, fn_pen=fn_pen,
niter=10, niter_gr=0, niter_rf=2,
b1Map=b1map)
rf = pIni1.rf # (1, xy, nT)
rf = (rf + torch.flip(rf, dims=[2]))/2
pIni1.rf = rf
pIni0 = pIni1
return
arg['niter_rf'] = dflt_arg('niter_rf', 2, lambda k: arg[k].item())
eta = dflt_arg('eta', 4, lambda k: float(arg[k].item()))
print('eta: ', eta)
err_meth = dflt_arg('err_meth', 'l2xy', lambda k: arg[k].item())
pen_meth = dflt_arg('pen_meth', 'l2', lambda k: arg[k].item())
err_hash = {'null': metrics.err_null,
'l2xy': metrics.err_l2xy, 'ml2xy': metrics.err_ml2xy,
'l2z': metrics.err_l2z}
pen_hash = {'null': penalties.pen_null, 'l2': penalties.pen_l2}
fn_err, fn_pen = err_hash[err_meth], pen_hash[pen_meth]
# %% pulse design
kw = {k: arg[k] for k in ('b1Map_', 'niter', 'niter_gr', 'niter_rf',
'doRelax')}
if False:
# optimize spiral shape parameters
pulse, optInfos = optimizers.arctanLBFGS(target, cube, pulse,
fn_err, fn_pen, eta=eta, **kw)
else:
# optimize directly on gradient samples (for fine tuning)
pulse, optInfos = optimizers.arctanLBFGS_orig(target, cube, pulse,
fn_err, fn_pen, eta=eta, **kw)
# %% saving
io.p2m(p2mName, pulse, {'optInfos': optInfos})