def optimize(self, agd=True, batch_size=None, batch_number=None, lr=0.01,
**kwargs):
"""\
Optimize stress function using gradient-based methods. If batch size or
number are given, optimization begins with stochastic gradient descent.
If agd is set to True, optimization ends with adaptive gradient descent.
"""
if self.verbose > 0:
print('- MDS.optimize():')
if batch_number is not None or batch_size is not None:
F = lambda X: self.F(X,batch_number=batch_number,
batch_size=batch_size)
if self.verbose > 0:
print(' method : stochastic gradient descent')
if batch_number is None:
print(f' batch size : {batch_size}')
else:
print(f' batch number : {batch_number}')
self.X, H = gd.mgd(self.X,F,lr=lr,**kwargs)
self.update(H=H)
if agd is True:
F = lambda X: self.F(X)
if self.verbose > 0:
print(' method : exact gradient & adaptive gradient descent')
self.X, H = gd.agd(self.X,F,**kwargs,**self.H)
self.update(H=H)
if self.verbose > 0:
print(f' final stress : {self.cost:0.2e}[{self.ncost:0.2e}]')
def optimize_all(self,
agd=True,
batch_size=None,
batch_number=None,
lr=0.01,
**kwargs):
if self.verbose:
print('- Multiview.optimize_all(): ')
p = [None] + [self.persp.c] * self.K
if batch_number is not None or batch_size is not None:
XQ = [self.X] + self.Q
F = lambda XQ: self.F(XQ[0],
XQ[1::],
batch_number=batch_number,
batch_size=batch_size)
XQ, H = gd.mgd(XQ, F, lr=lr, **kwargs)
self.X = XQ[0]
self.Q = XQ[1::]
self.update(H=H)
if agd is True:
XQ = [self.X] + self.Q
F = lambda XQ: self.F(XQ[0], XQ[1::])
XQ, H = gd.cagd(XQ, F, **kwargs, **self.H)
self.X = XQ[0]
self.Q = XQ[1::]
self.update(H=H)
if self.verbose > 0:
print(f' Final stress : {self.cost:0.2e}')
def optimize_X(self, agd=True, batch_size=None, batch_number=None, lr=0.01,
**kwargs):
if self.verbose > 0:
print('- Multiview.optimize_X():')
if batch_number is not None or batch_size is not None:
if self.verbose > 0:
print(' method : stochastic gradient descent')
if batch_number is None:
print(f' batch size : {batch_size}')
else:
print(f' batch number : {batch_number}')
F = lambda X: self.FX(X,self.Q,batch_number=batch_number,
batch_size=batch_size)
self.X, H = gd.mgd(self.X,F,lr=lr,**kwargs)
self.update(H=H)
if agd is True:
F = lambda X: self.FX(X,self.Q)
if self.verbose > 0:
print(' method : exact gradient & adaptive gradient descent')
self.X, H = gd.agd(self.X,F,**kwargs,**self.H)
self.update(H=H)
if self.verbose > 0:
print(f' Final stress : {self.cost:0.2e}[{self.ncost:0.2e}]')
def optimize_Q(self,batch_size=None,batch_number=None,lr=0.01,**kwargs):
if self.verbose > 0:
print('- Multiview.optimize_Q():')
F = lambda Q: self.FQ(self.X,Q,batch_number=batch_number,
batch_size=batch_size)
if batch_number is None and batch_size is None:
self.Q, H = gd.cagd(self.Q,F,**kwargs)
else:
self.Q, H = gd.mgd(self.Q,F,lr=lr,**kwargs)
self.update(H=H)
if self.verbose > 0:
print(f' Final stress : {self.cost:0.2e}[{self.ncost:0.2e}]')
def optimize_X(self, agd=True, approx=0.5, lr=5., **kwargs):
if self.verbose > 0:
print('- Multiview.optimize_X():')
if approx is not None:
if self.verbose > 0:
print(' method : stochastic gradient descent')
print(' approx =', approx)
F = lambda X: self.FX(X, self.Q, approx=approx)
self.X, H = gd.mgd(self.X, F, lr=lr, **kwargs)
self.update(H=H)
if agd is True:
F = lambda X: self.FX(X, self.Q)
if self.verbose > 0:
print(' method : exact gradient & adaptive gradient descent')
self.X, H = gd.agd(self.X, F, **kwargs, **self.H)
self.update(H=H)
if self.verbose > 0:
print(f' Final stress : {self.cost:0.2e}')