vsgd.py

from scipy import mean, ones_like, clip, logical_or, median, zeros_like
from sgd import SGD
from core.gradientalgos import BbpropHessians, FiniteDifferenceHessians


class vSGD(SGD, BbpropHessians):
    """ vSGD: SGD with variance-adapted learning rates,
    as described in Schaul, Zhang & LeCun 2012. """

    # how to initialize the running averages
    slow_constant = 2
    init_samples = 10
    
    # avoiding numerical instability
    epsilon = 1e-9
    outlier_level = 1

    def _additionalInit(self):
        # default setting
        if self.slow_constant is None:
            self.slow_constant = max(1, int(self.paramdim / 10.))
        
        if self.init_samples > 0:
            # get a few initial samples to work with
            tmp = self.batch_size
            self.batch_size = self.init_samples * tmp
            self._collectGradients()
            self._num_updates += self.init_samples
            self.batch_size = tmp
        
            # mean gradient vector
            self._gbar = mean(self._last_gradients, axis=0)
            # mean squared gradient
            self._vbar = (mean(self._last_gradients ** 2, axis=0) + self.epsilon) * self.slow_constant
        
            # mean diagonal Hessian
            #hs = clip(mean(self._last_diaghessians, axis=0), 1, 1 / self.epsilon)
            #self._hbar = hs * self.slow_constant
            self._hbar = mean(self._last_diaghessians, axis=0)
            
        else:
            # just start small, in the hope that the statistics will become accurate quickly enough
            self._gbar = zeros_like(self.parameters)
            self._vbar = ones_like(self.parameters) * self.epsilon
            self._hbar = ones_like(self.parameters) 
            
        self._vpart = self._gbar ** 2 / self._vbar
        
        # time constants
        self._taus = (ones_like(self.parameters) + self.epsilon) * 2#* self.init_samples
        
        # for debugging
        self._print_quantities = [('p', self.parameters),
                                  ('tau', self._taus),
                                  ('g', self._gbar),
                                  ('v', self._vbar),
                                  ('vpa', self._vpart),
                                  ('h', self._hbar),
                                  ]
        
    
    def _detectOutliers(self):
        """ Binary vector for which dimension saw an outlier gradient. """
        var = (self._vbar - self._gbar ** 2) / self.batch_size
        return (self._last_gradient - self._gbar) ** 2 > self.outlier_level ** 2 * var 


    def _computeStatistics(self):
        grads = self._last_gradients 
        hs = mean(self._last_diaghessians, axis=0) 
        
        # slow down updates if the last sample was an outlier
        if self.outlier_level is not None:
            self._taus[self._detectOutliers()] += 1
        
        # update statistics
        fract = 1. / self._taus
        self._gbar *= 1. - fract
        self._gbar += fract * mean(grads, axis=0)
        self._vbar *= 1. - fract
        self._vbar += fract * mean(grads ** 2, axis=0) + self.epsilon            
        self._hbar *= 1. - fract
        self._hbar += fract * hs
        
        # update time constants based on the variance-part of the learning rate
        self._vpart *= 0  
        if self.batch_size > 1:                
            self._vpart += self._gbar ** 2 / (1. / self.batch_size * self._vbar + 
                                       (self.batch_size - 1.) / self.batch_size * self._gbar ** 2)
        else:
            self._vpart += self._gbar ** 2 / self._vbar
            
        self._taus *= 1. - self._vpart
        self._taus += 1 + self.epsilon
        del hs
        return fract
    
    @property
    def learning_rate(self):
        return self._vpart / (self._hbar + self.epsilon)
    
    def _printStuff(self):
        print self._num_updates,
        for n, a in self._print_quantities:
            #print n, type(a)
            if abs(median(a)) > 1e4 or abs(median(a)) < 1e-3:
                print n, median(a), '\t',
            else:
                print n, round(median(a), 4), '\t',
        print
        
                                  
                                  
class vSGD_original(vSGD):
    """ The original version from 2012 had not outlier detection. """
    outlier_level = None    
                                  
                                  
class vSGDfd(FiniteDifferenceHessians, vSGD):
    """ vSGD with finite-difference estimate of diagonal Hessian,
    as described in Schaul & LeCun 2013. """
            
    @property
    def learning_rate(self):            
        return self._vpart * self._hpart
    
    def _additionalInit(self):
        vSGD._additionalInit(self)
        #h2s = clip(mean(self._last_diaghessians ** 2, axis=0), 1, 1 / self.epsilon)
        #self._vhbar = h2s * self.slow_constant #** 2     
        if self.init_samples > 0:
            self._vhbar = mean(self._last_diaghessians ** 2, axis=0)
        else:
            self._hbar = zeros_like(self.parameters)
            self._vhbar = ones_like(self.parameters) * self.epsilon 
        self._hpart = (self._hbar+self.epsilon) / (self._vhbar + self.epsilon)
        self._print_quantities.extend([('vh', self._vhbar),
                                       ('hpa', self._hpart),
                                       ])   
    
    def _computeStatistics(self):
        fract = vSGD._computeStatistics(self)
        self._vhbar *= (1. - fract)
        self._vhbar += fract * mean(self._last_diaghessians ** 2, axis=0) 
        self._hpart *= 0  
        self._hpart += (self._hbar+self.epsilon) / (self._vhbar + self.epsilon) 
        del fract        
    
    def _fdDirection(self):
        if self._num_updates == 0:
            return self._last_gradient + self.epsilon
        else:
            return self._gbar + self.epsilon
        
    def _detectOutliers(self):
        """ Binary vector for which dimension saw an outlier gradient:
        considers oultier curvature as well. """
        hs = mean(self._last_diaghessians, axis=0) 
        var = (self._vhbar - self._hbar ** 2) / self.batch_size
        res = logical_or(vSGD._detectOutliers(self),
                         (hs - self._hbar) ** 2 > self.outlier_level ** 2 * var)
        del hs, var
        return res