ff0d04f231cc8cd912a99982269dca0c41a68316,thinc/neural/_classes/batchnorm.py,BatchNorm,begin_update,#BatchNorm#Any#Any#,43

Before Change

        diff = X - self.m
        incr = (1-alpha) * diff
        self.m += incr.mean(axis=0)
        self.v += (diff * incr).mean(axis=0)
        self.v *= alpha

        Xhat = _forward(self.ops, X, mu, var)

        // Batch "renormalization"
        if self.nr_upd >= 7500:
            Xhat *= var / (self.v+1e-08)
            Xhat += (mu - self.m) / (self.v+1e-08)

        y, backprop_rescale = self._begin_update_scale_shift(Xhat)

After Change

        return y

    def begin_update(self, X, drop=0.):
        if drop is None:
            return self.predict(X), None
        assert X.dtype == "float32"
        X, backprop_child = self.child.begin_update(X, drop=0.)
        N, mu, var = _get_moments(self.ops, X)
        var += self.eps
        self.r = min(self.rmax, max(1. / self.rmax, var / self.v))
        self.d = min(self.dmax, max(-self.dmax, (mu-self.m) / self.v))
        self.nr_upd += 1

        // I"m not sure this is the best thing to do --
        // Should we consider a sample be the instance, or the batch?
        // If we want the variance of the inputs it should be like:
        """
        diff = X - self.m
        incr = (1-alpha) * diff
        self.m += incr.mean(axis=0)
        self.v += (diff * incr).mean(axis=0)
        self.v *= alpha
        """
        self.m += self.alpha * (mu - self.m)
        self.v += self.alpha * (var - self.v)
        Xhat = _forward(self.ops, X, mu, var)
        Xhat *= self.r
        Xhat += self.d

Instances