models_reservoirs.py 文件源码

def mixtureMV(self, mu, sig, ps):
        """LearningRules.mixtureMV

        Sample from a multivariate gaussian mixture with \mu = means, \Sigma = cov matrix
        """
        # print "%s.mixtureMV: Implement me" % (self.__class__.__name__)

        # print "mixtureMV", "mu", mu.shape, "sig", sig.shape, "ps", ps.shape

        mixcomps = self.mixcomps
        d = self.dim

        assert mixcomps == ps.shape[0]
        assert not np.isnan(mu).any()
        assert not np.isnan(sig).any()
        assert not np.isnan(ps).any()

        # check flat inputs
        if mu.shape[1] == 1:
            mu = mu.reshape((self.mixcomps, self.dim))
            # mu_ = mu.reshape((mixcomps, d))
        if sig.shape[1] == 1:
            # S_diag = sig[:(mixcomps * self.dim)]
            # S_triu = sig[(mixcomps * self.dim):]
            S_diag = sig[:(mixcomps * d)].reshape((mixcomps, d))
            S_triu = sig[(mixcomps * d):].reshape((mixcomps, -1))
            S = []
            for i in range(mixcomps):
                S_ = np.diag(S_diag[i])
                S_[np.triu_indices(d,  1)] = S_triu[i]
                S_[np.tril_indices(d, -1)] = S_triu[i] # check if thats correct
                S.append(S_)
            sig = np.array(S)

        # print "mixtureMV", "mu", mu.shape, "sig", sig.shape, "ps", ps.shape

        # d = mu.shape[0]
        compidx = self.mixture_sample_prior(ps)

        mu_ = mu[compidx]
        S_  = sig[compidx]

        # print "compidx", compidx, "mu_", mu_, "S_", S_

        y = np.random.multivariate_normal(mu_, S_)
        return y

    # mixture, mdn_loss, and softmax are taken from karpathy's MixtureDensityNets.py