コード例 #1
0
def set_Q_noise_solve(rule, vk, y, virusobjs, p_theta_eta, Na):
    t_Q=tm.time()

    WEIGHT_INDEX = 5 # the index of the abscissa weights in the rule data structure

    vo = virusobjs # use a shorter name inside the function

    """ initialize variables """

    Ny = y.shape[0]
    Nv = y.shape[1]
    Nzi = rule.shape[0]
    Nc = vo[0].cbar.shape[0]

    """ compute the integral $$Q$$ Eq. (178) its gradient and Hessian """
    Qem = 0
    A=0
    B=0

    # compute the expectation with numerical integrations
    eta=0

    nu=vo[eta].nu
    # unneeded since the lndet_fast function is commented out detV=prod(nu);
    # read tilde b
    b_fn = 'callti.out.read_by_C' # file contains \tilde{b} values computed from mathematica
#    lmax = np.max(vo[eta].clnp.l)
    lmax = find_lmax(vo[eta].clnp.l)
    print lmax
    tilde_b = rd_b(b_fn, lmax)

    for n in xrange(0,Nzi):
        Rabc = euler2R(rule[n, 0:3])
        L=setL_nord(Rabc, vo[eta].clnp.l, vo[eta].clnp.n, vk, vo[eta].Htable, vo[eta].map_unique2lp,tilde_b)
        """ Use fast lndet algorithm """
        Lc = np.dot(L,vo[eta].cbar) # \mu = L*c;
        y_Lc =  y - Lc
        wri = np.dot(rule[n,WEIGHT_INDEX],p_theta_eta[eta][0][:,n])

        A=A+np.sum(wri)
        B=B+np.dot(np.sum(y_Lc**2,axis=0),wri)

    A=A*Na*Na
    noise_var=B/A

    sys.stdout.write('set_Q_noise_solve time: %d\n'%(tm.time()-t_Q))
    return noise_var

    # copied from funcs for debugging and modification
    # omitted here for simplicity
    # returns a matrix L
    pass
コード例 #2
0
def setHQem(nu):  # rule 2D matrix; vk 2D
    data = si.io.loadmat('fmin.mat', squeeze_me=True, struct_as_record=False)

    t_Q = tm.time()  #clock:matlab function, return the date and time

    WEIGHT_INDEX = 5

    vo = data['vobj']

    Ny = data['y'].shape[0]
    Nv = data['y'].shape[1]
    rule = data['EM_iter'].rule
    vk = data['vk']
    noise_var = data['pixelnoisevar']
    p_theta_eta = data['p_theta_eta']
    y = data['y']
    Nzi = data['EM_iter'].rule.shape[0]
    tilde_b = data['tilde_b']
    eta = data['eta']

    # print (type(vo[0].flags[0]))
    Nc = vo[0].cbar.shape[0]
    #vo, a cell array, each element is a virus object
    #cbar, a vector of double precision, attribute of the object
    #matlab mathwork

    HQem = np.zeros((Nc, Nc), float)  #numpy array

    #eta = 0
    nu = vo[eta].nu
    if len(nu.shape)<2:
        d1 = nu.shape[0]
        nu = si.reshape(nu, [d1, 1])
    lndetV = np.sum(np.log(nu))

    #b_fn = 'callti.out.read_by_C'
    #lmax = vo[eta].clnp.l.max(0)  #il in matlab is l in python, ndarray
    #tilde_b = fun.rd_b(b_fn, lmax)

    for n in range(Nzi):  #n an index for a matrix
        Rabc = fun.euler2R(rule[n, 0:3])  #rule is a matrix
        L = fun.setL_nord(Rabc, vo[eta].clnp.l, vo[eta].clnp.n, vk, vo[eta].Htable, vo[eta].map_unique2lp,
                      tilde_b)  #L is a ndarray
        #clnp.in happens to be a keyword in python, how to fix this.
        lndet_Sigma = fun.lndet_fast(noise_var, L, nu, lndetV)  #ndarray

        Lc = np.dot(L, vo[eta].cbar)  #cbar is a list. Transferred to a ndarray before multiplication. Output is a ndarray.
        y_Lc = (y.transpose()-Lc).transpose()  #np.substract(y, Lc)  Good! Translate bxsfun() in Python
        wri = np.dot(rule[n][WEIGHT_INDEX], p_theta_eta[eta][:, n])  #.*, element-wise; *, dot product in matlab ???No matrix, all ndarrays???
        sum_wri = np.sum(wri)
        D, M = fun.LT_SigmaInv_fast(noise_var, L, vo[eta].nu, y_Lc)  #D, M is ndarray according to translation of Yu
        repwri = np.tile(np.sqrt(wri.T), (Nc, 1))
        D = D * repwri
        D = np.dot(D, D.T)

        y_Lc_Sigma = fun.y_Lc_Sigma_fast(noise_var, L, nu, y_Lc)  #ndarray

        HQem = HQem + (-2 * D + M * sum_wri) * M

    HQem = -HQem
    # sys.stdout.write('set_Q_dV_unc time: %d\n' % (tm.time() - t_Q))

    return HQem
コード例 #3
0
def setQem(nu ):  # rule 2D matrix; vk 2D
    data = si.io.loadmat('fmin.mat', squeeze_me=True, struct_as_record=False)


    if any( nuele <= 0 for nuele in nu ):
        print("element in nu <=0")

    t_Q = tm.time()  #clock:matlab function, return the date and time

    WEIGHT_INDEX = 5

    vo = data['vobj']
    rule = data['EM_iter'].rule
    vk = data['vk']
    noise_var = data['pixelnoisevar']
    p_theta_eta = data['p_theta_eta']
    y = data['y']
    Nzi = data['EM_iter'].rule.shape[0]
    tilde_b = data['tilde_b']
    eta = data['eta']


    # vo = vobj
    # rule = rule
    # vk = vk
    # noise_var = pixelnoisevar
    # p_theta_eta = p_theta_eta
    # y = y
    # Nzi = shape





    # print (type(vo[0].flags[0]))
    Nc = vo[0].cbar.shape[0]
    #vo, a cell array, each element is a virus object
    #cbar, a vector of double precision, attribute of the object
    #matlab mathwork

    Qem = 0
    gQem = np.zeros((Nc,), float)  #numpy array

    #eta = 0
    # nu = vo[eta].nu

    #easy for testing with matlab data, I think we should
    # delete this code after testing
    if len(nu.shape)<2:
        d1 = nu.shape[0]
        nu = si.reshape(nu, [d1, 1])

    lndetV = np.sum(np.log(nu))

    #b_fn = 'callti.out.read_by_C'
    #lmax = vo[eta].clnp.l.max(0)  #il in matlab is l in python, ndarray
    #############################
    #tilde_b = fun.rd_b(b_fn, lmax)

    for n in range(Nzi):  #n an index for a matrix
        ################################
        Rabc = fun.euler2R(rule[n, 0:3])  #rule is a matrix
        L = fun.setL_nord(Rabc, vo[eta].clnp.l, vo[eta].clnp.n, vk, vo[eta].Htable, vo[eta].map_unique2lp,
                      tilde_b)  #L is a ndarray
        #clnp.in happens to be a keyword in python, how to fix this.
        lndet_Sigma = fun.lndet_fast(noise_var, L, nu, lndetV)  #ndarray

        Lc = np.dot(L, vo[eta].cbar)  #cbar is a list. Transferred to a ndarray before multiplication. Output is a ndarray.
        y_Lc = (y.transpose()-Lc).transpose()
        #change [eta][0][:,n]->[eta][:,n]
        wri = np.dot(rule[n][WEIGHT_INDEX], p_theta_eta[eta][:, n])  #.*, element-wise; *, dot product in matlab ???No matrix, all ndarrays???
        sum_wri = np.sum(wri)
        D, M = fun.LT_SigmaInv_fast(noise_var, L, vo[eta].nu, y_Lc)  #D, M is ndarray according to translation of Yu
        repwri = np.tile(np.sqrt(wri.T), (Nc, 1))
        D = D * repwri
        D = np.dot(D, D.T)

        y_Lc_Sigma = fun.y_Lc_Sigma_fast(noise_var, L, nu, y_Lc)  #ndarray

        Qem = Qem + (-0.5) * (np.dot(lndet_Sigma + y_Lc_Sigma, wri))
        gQem = gQem + np.diag(D - np.dot(sum_wri, M))

    Qem = -Qem
    gQem = -gQem
    temp = np.around(-Qem, decimals=16)

    sys.stdout.write('Q = %.16f\n'% temp)
    sys.stdout.write('set_Q_dV_unc time: %d\n' % (tm.time() - t_Q))

    return Qem, gQem