Exemplo n.º 1
0
def plot_surface(tdata_agg, surface_result, num_levels=11):
    """plot the 3D surface and the level set"""

    # calculate the 3D surface
    num_mt = 101
    num_dt = 101
    mean_temp = np.linspace(tdata_agg.mean_temp.min(),
                            tdata_agg.mean_temp.max(), num_mt)
    daily_temp = np.linspace(tdata_agg.daily_temp.min(),
                             tdata_agg.daily_temp.max(), num_dt)
    MT, DT = np.meshgrid(mean_temp, daily_temp)
    mt = MT.flatten()
    dt = DT.flatten()

    # compute residual surface
    scaled_dt = utils.scale_daily_temp(mt, dt, surface_result.scale_params)
    invalid_id = (scaled_dt <= -1.0) | (scaled_dt > 0.75)

    surface = surface_result.surface_func(mt, dt)
    surface[invalid_id] = np.inf
    surface = surface.reshape(num_mt, num_dt)

    # plot 3D surface
    fig = plt.figure(figsize=(16, 8))
    ax = fig.add_subplot(121, projection='3d')
    ax.scatter(tdata_agg.mean_temp,
               tdata_agg.daily_temp,
               tdata_agg.obs_mean,
               s=1.0 / tdata_agg.obs_std,
               c=tdata_agg.obs_mean)

    ax.plot_surface(MT, DT, surface)
    ax.set_xlabel('mean temp degree')
    ax.set_ylabel('daily temp cat')
    ax.set_zlabel('ln rr')

    # plot the level set
    ax = fig.add_subplot(122)
    ax.scatter(tdata_agg.mean_temp,
               tdata_agg.daily_temp,
               s=1.0 / tdata_agg.obs_std,
               c=tdata_agg.obs_mean)
    c = ax.contour(MT, DT, surface, levels=np.linspace(-0.5, 0.5, num_levels))
    ax.set_xlabel('mean temp degree')
    ax.set_ylabel('daily temp cat')
    ax.clabel(c)

    # min_daily_temp_id = np.array([np.argmin(surface[:, i])
    #                               for i in range(num_mt)])
    # min_daily_temp = daily_temp[min_daily_temp_id]
    min_daily_temp = np.array([
        surface_result.tmrl_at_mean_temp(mean_temp[i]) for i in range(num_mt)
    ])
    ax.scatter(mean_temp, min_daily_temp, c='r', marker='.')
Exemplo n.º 2
0
def fit_surface(tdata_agg, scale_params=[40.0, 1.25], linear_no_mono=False):
    """fit the mean surface"""

    # unpack data
    mt = tdata_agg.mean_temp
    dt = tdata_agg.daily_temp
    scaled_dt = utils.scale_daily_temp(mt, dt, scale_params)

    obs_mean = tdata_agg.obs_mean
    obs_std = tdata_agg.obs_std
    study_sizes = tdata_agg.study_sizes

    # create spline
    mt_knots = np.linspace(mt.min(), mt.max(), 2)
    mt_degree = 3
    if linear_no_mono:
        dt_degree = 1
        dt_knots = np.linspace(scaled_dt.min(), scaled_dt.max(), 2)
    else:
        dt_degree = 3
        dt_knots = np.linspace(scaled_dt.min(), scaled_dt.max(), 3)
    spline_list = [
        xspline.ndxspline(2, [mt_knots, dt_knots], [mt_degree, dt_degree])
    ]

    # create mrbrt object
    x_cov_list = [{
        'cov_type': 'ndspline',
        'spline_id': 0,
        'mat': np.vstack((mt, scaled_dt))
    }]
    z_cov_list = [{'cov_type': 'linear', 'mat': np.ones(mt.size)}]

    mr = mrbrt.MR_BRT(obs_mean, obs_std, study_sizes, x_cov_list, z_cov_list,
                      spline_list)

    # add priors
    prior_list = [{
        'prior_type':
        'ndspline_shape_function_uprior',
        'x_cov_id':
        0,
        'interval': [[mt.min(), mt.max()], [scaled_dt.min(),
                                            scaled_dt.max()]],
        'indicator': [-1.0, 1.0],
        'num_points': [20, 20]
    }, {
        'prior_type': 'ndspline_shape_monotonicity',
        'x_cov_id': 0,
        'dim_id': 1,
        'interval': [[mt.min(), mt.max()], [0.7, scaled_dt.max()]],
        'indicator': 'increasing',
        'num_points': [20, 10]
    }, {
        'prior_type': 'z_cov_uprior',
        'z_cov_id': 0,
        'prior': np.array([[0.0] * mr.k_gamma, [0.0] * mr.k_gamma])
    }]
    if not linear_no_mono:
        prior_list += [
            {
                'prior_type': 'ndspline_shape_monotonicity',
                'x_cov_id': 0,
                'dim_id': 1,
                'interval': [[mt.min(), mt.max()], [scaled_dt.min(), -0.75]],
                'indicator': 'decreasing',
                'num_points': [20, 10]
            },
        ]

    mr.addPriors(prior_list)

    # initialization
    k_beta = mr.lt.k_beta
    k_gamma = mr.lt.k_gamma
    X = mr.lt.JF(mr.lt.beta)
    S = mr.lt.S
    V = S**2
    Y = mr.lt.Y
    beta0 = np.linalg.solve((X.T / V).dot(X), (X.T / V).dot(Y))
    gamma0 = np.zeros(k_gamma)
    x0 = np.hstack((beta0, gamma0))

    # fit the model and store the result
    mr.fitModel(x0=x0)

    # compute posterior variance for the beta
    # extract the matrix
    beta_var = (X.T / V).dot(X)

    if mr.lt.use_regularizer:
        H = np.vstack([
            mr.lt.H(
                np.hstack(
                    np.eye(1, k_beta, i).reshape(k_beta, ), np.zeros(k_gamma)))
            for i in range(k_beta)
        ]).T
        SH = mr.lt.h[1]
        VH = SH**2
        beta_var += (H.T / VH).dot(VH)

    beta_var = np.linalg.inv(beta_var)

    mean_temp = tdata_agg.unique_mean_temp
    daily_temp_range = []
    for mt in mean_temp:
        tdata_agg_mt = mtslice.extract_mtslice(tdata_agg, mt)
        daily_temp_range.append(
            [tdata_agg_mt.daily_temp.min(),
             tdata_agg_mt.daily_temp.max()])

    surface_result = utils.SurfaceResult(mr.beta_soln,
                                         beta_var,
                                         spline_list[0],
                                         mean_temp,
                                         daily_temp_range,
                                         scale_params=scale_params)

    return surface_result
Exemplo n.º 3
0
def fit_surface(tdata, scale_params=[40.0, 1.25]):
    """fit the mean surface"""
    agg_tdata = aggregate_data(tdata)

    # unpack data
    mt = agg_tdata.mean_temp
    dt = agg_tdata.daily_temp
    scaled_dt = utils.scale_daily_temp(mt, dt, scale_params)

    obs_mean = agg_tdata.obs_mean
    obs_std = agg_tdata.obs_std * 15.0
    study_sizes = agg_tdata.study_sizes

    # mt = tdata.mean_temp
    # dt = tdata.daily_temp
    # scaled_dt = utils.scale_daily_temp(mt, dt, scale_params)

    # obs_mean = tdata.obs_mean
    # obs_std = tdata.obs_std
    # study_sizes = np.array([1]*obs_mean.size)

    # create spline
    mt_knots = np.linspace(mt.min(), mt.max(), 2)
    dt_knots = np.linspace(scaled_dt.min(), scaled_dt.max(), 2)
    mt_degree = 3
    dt_degree = 3
    spline_list = [
        xspline.ndxspline(2, [mt_knots, dt_knots], [mt_degree, dt_degree])
    ]

    # create mrbrt object
    x_cov_list = [{
        'cov_type': 'ndspline',
        'spline_id': 0,
        'mat': np.vstack((mt, scaled_dt))
    }]
    z_cov_list = [{'cov_type': 'linear', 'mat': np.ones(mt.size)}]

    mr = mrbrt.MR_BRT(obs_mean, obs_std, study_sizes, x_cov_list, z_cov_list,
                      spline_list)

    # add priors
    prior_list = [
        {
            'prior_type': 'ndspline_shape_function_uprior',
            'x_cov_id': 0,
            'interval': [[mt.min(), mt.max()],
                         [scaled_dt.min(), scaled_dt.max()]],
            'indicator': [-1.0, 1.0],
            'num_points': [20, 20]
        },
        # {
        #     'prior_type': 'x_cov_gprior',
        #     'x_cov_id': 0,
        #     'prior': np.array([[0.0]*mr.k_beta, [10.0]*mr.k_beta])
        # },
        {
            'prior_type': 'ndspline_shape_monotonicity',
            'x_cov_id': 0,
            'dim_id': 1,
            'interval': [[mt.min(), mt.max()], [0.5, scaled_dt.max()]],
            'indicator': 'increasing',
            'num_points': [20, 10]
        },
        {
            'prior_type': 'ndspline_shape_monotonicity',
            'x_cov_id': 0,
            'dim_id': 1,
            'interval': [[mt.min(), mt.max()], [scaled_dt.min(), -0.75]],
            'indicator': 'decreasing',
            'num_points': [20, 10]
        },
        {
            'prior_type': 'z_cov_uprior',
            'z_cov_id': 0,
            'prior': np.array([[1e-6] * mr.k_gamma, [1e-6] * mr.k_gamma])
        }
    ]
    mr.addPriors(prior_list)

    # fit the model and store the result
    mr.fitModel()

    # compute posterior variance for the beta
    # extract the matrix
    k_beta = mr.lt.k_beta
    k_gamma = mr.lt.k_gamma
    X = mr.lt.JF(mr.lt.beta)
    S = mr.lt.S
    V = S**2

    beta_var = (X.T / V).dot(X)

    if mr.lt.use_regularizer:
        H = np.vstack([
            mr.lt.H(
                np.hstack(
                    np.eye(1, k_beta, i).reshape(k_beta, ), np.zeros(k_gamma)))
            for i in range(k_beta)
        ]).T
        SH = mr.lt.h[1]
        VH = SH**2
        beta_var += (H.T / VH).dot(VH)

    beta_var = np.linalg.inv(beta_var)

    surface_result = utils.SurfaceResult(mr.beta_soln,
                                         beta_var,
                                         spline_list[0],
                                         scale_params=scale_params)

    return surface_result, agg_tdata