def plot_contour_trajectory(surf_file, dir_file, proj_file, surf_name='loss_vals',
                            vmin=0.1, vmax=10, vlevel=0.5, show=False):
    """2D contour + trajectory"""

    assert exists(surf_file) and exists(proj_file) and exists(dir_file)

    # plot contours
    f = h5py.File(surf_file,'r')
    x = np.array(f['xcoordinates'][:])
    y = np.array(f['ycoordinates'][:])
    X, Y = np.meshgrid(x, y)
    if surf_name in f.keys():
        Z = np.array(f[surf_name][:])

    fig = plt.figure()
    CS1 = plt.contour(X, Y, Z, levels=np.arange(vmin, vmax, vlevel))
    CS2 = plt.contour(X, Y, Z, levels=np.logspace(1, 8, num=8))

    # plot trajectories
    pf = h5py.File(proj_file, 'r')
    plt.plot(pf['proj_xcoord'], pf['proj_ycoord'], marker='.')

    # plot red points when learning rate decays
    # for e in [150, 225, 275]:
    #     plt.plot([pf['proj_xcoord'][e]], [pf['proj_ycoord'][e]], marker='.', color='r')

    # add PCA notes
    df = h5py.File(dir_file,'r')
    ratio_x = df['explained_variance_ratio_'][0]
    ratio_y = df['explained_variance_ratio_'][1]
    plt.xlabel('1st PC: %.2f %%' % (ratio_x*100), fontsize='xx-large')
    plt.ylabel('2nd PC: %.2f %%' % (ratio_y*100), fontsize='xx-large')
    df.close()
    plt.clabel(CS1, inline=1, fontsize=6)
    plt.clabel(CS2, inline=1, fontsize=6)
    fig.savefig(proj_file + '_' + surf_name + '_2dcontour_proj.pdf', dpi=300,
                bbox_inches='tight', format='pdf')
    pf.close()
    if show: plt.show()