def plot_variance_ratios()

in reproduce/plot_variance_ratio.py [0:0]

• 62 lines of code
• 13 McCabe index (conditional complexity)

def plot_variance_ratios(plot_name, data_files_grob, xvals):
    ylabel = "SVR Variance / SGD Variance"

    keys_to_show = ['2%', '11%', '33%', '100%']
     # Position of in-plot labels along the x axis

    epochs = []
    ratios = []
    vr_variances = []
    gradient_variances = []

    trace_data = {}

    data_files = glob.glob(data_files_grob)

    def atoi(text):
        return int(text) if text.isdigit() else text

    def natural_keys(text):
        return [ atoi(c) for c in re.split('(\d+)', text) ]

    data_files.sort(key=natural_keys)

    #pdb.set_trace()

    for fname in data_files:
        print("(ALL) processing file ", fname)
        with open(fname, 'rb') as fdata:
            rd = pickle.load(fdata)
            #pdb.set_trace()
            if 'batch_indices' in rd:
                print("Has batch indices")
                # Calculate x axis for plotting
                batch_indices = np.array(rd["batch_indices"])
                nk = len(batch_indices)
                if max(batch_indices) == min(batch_indices):
                    eval_points = np.array(range(nk))/nk
                else:
                    eval_points = batch_indices/max(batch_indices)

                epochs.append(rd["epoch"])
                #pdb.set_trace()

                ratio_points = (np.array(rd["vr_step_variances"])/np.array(rd["gradient_variances"])).tolist()

                for i, ep in enumerate(eval_points):
                    ep_name = "{0:.0f}%".format(100*ep)
                    if ep_name not in trace_data.keys():
                        trace_data[ep_name] = [ratio_points[i]]
                    else:
                        trace_data[ep_name].append(ratio_points[i])

    plt.cla()
    fig = plt.figure(figsize=(3.2,2))
    ax = fig.add_subplot(111)
    ax.set_prop_cycle("color", colors)

    #pdb.set_trace()
    for ep_name, data in trace_data.items():
        if ep_name in keys_to_show:
            ax.plot(epochs, data, ".",
                label=ep_name) #, linestyle=next(linestyles))
        if ep_name == "100%":
            print("100p epochs:", epochs)
            print("ratios: ", data)


    print("Finalizing plot")
    plt.xlabel('Epoch')
    plt.ylabel(ylabel)
    ax.set_yscale("log", basey=2)
    ax.set_yticks([2**(-i) for i in range(0, 11)])
    plt.ylim([1e-3, 3])
    plt.xlim([0.0, 240])

    # Horizontal line at 1
    #plt.axhline(y=1.0, color="#000000", linestyle='--')
    #plt.axhline(y=2.0, color="#000000", linestyle='--')
    ax.axhspan(1, 2, alpha=0.3, facecolor='red', edgecolor=None)
    # Step size reduction indicators
    plt.axvline(x=150.0, color="brown", linestyle='--')
    plt.axvline(x=220.0, color="brown", linestyle='--')


    #loc = plticker.LogLocator(base=2.0)
    #ax.yaxis.set_major_locator(loc)
    #plt.tick_params(axis='y', which='minor')

    ax.grid(False)
    ax.xaxis.set_tick_params(direction='in')
    ax.yaxis.set_tick_params(direction='in', right="on")
    labelLines(plt.gca().get_lines(), align=False, fontsize=label_fontsize, xvals=xvals)
    figname = "{}/{}.pdf".format(plot_dir, plot_name)

    fig.savefig(figname, bbox_inches='tight', pad_inches=0)
    print("saved", figname)