使用matplotlib绘制大量点并耗尽内存

Question

我有一个简单格式的大(~6GB)文本文件

x1 y1 z1
x2 y2 z2
...

由于我可能不止一次加载这些数据,因此我np.memmap出于效率原因创建了一个文件:

X,Y,Z = np.memmap(f_np_mmap,dtype='float32',mode='r',shape=shape).T

我想做的是情节:

plt.scatter(X, Y, 
           color=custom_colorfunction(Z), 
           alpha=.01, s=.001, marker='s', linewidth=0)

这适用于较小的数据集.但是,对于这个更大的数据集,我的内存不足.我已经检查过plt.scatter它占用了所有内存; 我可以顺利通过X,Y,Z.有没有一种方法我"光栅化"画布,所以我不会耗尽内存？我不需要缩放和平移图像,它将直接进入磁盘.我意识到我可以对数据进行分区并对其进行绘制,但我不确定如何使用自定义色彩映射和 alpha值来完成此操作.

Answer 1

@tcaswell建议覆盖该Axes.draw方法绝对是最灵活的方法.

但是,您可以使用/滥用blitting来执行此操作而无需子类化Axes.只需使用draw_artist每次都不需要恢复画布.

还有一个额外的技巧:我们需要一个特殊的save方法,因为所有其他方法在保存之前绘制画布,这将消除我们之前绘制的所有内容.

此外,正如tcaswell注意到,调用draw_artist每个项目的速度相当慢,因此对于大量的点,您将需要对输入数据进行分块.分块将提供显着的加速,但这种方法总是比绘制单一的慢PathCollection.

无论如何,这些答案中的任何一个都应该可以缓解你的记忆问题.这是一个简单的例子.

import numpy as np
import matplotlib.pyplot as plt
from matplotlib import _png
from itertools import izip

def main():
    # We'll be saving the figure's background, so let's make it transparent.
    fig, ax = plt.subplots(facecolor='none')

    # You'll have to know the extent of the input beforehand with this method.
    ax.axis([0, 10, 0, 10])

    # We need to draw the canvas before we start adding points.
    fig.canvas.draw()

    # This won't actually ever be drawn. We just need an artist to update.
    col = ax.scatter([5], [5], color=[0.1, 0.1, 0.1], alpha=0.3)

    for xy, color in datastream(int(1e6), chunksize=int(1e4)):
        col.set_offsets(xy)
        col.set_color(color)
        ax.draw_artist(col)

    save(fig, 'test.png')

def datastream(n, chunksize=1):
    """Returns a generator over "n" random xy positions and rgb colors."""
    for _ in xrange(n//chunksize):
        xy = 10 * np.random.random((chunksize, 2))
        color = np.random.random((chunksize, 3))
        yield xy, color

def save(fig, filename):
    """We have to work around `fig.canvas.print_png`, etc calling `draw`."""
    renderer = fig.canvas.renderer
    with open(filename, 'w') as outfile:
        _png.write_png(renderer._renderer.buffer_rgba(),
                       renderer.width, renderer.height,
                       outfile, fig.dpi)

main()

此外,您可能会注意到顶部和左侧棘刺都被拉过来了.你可以通过ax.draw_artist(ax.spines['top'])在保存之前重新绘制这两个刺(等)来解决这个问题.

Answer 2

这样的事情(对于长代码很抱歉,大部分是从标准中复制的axes.Axes.draw):

from operator import itemgetter
class generator_scatter_axes(matplotlib.axes.Axes):
    def __init__(self, *args, **kwargs):
        matplotlib.axes.Axes.__init__(self, *args, **kwargs)
        self._big_data = None
    def draw(self, renderer=None, inframe=None):
        # copied from original draw (so you can still add normal artists ect)
        if renderer is None:
            renderer = self._cachedRenderer

        if renderer is None:
            raise RuntimeError('No renderer defined')
        if not self.get_visible():
            return
        renderer.open_group('axes')

        locator = self.get_axes_locator()
        if locator:
            pos = locator(self, renderer)
            self.apply_aspect(pos)
        else:
            self.apply_aspect()


        artists = []

        artists.extend(self.collections)
        artists.extend(self.patches)
        artists.extend(self.lines)
        artists.extend(self.texts)
        artists.extend(self.artists)
        if self.axison and not inframe:
            if self._axisbelow:
                self.xaxis.set_zorder(0.5)
                self.yaxis.set_zorder(0.5)
            else:
                self.xaxis.set_zorder(2.5)
                self.yaxis.set_zorder(2.5)
            artists.extend([self.xaxis, self.yaxis])
        if not inframe:
            artists.append(self.title)
            artists.append(self._left_title)
            artists.append(self._right_title)
        artists.extend(self.tables)
        if self.legend_ is not None:
            artists.append(self.legend_)

        # the frame draws the edges around the axes patch -- we
        # decouple these so the patch can be in the background and the
        # frame in the foreground.
        if self.axison and self._frameon:
            artists.extend(self.spines.itervalues())

        if self.figure.canvas.is_saving():
            dsu = [(a.zorder, a) for a in artists]
        else:
            dsu = [(a.zorder, a) for a in artists
                   if not a.get_animated()]

        # add images to dsu if the backend support compositing.
        # otherwise, does the manaul compositing  without adding images to dsu.
        if len(self.images) <= 1 or renderer.option_image_nocomposite():
            dsu.extend([(im.zorder, im) for im in self.images])
            _do_composite = False
        else:
            _do_composite = True

        dsu.sort(key=itemgetter(0))

        # rasterize artists with negative zorder
        # if the minimum zorder is negative, start rasterization
        rasterization_zorder = self._rasterization_zorder
        if (rasterization_zorder is not None and
            len(dsu) > 0 and dsu[0][0] < rasterization_zorder):
            renderer.start_rasterizing()
            dsu_rasterized = [l for l in dsu if l[0] < rasterization_zorder]
            dsu = [l for l in dsu if l[0] >= rasterization_zorder]
        else:
            dsu_rasterized = []

        # the patch draws the background rectangle -- the frame below
        # will draw the edges
        if self.axison and self._frameon:
            self.patch.draw(renderer)

        if _do_composite:
            # make a composite image blending alpha
            # list of (mimage.Image, ox, oy)

            zorder_images = [(im.zorder, im) for im in self.images
                             if im.get_visible()]
            zorder_images.sort(key=lambda x: x[0])

            mag = renderer.get_image_magnification()
            ims = [(im.make_image(mag), 0, 0, im.get_alpha()) for z, im in zorder_images]

            l, b, r, t = self.bbox.extents
            width = mag * ((round(r) + 0.5) - (round(l) - 0.5))
            height = mag * ((round(t) + 0.5) - (round(b) - 0.5))
            im = mimage.from_images(height,
                                    width,
                                    ims)

            im.is_grayscale = False
            l, b, w, h = self.bbox.bounds
            # composite images need special args so they will not
            # respect z-order for now

            gc = renderer.new_gc()
            gc.set_clip_rectangle(self.bbox)
            gc.set_clip_path(mtransforms.TransformedPath(
                    self.patch.get_path(),
                    self.patch.get_transform()))

            renderer.draw_image(gc, round(l), round(b), im)
            gc.restore()

        if dsu_rasterized:
            for zorder, a in dsu_rasterized:
                a.draw(renderer)
            renderer.stop_rasterizing()

        for zorder, a in dsu:
            a.draw(renderer)
        ############################    
        # new bits
        ############################
        if self._big_data is not None:

            for x, y, z in self._big_data:
                # add the (single point) to the axes
                a = self.scatter(x, y, color='r',
                            alpha=1, s=10, marker='s', linewidth=0)
                # add the point, in Agg this will render + composite
                a.draw(renderer)
                # remove the artist from the axes, shouldn't let the render know
                a.remove()
                # delete the artist for good measure
                del a
        #######################
        # end new bits
        #######################    
        # again, from original to clean up
        renderer.close_group('axes')
        self._cachedRenderer = renderer

像这样使用它:

In [42]: fig = figure()

In [43]: ax = generator_scatter_axes(fig, [.1, .1, .8, .8])

In [44]: fig.add_axes(ax)
Out[44]: <__main__.generator_scatter_axes at 0x56fe090>

In [45]: ax._big_data = rand(500, 3)

In [46]: draw()

我改变了散射函数,使其具有少量可见的形状.每次设置scatter对象时,这将非常慢.我要么采取合理的数据块并绘制这些数据,或者将调用替换为基础艺术家对象,或者使用Joe的建议并更新单个艺术家.scatter