Python has some great data visualization librairies, but few can render GIFs or video animations. This post shows how to use MoviePy as a generic animation plugin for any other library.git
MoviePy lets you define custom animations with a functionmake_frame(t), which returns the video frame corresponding to time t (in seconds):github
1 2 3 4 5 6 7 8 9 10 11 12 |
from moviepy.editor import VideoClip def make_frame(t): """ returns an image of the frame at time t """ # ... create the frame with any library return frame_for_time_t # (Height x Width x 3) Numpy array animation = VideoClip(make_frame, duration=3) # 3-second clip # For the export, many options/formats/optimizations are supported animation.write_videofile("my_animation.mp4", fps=24) # export as video animation.write_gif("my_animation.gif", fps=24) # export as GIF (slow) |
In previous posts I used this method to animate vector graphics (with the library Gizeh), and ray-traced 3D scenes (generated by POV-Ray). This post covers the scientific libraries Mayavi, Vispy, Matplotlib, Numpy, and Scikit-image.canvas
Animations with Mayavi
Mayavi is a Python module for interactive 3D data visualization with a simple interface. In this first example we animate a surface whose elevation depends on the timet:app
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 |
import numpy as np import mayavi.mlab as mlab import moviepy.editor as mpy duration= 2 # duration of the animation in seconds (it will loop) # MAKE A FIGURE WITH MAYAVI fig_myv = mlab.figure(size=(220,220), bgcolor=(1,1,1)) X, Y = np.linspace(-2,2,200), np.linspace(-2,2,200) XX, YY = np.meshgrid(X,Y) ZZ = lambda d: np.sinc(XX**2+YY**2)+np.sin(XX+d) # ANIMATE THE FIGURE WITH MOVIEPY, WRITE AN ANIMATED GIF def make_frame(t): mlab.clf() # clear the figure (to reset the colors) mlab.mesh(YY,XX,ZZ(2*np.pi*t/duration), figure=fig_myv) return mlab.screenshot(antialiased=True) animation = mpy.VideoClip(make_frame, duration=duration) animation.write_gif("sinc.gif", fps=20) |
Another example with a wireframe mesh whose coordinates and view angle depend on the time :dom
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 |
import numpy as np import mayavi.mlab as mlab import moviepy.editor as mpy duration = 2 # duration of the animation in seconds (it will loop) # MAKE A FIGURE WITH MAYAVI fig = mlab.figure(size=(500, 500), bgcolor=(1,1,1)) u = np.linspace(0,2*np.pi,100) xx,yy,zz = np.cos(u), np.sin(3*u), np.sin(u) # Points l = mlab.plot3d(xx,yy,zz, representation="wireframe", tube_sides=5, line_width=.5, tube_radius=0.2, figure=fig) # ANIMATE THE FIGURE WITH MOVIEPY, WRITE AN ANIMATED GIF def make_frame(t): """ Generates and returns the frame for time t. """ y = np.sin(3*u)*(0.2+0.5*np.cos(2*np.pi*t/duration)) l.mlab_source.set(y = y) # change y-coordinates of the mesh mlab.view(azimuth= 360*t/duration, distance=9) # camera angle return mlab.screenshot(antialiased=True) # return a RGB image animation = mpy.VideoClip(make_frame, duration=duration).resize(0.5) # Video generation takes 10 seconds, GIF generation takes 25s animation.write_videofile("wireframe.mp4", fps=20) animation.write_gif("wireframe.gif", fps=20) |
As Mayavi relies on the powerful ITK visualization engine it can also process complex datasets. Here is an animation derived from aMayavi example:ide
Animations with Vispy
Vispy is another interactive 3D data visualization library, based on OpenGL. As for Mayavi, we first create a figure and a mesh, that we animate with MoviePy.oop
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 |
from moviepy.editor import VideoClip import numpy as np from vispy import app, scene from vispy.gloo.util import _screenshot canvas = scene.SceneCanvas(keys='interactive') view = canvas.central_widget.add_view() view.set_camera('turntable', mode='perspective', up='z', distance=2, azimuth=30., elevation=65.) xx, yy = np.arange(-1,1,.02),np.arange(-1,1,.02) X,Y = np.meshgrid(xx,yy) R = np.sqrt(X**2+Y**2) Z = lambda t : 0.1*np.sin(10*R-2*np.pi*t) surface = scene.visuals.SurfacePlot(x= xx-0.1, y=yy+0.2, z= Z(0), shading='smooth', color=(0.5, 0.5, 1, 1)) view.add(surface) canvas.show() # ANIMATE WITH MOVIEPY def make_frame(t): surface.set_data(z = Z(t)) # Update the mathematical surface canvas.on_draw(None) # Update the image on Vispy's canvas return _screenshot((0,0,canvas.size[0],canvas.size[1]))[:,:,:3] animation = VideoClip(make_frame, duration=1).resize(width=350) animation.write_gif('sinc_vispy.gif', fps=20, opt='OptimizePlus') |
Here are more advanced examples (derived from the Vispy gallery) where C code snippets are embedded in the Python code to fine-tune the 3D shaders:post
Animations with Matplotlib
The 2D/3D plotting library Matplotlib already has an animation module, but I found that MoviePy produces lighter, better quality videos, while being up to two times faster (not sure why, seehere for more details). Here is how you animate Matplotlib with MoviePy:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 |
import matplotlib.pyplot as plt import numpy as np from moviepy.video.io.bindings import mplfig_to_npimage import moviepy.editor as mpy # DRAW A FIGURE WITH MATPLOTLIB duration = 2 fig_mpl, ax = plt.subplots(1,figsize=(5,3), facecolor='white') xx = np.linspace(-2,2,200) # the x vector zz = lambda d: np.sinc(xx**2)+np.sin(xx+d) # the (changing) z vector ax.set_title("Elevation in y=0") ax.set_ylim(-1.5,2.5) line, = ax.plot(xx, zz(0), lw=3) # ANIMATE WITH MOVIEPY (UPDATE THE CURVE FOR EACH t). MAKE A GIF. def make_frame_mpl(t): line.set_ydata( zz(2*np.pi*t/duration)) # <= Update the curve return mplfig_to_npimage(fig_mpl) # RGB image of the figure animation =mpy.VideoClip(make_frame_mpl, duration=duration) animation.write_gif("sinc_mpl.gif", fps=20) |
Matplotlib has many beautiful themes and works well with numerical modules like Pandas or Scikit-Learn. Let us watch a SVM classifier getting a better understanding of the map as the number of training point increases.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 |
import numpy as np import matplotlib.pyplot as plt from sklearn import svm # sklearn = scikit-learn from sklearn.datasets import make_moons from moviepy.editor import VideoClip from moviepy.video.io.bindings import mplfig_to_npimage X, Y = make_moons(50, noise=0.1, random_state=2) # semi-random data fig, ax = plt.subplots(1, figsize=(4, 4), facecolor=(1,1,1)) fig.subplots_adjust(left=0, right=1, bottom=0) xx, yy = np.meshgrid(np.linspace(-2,3,500), np.linspace(-1,2,500)) def make_frame(t): ax.clear() ax.axis('off') ax.set_title("SVC classification", fontsize=16) classifier = svm.SVC(gamma=2, C=1) # the varying weights make the points appear one after the other weights = np.minimum(1, np.maximum(0, t**2+10-np.arange(50))) classifier.fit(X, Y, sample_weight=weights) Z = classifier.decision_function(np.c_[xx.ravel(), yy.ravel()]) Z = Z.reshape(xx.shape) ax.contourf(xx, yy, Z, cmap=plt.cm.bone, alpha=0.8, vmin=-2.5, vmax=2.5, levels=np.linspace(-2,2,20)) ax.scatter(X[:,0], X[:,1], c=Y, s=50*weights, cmap=plt.cm.bone) return mplfig_to_npimage(fig) animation = VideoClip(make_frame, duration = 7) animation.write_gif("svm.gif", fps=15) |
Put simply, the background colors tell us where the classifier thinks the black points and white points belong. At the begining it has no real clue, but as more points appear it progressively understands that they are distributed along moon-shaped regions.
Animations with Numpy
If you are working with Numpy arrays (Numpy is the central numerical library in Python), you don’t need any external plotting library, you can feed the arrays directly to MoviePy.
This is well illustrated by this simulation of a zombie outbreak in France (inspired bythis blog post by Max Berggren). France is modelled as a grid (Numpy array) on which all the computations for dispersion and infection are done. At regular intervals, a few Numpy operations tranform the grid into a valid RGB image, and send it to MoviePy.
Putting animations together
What is better than an animation ? Two animations ! You can take advantage of MoviePy’s video composition capabilities to mix animations from different libraries:
1 2 3 4 5 6 |
import moviepy.editor as mpy # We use the GIFs generated earlier to avoid recomputing the animations. clip_mayavi = mpy.VideoFileClip("sinc.gif") clip_mpl = mpy.VideoFileClip("sinc_mpl.gif").resize(height=clip_mayavi.h) animation = mpy.clips_array([[clip_mpl, clip_mayavi]]) animation.write_gif("sinc_plot.gif", fps=20) |
Or for something more artistic:
1 2 3 4 5 6 7 8 |
# Make the white color transparent in clip_mayavi clip_mayavi2 = (clip_mayavi.fx( mpy.vfx.mask_color, [255,255,255]) .set_opacity(.4) # whole clip is semi-transparent .resize(height=0.85*clip_mpl.h) .set_pos('center')) animation = mpy.CompositeVideoClip([clip_mpl, clip_mayavi2]) animation.write_gif("sinc_plot2.gif", fps=20) |
It may be a tad too flashy, but sometimes you must give your audience something they can tweet.
You can also annotate the animations, which is useful when comparing different filters or algorithms. Let’s display four image transformations from the libraryScikit-image:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 |
import moviepy.editor as mpy import skimage.exposure as ske # rescaling, histogram eq. import skimage.filter as skf # gaussian blur clip = mpy.VideoFileClip("sinc.gif") gray = clip.fx(mpy.vfx.blackwhite).to_mask() def apply_effect(effect, title, **kw): """ Returns a clip with the effect applied and a title""" filtr = lambda im: effect(im, **kw) new_clip = gray.fl_image(filtr).to_RGB() txt = (mpy.TextClip(title, font="Purisa-Bold", fontsize=15) .set_position(("center","top")) .set_duration(clip.duration)) return mpy.CompositeVideoClip([new_clip,txt]) # Apply 4 different effects to the original animation equalized = apply_effect(ske.equalize_hist, "Equalized") rescaled = apply_effect(ske.rescale_intensity, "Rescaled") adjusted = apply_effect(ske.adjust_log, "Adjusted") blurred = apply_effect(skf.gaussian_filter, "Blurred", sigma=4) # Put the clips together on a 2x2 grid, and write to a file. finalclip = mpy.clips_array([[ equalized, adjusted ], [ blurred, rescaled ]]) final_clip.write_gif("test2x2.gif", fps=20) |
If we replace CompositeVideoClip and clips_array by concatenate_videoclips we get a title-effect type animation:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 |
import moviepy.editor as mpy import skimage.exposure as ske import skimage.filter as skf clip = mpy.VideoFileClip("sinc.gif") gray = clip.fx(mpy.vfx.blackwhite).to_mask() def apply_effect(effect, label, **kw): """ Returns a clip with the effect applied and a top label""" filtr = lambda im: effect(im, **kw) new_clip = gray.fl_image(filtr).to_RGB() txt = (mpy.TextClip(label, font="Amiri-Bold", fontsize=25, bg_color='white', size=new_clip.size) .set_position(("center")) .set_duration(1)) return mpy.concatenate_videoclips([txt, new_clip]) equalized = apply_effect(ske.equalize_hist, "Equalized") rescaled = apply_effect(ske.rescale_intensity, "Rescaled") adjusted = apply_effect(ske.adjust_log, "Adjusted") blurred = apply_effect(skf.gaussian_filter, "Blurred", sigma=4) clips = [equalized, adjusted, blurred, rescaled] animation = mpy.concatenate_videoclips(clips) animation.write_gif("sinc_cat.gif", fps=15) |
Finally, MoviePy will be particularly practical when dealing with video data, as it is its first job. For our last example we estimate the size of a growing bacterial population by thresholding the video frames and counting the white pixels. The third panel shows that the population size grows exponentially in time.
One library to animate them all ?
I hope to have given you enough recipes to impress your colleagues at your next presentation. Any other library could be animated with MoviePy, as long as its output can be converted to a Numpy array.
Some libraries have their own animation modules, but these are usually a pain to fix and maintain. Thanks to the many users who have tested it in very different contexts, MoviePy seems to have become stable (or people stopped reporting bugs), and can be adapted to many situations. There is still a lot to do, but it would be nice if authors started relying on it for video and GIF rendering, like Pandas and Scikit-Learn rely on Matplotlib for plotting.
For completeness, and because it may better fit your needs, I must mention ImageIO, another Python library with video writing capabilities which focuses on providing a very simple interface to read or write any kind of image, video or volumetric data. For instance you useimwrite() to write any image, mimwrite() for any video/GIF,volwrite() for volumetric data, or simply write() for streamed data.
Cheers, and happy GIFing !