🧵🧶🪡Weave your frames into matplotlib animations!
matplotloom is published on PyPI so you can install matplotloom via pip
pip install matplotloom
or poetry
poetry add matplotloom
or conda
conda install matplotloom
matplotloom requires Python 3.9+ and is continuously tested on Linux, Windows, and Mac. Ensure you have ffmpeg
installed so that animations can be generated.
To visualize simulation output for computational fluid dynamics I've had to make long animations with complex figures for a long time. The animations consist of thousands of frames and the figures are too complex for FuncAnimation
and ArtistAnimation
. This package aims to simplify and massively speed up the process of making these kinds of animations.
- The main idea behind matplotloom is to describe how to generate each frame of your animation from scratch, instead of generating an animation by modifying one existing plot. This simplifies generating animations. See the examples below and how the code inside the
for
loops is plain and familiar matplotlib. It also ensures that every feature can be animated and that the generation process can be easily parallelized. - matplotlib has two tools for making animations:
FuncAnimation
andArtistAnimation
. But to use them you have to write your plotting code differently to modify an existing frame. This makes it difficult to go from plotting still figures to making animations. And some features are non-trivial to animate. - celluloid is a nice package for making matplotlib animations easily, but as it relies on
ArtistAnimation
under the hood it does come with some limitations such as not being able to animate titles. It also hasn't been maintained since 2018. - Plotting many frames (hundreds to thousands+) can be slow but with matplotloom you can use a parallel
Loom
to plot each frame in parallel, speeding up the animation process significantly especially if you can dedicate many cores to plotting.
- You can use
loom.show()
to display animations in Jupyter notebooks. - Anxious about animation progress? Pass
verbose=True
or use tqdm to monitor progress. - Animations taking too long to make or do you have tons of frames? You can parallelize frame creating by looming in parallel.
- You have to call
loom.save_frame(fig)
for each frame (see the examples). While theLoom
object can be made to do this automatically it would have to create and own theFigure
instance and I wanted full control over the creation of theFigure
for maximum flexibility. - matplotloom is going to be slow. But it's flexible and compatible with all of matplotlib! The real speedup comes from parallelizing frame creation, especially if you have a ton of frames to make.
import numpy as np
import matplotlib.pyplot as plt
from matplotloom import Loom
with Loom("sine_wave_animation.gif", fps=30) as loom:
for phase in np.linspace(0, 2*np.pi, 100):
fig, ax = plt.subplots()
x = np.linspace(0, 2*np.pi, 200)
y = np.sin(x + phase)
ax.plot(x, y)
ax.set_xlim(0, 2*np.pi)
loom.save_frame(fig)
import numpy as np
import matplotlib.pyplot as plt
from matplotloom import Loom
with Loom("rotating_circular_sine_wave.mp4", fps=10) as loom:
for i in range(36):
fig, ax = plt.subplots(figsize=(12, 8), subplot_kw={"projection": "3d"})
X = np.arange(-5, 5, 0.25)
Y = np.arange(-5, 5, 0.25)
X, Y = np.meshgrid(X, Y)
R = np.sqrt(X**2 + Y**2)
Z = np.sin(R)
surf = ax.plot_surface(X, Y, Z, cmap="coolwarm")
ax.view_init(azim=i*10)
ax.set_zlim(-1.01, 1.01)
fig.colorbar(surf, shrink=0.5, aspect=5)
loom.save_frame(fig)
rotating_circular_sine_wave.mp4
Compare with animatplot's blocks example. With matplotloom you just use regular matplotlib abstractions.
import numpy as np
import matplotlib.pyplot as plt
from scipy.special import j0
from cmocean import cm
from matplotloom import Loom
def bessel_wave(r, t, k, omega, A):
return A * j0(k*r - omega*t)
def create_frame(x, y, t):
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 5))
r = np.sqrt(x**2 + y**2)
z = bessel_wave(r, t, k=2, omega=1, A=1)
pcm = ax1.pcolormesh(x, y, z, cmap=cm.balance, shading='auto', vmin=-1, vmax=1)
fig.colorbar(pcm, ax=ax1)
ax1.set_title(f"Bessel wave: t = {t:.3f}")
ax1.set_xlabel("x")
ax1.set_ylabel("y")
ax1.set_xlim(-10, 10)
ax1.set_ylim(-10, 10)
ax1.set_aspect("equal", adjustable="box")
mid = z.shape[0] // 2
ax2.plot(x[mid], z[mid])
ax2.set_xlim(x.min(), x.max())
ax2.set_ylim(-1.1, 1.1)
ax2.set_title("Cross-section at y = 0")
ax2.set_xlabel("x")
ax2.set_ylabel("z")
return fig
loom = Loom(
"bessel_wave.mp4",
fps = 30,
overwrite = True,
verbose = True,
savefig_kwargs = {
"dpi": 100,
"bbox_inches": "tight"
}
)
with loom:
x = np.linspace(-10, 10, 500)
y = np.linspace(-10, 10, 500)
x, y = np.meshgrid(x, y)
for t in np.linspace(0, 50, 300):
fig = create_frame(x, y, t)
loom.save_frame(fig)
bessel_wave.mp4
Compare with matplotlib's double pendulum example.
import numpy as np
import matplotlib.pyplot as plt
from scipy.integrate import solve_ivp
from tqdm import tqdm
from matplotloom import Loom
g = 9.80665 # standard acceleration of gravity [m/s²]
l1, l2 = 1, 1 # pendulum arms lengths [m]
m1, m2 = 1, 1 # pendulum masses [kg]
# Calculate dy/dt where y = [θ1, ω1, θ2, ω2].
def derivatives(t, state):
θ1, ω1, θ2, ω2 = state
dydt = np.zeros_like(state)
dydt[0] = ω1
dydt[2] = ω2
Δθ = θ2 - θ1
denominator1 = (m1 + m2) * l1 - m2 * l1 * np.cos(Δθ)**2
dydt[1] = (m2 * l1 * ω1**2 * np.sin(Δθ) * np.cos(Δθ)
+ m2 * g * np.sin(θ2) * np.cos(Δθ)
+ m2 * l2 * ω2**2 * np.sin(Δθ)
- (m1 + m2) * g * np.sin(θ1)) / denominator1
denominator2 = (l2 / l1) * denominator1
dydt[3] = (-m2 * l2 * ω2**2 * np.sin(Δθ) * np.cos(Δθ)
+ (m1 + m2) * g * np.sin(θ1) * np.cos(Δθ)
- (m1 + m2) * l1 * ω1**2 * np.sin(Δθ)
- (m1 + m2) * g * np.sin(θ2)) / denominator2
return dydt
t_span = (0, 20)
y0 = [np.pi/2, 0, np.pi/2, 0]
sol = solve_ivp(derivatives, t_span, y0, dense_output=True)
times = np.linspace(t_span[0], t_span[1], 1000)
θ1, ω1, θ2, ω2 = sol.sol(times)
x1 = l1 * np.sin(θ1)
y1 = -l1 * np.cos(θ1)
x2 = x1 + l2 * np.sin(θ2)
y2 = y1 - l2 * np.cos(θ2)
loom = Loom(
"double_pendulum.mp4",
fps = 60,
overwrite = True,
savefig_kwargs = {"bbox_inches": "tight"}
)
with loom:
for i, t in tqdm(enumerate(times), total=len(times)):
fig, ax = plt.subplots(figsize=(8, 8))
ax.plot(
[0, x1[i], x2[i]],
[0, y1[i], y2[i]],
linestyle = "solid",
marker = "o",
color = "black",
linewidth = 3
)
ax.plot(
x2[:i+1],
y2[:i+1],
linestyle = "solid",
linewidth = 2,
color = "red",
alpha = 0.5
)
ax.set_title(f"Double Pendulum: t = {t:.3f}s")
ax.set_xlim(-2.2, 2.2)
ax.set_ylim(-2.2, 2.2)
ax.set_aspect("equal", adjustable="box")
loom.save_frame(fig)
double_pendulum.mp4
matplotloom works out of the box with anything that is built on top of matplotlib. Here we're extending a Cartopy example.
import datetime
import matplotlib.pyplot as plt
import cartopy.crs as ccrs
from cartopy.feature.nightshade import Nightshade
from joblib import Parallel, delayed
from matplotloom import Loom
def plot_frame(day_of_year, loom, frame_number):
date = datetime.datetime(2024, 1, 1, 12) + datetime.timedelta(days=day_of_year-1)
fig = plt.figure(figsize=(15, 5))
proj1 = ccrs.Orthographic(central_longitude=0, central_latitude=30)
proj2 = ccrs.Orthographic(central_longitude=120, central_latitude=0)
proj3 = ccrs.Orthographic(central_longitude=240, central_latitude=-30)
ax1 = fig.add_subplot(1, 3, 1, projection=proj1)
ax2 = fig.add_subplot(1, 3, 2, projection=proj2)
ax3 = fig.add_subplot(1, 3, 3, projection=proj3)
fig.suptitle(f"Night time shading for {date} UTC")
ax1.stock_img()
ax1.add_feature(Nightshade(date, alpha=0.2))
ax2.stock_img()
ax2.add_feature(Nightshade(date, alpha=0.2))
ax3.stock_img()
ax3.add_feature(Nightshade(date, alpha=0.2))
loom.save_frame(fig, frame_number)
loom = Loom(
"night_time_shading.mp4",
fps = 10,
overwrite = True,
parallel = True,
verbose = True,
savefig_kwargs = {
"bbox_inches": "tight"
}
)
with loom:
n_days_2024 = 366
days_of_year = range(1, n_days_2024 + 1)
Parallel(n_jobs=-1)(
delayed(plot_frame)(day_of_year, loom, i)
for i, day_of_year in enumerate(days_of_year)
)
night_time_shading.mp4
By passing parallel=True
when creating a Loom
, you can save frames using loom.save_frame(fig, frame_number)
which allows you to plot and save all your frames in parallel. One easy way to leverage this is by using joblib to parallelize the for loop. For example, here's how you can parallelize the simple sine wave example:
import numpy as np
import matplotlib.pyplot as plt
from matplotloom import Loom
from joblib import Parallel, delayed
def plot_frame(phase, frame_number, loom):
fig, ax = plt.subplots()
x = np.linspace(0, 2*np.pi, 200)
y = np.sin(x + phase)
ax.plot(x, y)
ax.set_xlim(0, 2*np.pi)
loom.save_frame(fig, frame_number)
with Loom("parallel_sine_wave.gif", fps=30, parallel=True) as loom:
phases = np.linspace(0, 2*np.pi, 100)
Parallel(n_jobs=-1)(
delayed(plot_frame)(phase, i, loom)
for i, phase in enumerate(phases)
)