3b1b-manim/from_3b1b/active/bayes/beta3.py
2020-04-12 09:54:20 -07:00

2150 lines
63 KiB
Python

from manimlib.imports import *
from from_3b1b.active.bayes.beta_helpers import *
from from_3b1b.active.bayes.beta1 import *
from from_3b1b.active.bayes.beta2 import ShowLimitToPdf
import scipy.stats
OUTPUT_DIRECTORY = "bayes/beta3"
class RemindOfWeightedCoin(Scene):
def construct(self):
# Largely copied from beta2
# Prob label
p_label = get_prob_coin_label()
p_label.set_height(0.7)
p_label.to_edge(UP)
rhs = p_label[-1]
q_box = get_q_box(rhs)
p_label.add(q_box)
self.add(p_label)
# Coin grid
def get_random_coin_grid(p):
bools = np.random.random(100) < p
grid = get_coin_grid(bools)
return grid
grid = get_random_coin_grid(0.5)
grid.next_to(p_label, DOWN, MED_LARGE_BUFF)
self.play(LaggedStartMap(
FadeIn, grid,
lag_ratio=2 / len(grid),
run_time=3,
))
self.wait()
# Label as h
brace = Brace(q_box, DOWN, buff=SMALL_BUFF)
h_label = TexMobject("h")
h_label.next_to(brace, DOWN)
eq = TexMobject("=")
eq.next_to(h_label, RIGHT)
h_decimal = DecimalNumber(0.5)
h_decimal.next_to(eq, RIGHT)
self.play(
GrowFromCenter(brace),
FadeInFrom(h_label, UP),
grid.scale, 0.8, {"about_edge": DOWN},
)
self.wait()
# Alternate weightings
tail_grid = get_random_coin_grid(0)
head_grid = get_random_coin_grid(1)
grid70 = get_random_coin_grid(0.7)
alt_grids = [tail_grid, head_grid, grid70]
for ag in alt_grids:
ag.replace(grid)
for coins in [grid, *alt_grids]:
for coin in coins:
coin.generate_target()
coin.target.rotate(90 * DEGREES, axis=UP)
coin.target.set_opacity(0)
def get_grid_swap_anims(g1, g2):
return [
LaggedStartMap(MoveToTarget, g1, lag_ratio=0.02, run_time=1.5, remover=True),
LaggedStartMap(MoveToTarget, g2, lag_ratio=0.02, run_time=1.5, rate_func=reverse_smooth),
]
self.play(
FadeIn(eq),
UpdateFromAlphaFunc(h_decimal, lambda m, a: m.set_opacity(a)),
ChangeDecimalToValue(h_decimal, 0, run_time=2),
*get_grid_swap_anims(grid, tail_grid)
)
self.wait()
self.play(
ChangeDecimalToValue(h_decimal, 1, run_time=1.5),
*get_grid_swap_anims(tail_grid, head_grid)
)
self.wait()
self.play(
ChangeDecimalToValue(h_decimal, 0.7, run_time=1.5),
*get_grid_swap_anims(head_grid, grid70)
)
self.wait()
# Graph
axes = scaled_pdf_axes()
axes.to_edge(DOWN, buff=MED_SMALL_BUFF)
axes.y_axis.numbers.set_opacity(0)
axes.y_axis_label.set_opacity(0)
h_lines = VGroup()
for y in range(15):
h_line = Line(axes.c2p(0, y), axes.c2p(1, y))
h_lines.add(h_line)
h_lines.set_stroke(WHITE, 0.5, opacity=0.5)
axes.add(h_lines)
x_axis_label = p_label[:4].copy()
x_axis_label.set_height(0.4)
x_axis_label.next_to(axes.c2p(1, 0), UR, buff=SMALL_BUFF)
axes.x_axis.add(x_axis_label)
n_heads_tracker = ValueTracker(3)
n_tails_tracker = ValueTracker(3)
def get_graph(axes=axes, nht=n_heads_tracker, ntt=n_tails_tracker):
dist = scipy.stats.beta(nht.get_value() + 1, ntt.get_value() + 1)
graph = axes.get_graph(dist.pdf, step_size=0.05)
graph.set_stroke(BLUE, 3)
graph.set_fill(BLUE_E, 1)
return graph
graph = always_redraw(get_graph)
area_label = TextMobject("Area = 1")
area_label.set_height(0.5)
area_label.move_to(axes.c2p(0.5, 1))
# pdf label
pdf_label = TextMobject("probability ", "density ", "function")
pdf_label.next_to(axes.input_to_graph_point(0.5, graph), UP)
pdf_target_template = TextMobject("p", "d", "f")
pdf_target_template.next_to(axes.input_to_graph_point(0.7, graph), UR)
pdf_label.generate_target()
for part, letter2 in zip(pdf_label.target, pdf_target_template):
for letter1 in part:
letter1.move_to(letter2)
part[1:].set_opacity(0)
# Add plot
self.add(axes, *self.mobjects)
self.play(
FadeOut(eq),
FadeOut(h_decimal),
LaggedStartMap(MoveToTarget, grid70, run_time=1, remover=True),
FadeIn(axes),
)
self.play(
DrawBorderThenFill(graph),
FadeIn(area_label, rate_func=squish_rate_func(smooth, 0.5, 1), run_time=2),
Write(pdf_label, run_time=1),
)
self.wait()
# Region
lh_tracker = ValueTracker(0.7)
rh_tracker = ValueTracker(0.7)
def get_region(axes=axes, graph=graph, lh_tracker=lh_tracker, rh_tracker=rh_tracker):
lh = lh_tracker.get_value()
rh = rh_tracker.get_value()
region = get_region_under_curve(axes, graph, lh, rh)
region.set_fill(GREY, 0.85)
region.set_stroke(YELLOW, 1)
return region
region = always_redraw(get_region)
region_area_label = DecimalNumber(num_decimal_places=3)
region_area_label.next_to(axes.c2p(0.7, 0), UP, MED_LARGE_BUFF)
def update_ra_label(label, nht=n_heads_tracker, ntt=n_tails_tracker, lht=lh_tracker, rht=rh_tracker):
dist = scipy.stats.beta(nht.get_value() + 1, ntt.get_value() + 1)
area = dist.cdf(rht.get_value()) - dist.cdf(lht.get_value())
label.set_value(area)
region_area_label.add_updater(update_ra_label)
range_label = VGroup(
TexMobject("0.6 \\le"),
p_label[:4].copy(),
TexMobject("\\le 0.8"),
)
for mob in range_label:
mob.set_height(0.4)
range_label.arrange(RIGHT, buff=SMALL_BUFF)
pp_label = VGroup(
TexMobject("P("),
range_label,
TexMobject(")"),
)
for mob in pp_label[::2]:
mob.set_height(0.7)
mob.set_color(YELLOW)
pp_label.arrange(RIGHT, buff=SMALL_BUFF)
pp_label.move_to(axes.c2p(0.3, 3))
self.play(
FadeIn(pp_label[::2]),
MoveToTarget(pdf_label),
FadeOut(area_label),
)
self.wait()
self.play(TransformFromCopy(p_label[:4], range_label[1]))
self.wait()
self.play(TransformFromCopy(axes.x_axis.numbers[2], range_label[0]))
self.play(TransformFromCopy(axes.x_axis.numbers[3], range_label[2]))
self.wait()
self.add(region)
self.play(
lh_tracker.set_value, 0.6,
rh_tracker.set_value, 0.8,
UpdateFromAlphaFunc(
region_area_label,
lambda m, a: m.set_opacity(a),
rate_func=squish_rate_func(smooth, 0.25, 1)
),
run_time=3,
)
self.wait()
# 7/10 heads
bools = [True] * 7 + [False] * 3
random.shuffle(bools)
coins = VGroup(*[
get_coin("H" if heads else "T")
for heads in bools
])
coins.arrange(RIGHT)
coins.set_height(0.7)
coins.next_to(h_label, DOWN, buff=MED_LARGE_BUFF)
heads = [c for c in coins if c.symbol == "H"]
numbers = VGroup(*[
Integer(i + 1).set_height(0.2).next_to(coin, DOWN, SMALL_BUFF)
for i, coin in enumerate(heads)
])
for coin in coins:
coin.save_state()
coin.rotate(90 * DEGREES, UP)
coin.set_opacity(0)
pp_label.generate_target()
pp_label.target.set_height(0.5)
pp_label.target.next_to(axes.c2p(0, 2), RIGHT, MED_LARGE_BUFF)
self.play(
LaggedStartMap(Restore, coins),
MoveToTarget(pp_label),
run_time=1,
)
self.play(ShowIncreasingSubsets(numbers))
self.wait()
# Move plot
self.play(
n_heads_tracker.set_value, 7,
n_tails_tracker.set_value, 3,
FadeOut(pdf_label, rate_func=squish_rate_func(smooth, 0, 0.5)),
run_time=2
)
self.wait()
# How does the answer change with more data
new_bools = [True] * 63 + [False] * 27
random.shuffle(new_bools)
bools = [c.symbol == "H" for c in coins] + new_bools
grid = get_coin_grid(bools)
grid.set_height(3.5)
grid.next_to(axes.c2p(0, 3), RIGHT, MED_LARGE_BUFF)
self.play(
FadeOut(numbers),
ReplacementTransform(coins, grid[:10]),
)
self.play(
FadeIn(grid[10:], lag_ratio=0.1, rate_func=linear),
pp_label.next_to, grid, DOWN,
)
self.wait()
self.add(graph, region, region_area_label, p_label, q_box, brace, h_label)
self.play(
n_heads_tracker.set_value, 70,
n_tails_tracker.set_value, 30,
)
self.wait()
origin = axes.c2p(0, 0)
self.play(
axes.y_axis.stretch, 0.5, 1, {"about_point": origin},
h_lines.stretch, 0.5, 1, {"about_point": origin},
)
self.wait()
# Shift the shape around
pairs = [
(70 * 3, 30 * 3),
(35, 15),
(35 + 20, 15 + 20),
(7, 3),
(70, 30),
]
for nh, nt in pairs:
self.play(
n_heads_tracker.set_value, nh,
n_tails_tracker.set_value, nt,
run_time=2,
)
self.wait()
# End
self.embed()
class LastTimeWrapper(Scene):
def construct(self):
fs_rect = FullScreenFadeRectangle(fill_opacity=1, fill_color=GREY_E)
self.add(fs_rect)
title = TextMobject("Last Time")
title.scale(1.5)
title.to_edge(UP)
rect = ScreenRectangle()
rect.set_height(6)
rect.set_fill(BLACK, 1)
rect.next_to(title, DOWN)
self.play(
DrawBorderThenFill(rect),
FadeInFromDown(title),
)
self.wait()
class ComplainAboutSimplisticModel(ExternallyAnimatedScene):
pass
class BayesianFrequentistDivide(Scene):
def construct(self):
# Setup Bayesian vs. Frequentist divide
b_label = TextMobject("Bayesian")
f_label = TextMobject("Frequentist")
labels = VGroup(b_label, f_label)
for label, vect in zip(labels, [LEFT, RIGHT]):
label.set_height(0.7)
label.move_to(vect * FRAME_WIDTH / 4)
label.to_edge(UP, buff=0.35)
h_line = Line(LEFT, RIGHT)
h_line.set_width(FRAME_WIDTH)
h_line.next_to(labels, DOWN)
v_line = Line(UP, DOWN)
v_line.set_height(FRAME_HEIGHT)
v_line.center()
for label in labels:
label.save_state()
label.set_y(0)
self.play(
FadeInFrom(label, -normalize(label.get_center())),
)
self.wait()
self.play(
ShowCreation(VGroup(v_line, h_line)),
*map(Restore, labels),
)
self.wait()
# Overlay ShowBayesianUpdating in editing
# Frequentist list (ignore?)
kw = {
"tex_to_color_map": {
"$p$-value": YELLOW,
"$H_0$": PINK,
"$\\alpha$": BLUE,
},
"alignment": "",
}
freq_list = VGroup(
TextMobject("1. State a null hypothesis $H_0$", **kw),
TextMobject("2. Choose a test statistic,\\\\", "$\\qquad$ compute its value", **kw),
TextMobject("3. Calculate a $p$-value", **kw),
TextMobject("4. Choose a significance value $\\alpha$", **kw),
TextMobject("5. Reject $H_0$ if $p$-value\\\\", "$\\qquad$ is less than $\\alpha$", **kw),
)
freq_list.set_width(0.5 * FRAME_WIDTH - 1)
freq_list.arrange(DOWN, buff=MED_LARGE_BUFF, aligned_edge=LEFT)
freq_list.move_to(FRAME_WIDTH * RIGHT / 4)
freq_list.to_edge(DOWN, buff=LARGE_BUFF)
# Frequentist icon
axes = get_beta_dist_axes(y_max=5, y_unit=1)
axes.set_width(0.5 * FRAME_WIDTH - 1)
axes.move_to(FRAME_WIDTH * RIGHT / 4 + DOWN)
dist = scipy.stats.norm(0.5, 0.1)
graph = axes.get_graph(dist.pdf)
graphs = VGroup()
for x_min, x_max in [(0, 0.3), (0.3, 0.7), (0.7, 1.0)]:
graph = axes.get_graph(dist.pdf, x_min=x_min, x_max=x_max)
graph.add_line_to(axes.c2p(x_max, 0))
graph.add_line_to(axes.c2p(x_min, 0))
graph.add_line_to(graph.get_start())
graphs.add(graph)
graphs.set_stroke(width=0)
graphs.set_fill(RED, 1)
graphs[1].set_fill(GREY_D, 1)
H_words = VGroup(*[TextMobject("Reject\\\\$H_0$") for x in range(2)])
for H_word, graph, vect in zip(H_words, graphs[::2], [RIGHT, LEFT]):
H_word.next_to(graph, UP, MED_LARGE_BUFF)
arrow = Arrow(
H_word.get_bottom(),
graph.get_center() + 0.75 * vect,
buff=SMALL_BUFF
)
H_word.add(arrow)
H_words.set_color(RED)
self.add(H_words)
self.add(axes)
self.add(graphs)
self.embed()
# Transition to 2x2
# Go back to prior
# Label uniform prior
# Talk about real coin prior
# Update ad infinitum
class ArgumentBetweenBayesianAndFrequentist(Scene):
def construct(self):
pass
# From version 1
class ShowBayesianUpdating(Scene):
CONFIG = {
"true_p": 0.72,
"random_seed": 4,
"initial_axis_scale_factor": 3.5
}
def construct(self):
# Axes
axes = scaled_pdf_axes(self.initial_axis_scale_factor)
self.add(axes)
# Graph
n_heads = 0
n_tails = 0
graph = get_beta_graph(axes, n_heads, n_tails)
self.add(graph)
# Get coins
true_p = self.true_p
bool_values = np.random.random(100) < true_p
bool_values[1] = True
coins = self.get_coins(bool_values)
coins.next_to(axes.y_axis, RIGHT, MED_LARGE_BUFF)
coins.to_edge(UP, LARGE_BUFF)
# Probability label
p_label, prob, prob_box = self.get_probability_label()
self.add(p_label)
self.add(prob_box)
# Slow animations
def head_likelihood(x):
return x
def tail_likelihood(x):
return 1 - x
n_previews = 10
n_slow_previews = 5
for x in range(n_previews):
coin = coins[x]
is_heads = bool_values[x]
new_data_label = TextMobject("New data")
new_data_label.set_height(0.3)
arrow = Vector(0.5 * UP)
arrow.next_to(coin, DOWN, SMALL_BUFF)
new_data_label.next_to(arrow, DOWN, SMALL_BUFF)
new_data_label.shift(MED_SMALL_BUFF * RIGHT)
if is_heads:
line = axes.get_graph(lambda x: x)
label = TexMobject("\\text{Scale by } x")
likelihood = head_likelihood
n_heads += 1
else:
line = axes.get_graph(lambda x: 1 - x)
label = TexMobject("\\text{Scale by } (1 - x)")
likelihood = tail_likelihood
n_tails += 1
label.next_to(graph, UP)
label.set_stroke(BLACK, 3, background=True)
line.set_stroke(YELLOW, 3)
graph_copy = graph.copy()
graph_copy.unlock_triangulation()
scaled_graph = graph.copy()
scaled_graph.apply_function(
lambda p: axes.c2p(
axes.x_axis.p2n(p),
axes.y_axis.p2n(p) * likelihood(axes.x_axis.p2n(p))
)
)
scaled_graph.set_color(GREEN)
renorm_label = TextMobject("Renormalize")
renorm_label.move_to(label)
new_graph = get_beta_graph(axes, n_heads, n_tails)
renormalized_graph = scaled_graph.copy()
renormalized_graph.match_style(graph)
renormalized_graph.match_height(new_graph, stretch=True, about_edge=DOWN)
if x < n_slow_previews:
self.play(
FadeInFromDown(coin),
FadeIn(new_data_label),
GrowArrow(arrow),
)
self.play(
FadeOut(new_data_label),
FadeOut(arrow),
ShowCreation(line),
FadeIn(label),
)
self.add(graph_copy, line, label)
self.play(Transform(graph_copy, scaled_graph))
self.play(
FadeOut(line),
FadeOut(label),
FadeIn(renorm_label),
)
self.play(
Transform(graph_copy, renormalized_graph),
FadeOut(graph),
)
self.play(FadeOut(renorm_label))
else:
self.add(coin)
graph_copy.become(scaled_graph)
self.add(graph_copy)
self.play(
Transform(graph_copy, renormalized_graph),
FadeOut(graph),
)
graph = new_graph
self.remove(graph_copy)
self.add(new_graph)
# Rescale y axis
axes.save_state()
sf = self.initial_axis_scale_factor
axes.y_axis.stretch(1 / sf, 1, about_point=axes.c2p(0, 0))
for number in axes.y_axis.numbers:
number.stretch(sf, 1)
axes.y_axis.numbers[:4].set_opacity(0)
self.play(
Restore(axes, rate_func=lambda t: smooth(1 - t)),
graph.stretch, 1 / sf, 1, {"about_edge": DOWN},
run_time=2,
)
# Fast animations
for x in range(n_previews, len(coins)):
coin = coins[x]
is_heads = bool_values[x]
if is_heads:
n_heads += 1
else:
n_tails += 1
new_graph = get_beta_graph(axes, n_heads, n_tails)
self.add(coins[:x + 1])
self.add(new_graph)
self.remove(graph)
self.wait(0.25)
# self.play(
# FadeIn(new_graph),
# run_time=0.25,
# )
# self.play(
# FadeOut(graph),
# run_time=0.25,
# )
graph = new_graph
# Show confidence interval
dist = scipy.stats.beta(n_heads + 1, n_tails + 1)
v_lines = VGroup()
labels = VGroup()
x_bounds = dist.interval(0.95)
for x in x_bounds:
line = DashedLine(
axes.c2p(x, 0),
axes.c2p(x, 12),
)
line.set_color(YELLOW)
v_lines.add(line)
label = DecimalNumber(x)
label.set_height(0.25)
label.next_to(line, UP)
label.match_color(line)
labels.add(label)
true_graph = axes.get_graph(dist.pdf)
region = get_region_under_curve(axes, true_graph, *x_bounds)
region.set_fill(GREY_BROWN, 0.85)
region.set_stroke(YELLOW, 1)
label95 = TexMobject("95\\%")
fix_percent(label95.family_members_with_points()[-1])
label95.move_to(region, DOWN)
label95.shift(0.5 * UP)
self.play(*map(ShowCreation, v_lines))
self.play(
FadeIn(region),
Write(label95)
)
self.wait()
for label in labels:
self.play(FadeInFromDown(label))
self.wait()
# Show true value
self.wait()
self.play(FadeOut(prob_box))
self.play(ShowCreationThenFadeAround(prob))
self.wait()
# Much more data
many_bools = np.hstack([
bool_values,
(np.random.random(1000) < true_p)
])
N_tracker = ValueTracker(100)
graph.N_tracker = N_tracker
graph.bools = many_bools
graph.axes = axes
graph.v_lines = v_lines
graph.labels = labels
graph.region = region
graph.label95 = label95
label95.width_ratio = label95.get_width() / region.get_width()
def update_graph(graph):
N = int(graph.N_tracker.get_value())
nh = sum(graph.bools[:N])
nt = len(graph.bools[:N]) - nh
new_graph = get_beta_graph(graph.axes, nh, nt, step_size=0.05)
graph.become(new_graph)
dist = scipy.stats.beta(nh + 1, nt + 1)
x_bounds = dist.interval(0.95)
for x, line, label in zip(x_bounds, graph.v_lines, graph.labels):
line.set_x(graph.axes.c2p(x, 0)[0])
label.set_x(graph.axes.c2p(x, 0)[0])
label.set_value(x)
graph.labels[0].shift(MED_SMALL_BUFF * LEFT)
graph.labels[1].shift(MED_SMALL_BUFF * RIGHT)
new_simple_graph = graph.axes.get_graph(dist.pdf)
new_region = get_region_under_curve(graph.axes, new_simple_graph, *x_bounds)
new_region.match_style(graph.region)
graph.region.become(new_region)
graph.label95.set_width(graph.label95.width_ratio * graph.region.get_width())
graph.label95.match_x(graph.region)
self.add(graph, region, label95, p_label)
self.play(
N_tracker.set_value, 1000,
UpdateFromFunc(graph, update_graph),
Animation(v_lines),
Animation(labels),
Animation(graph.region),
Animation(graph.label95),
run_time=5,
)
self.wait()
#
def get_coins(self, bool_values):
coins = VGroup(*[
get_coin("H" if heads else "T")
for heads in bool_values
])
coins.arrange_in_grid(n_rows=10, buff=MED_LARGE_BUFF)
coins.set_height(5)
return coins
def get_probability_label(self):
head = get_coin("H")
p_label = TexMobject(
"P(00) = ",
tex_to_color_map={"00": WHITE}
)
template = p_label.get_part_by_tex("00")
head.replace(template)
p_label.replace_submobject(
p_label.index_of_part(template),
head,
)
prob = DecimalNumber(self.true_p)
prob.next_to(p_label, RIGHT)
p_label.add(prob)
p_label.set_height(0.75)
p_label.to_corner(UR)
prob_box = SurroundingRectangle(prob, buff=SMALL_BUFF)
prob_box.set_fill(GREY_D, 1)
prob_box.set_stroke(WHITE, 2)
q_marks = TexMobject("???")
q_marks.move_to(prob_box)
prob_box.add(q_marks)
return p_label, prob, prob_box
class HighlightReviewPartsReversed(HighlightReviewParts):
CONFIG = {
"reverse_order": True,
}
class Grey(Scene):
def construct(self):
self.add(FullScreenFadeRectangle(fill_color=GREY_D, fill_opacity=1))
class ShowBayesRule(Scene):
def construct(self):
hyp = "\\text{Hypothesis}"
data = "\\text{Data}"
bayes = TexMobject(
f"P({hyp} \\,|\\, {data})", "=", "{",
f"P({data} \\,|\\, {hyp})", f"P({hyp})",
"\\over", f"P({data})",
tex_to_color_map={
hyp: YELLOW,
data: GREEN,
}
)
title = TextMobject("Bayes' rule")
title.scale(2)
title.to_edge(UP)
self.add(title)
self.add(*bayes[:5])
self.wait()
self.play(
*[
TransformFromCopy(bayes[i], bayes[j], path_arc=30 * DEGREES)
for i, j in [
(0, 7),
(1, 10),
(2, 9),
(3, 8),
(4, 11),
]
],
FadeIn(bayes[5]),
run_time=1.5
)
self.wait()
self.play(
*[
TransformFromCopy(bayes[i], bayes[j], path_arc=30 * DEGREES)
for i, j in [
(0, 12),
(1, 13),
(4, 14),
(0, 16),
(3, 17),
(4, 18),
]
],
FadeIn(bayes[15]),
run_time=1.5
)
self.add(bayes)
self.wait()
hyp_word = bayes.get_part_by_tex(hyp)
example_hyp = TextMobject(
"For example,\\\\",
"$0.9 < s < 0.99$",
)
example_hyp[1].set_color(YELLOW)
example_hyp.next_to(hyp_word, DOWN, buff=1.5)
data_word = bayes.get_part_by_tex(data)
example_data = TexMobject(
"48\\,", CMARK_TEX,
"\\,2\\,", XMARK_TEX,
)
example_data.set_color_by_tex(CMARK_TEX, GREEN)
example_data.set_color_by_tex(XMARK_TEX, RED)
example_data.scale(1.5)
example_data.next_to(example_hyp, RIGHT, buff=1.5)
hyp_arrow = Arrow(
hyp_word.get_bottom(),
example_hyp.get_top(),
)
data_arrow = Arrow(
data_word.get_bottom(),
example_data.get_top(),
)
self.play(
GrowArrow(hyp_arrow),
FadeInFromPoint(example_hyp, hyp_word.get_center()),
)
self.wait()
self.play(
GrowArrow(data_arrow),
FadeInFromPoint(example_data, data_word.get_center()),
)
self.wait()
class VisualizeBayesRule(Scene):
def construct(self):
self.show_continuum()
self.show_arrows()
self.show_discrete_probabilities()
self.show_bayes_formula()
self.parallel_universes()
self.update_from_data()
def show_continuum(self):
axes = get_beta_dist_axes(y_max=1, y_unit=0.1)
axes.y_axis.add_numbers(
*np.arange(0.2, 1.2, 0.2),
number_config={
"num_decimal_places": 1,
}
)
p_label = TexMobject(
"P(s \\,|\\, \\text{data})",
tex_to_color_map={
"s": YELLOW,
"\\text{data}": GREEN,
}
)
p_label.scale(1.5)
p_label.to_edge(UP, LARGE_BUFF)
s_part = p_label.get_part_by_tex("s").copy()
x_line = Line(axes.c2p(0, 0), axes.c2p(1, 0))
x_line.set_stroke(YELLOW, 3)
arrow = Vector(DOWN)
arrow.next_to(s_part, DOWN, SMALL_BUFF)
value = DecimalNumber(0, num_decimal_places=4)
value.set_color(YELLOW)
value.next_to(arrow, DOWN)
self.add(axes)
self.add(p_label)
self.play(
s_part.next_to, x_line.get_start(), UR, SMALL_BUFF,
GrowArrow(arrow),
FadeInFromPoint(value, s_part.get_center()),
)
s_part.tracked = x_line
value.tracked = x_line
value.x_axis = axes.x_axis
self.play(
ShowCreation(x_line),
UpdateFromFunc(
s_part,
lambda m: m.next_to(m.tracked.get_end(), UR, SMALL_BUFF)
),
UpdateFromFunc(
value,
lambda m: m.set_value(
m.x_axis.p2n(m.tracked.get_end())
)
),
run_time=3,
)
self.wait()
self.play(
FadeOut(arrow),
FadeOut(value),
)
self.p_label = p_label
self.s_part = s_part
self.value = value
self.x_line = x_line
self.axes = axes
def show_arrows(self):
axes = self.axes
arrows = VGroup()
arrow_template = Vector(DOWN)
arrow_template.lock_triangulation()
def get_arrow(s, denom):
arrow = arrow_template.copy()
arrow.set_height(4 / denom)
arrow.move_to(axes.c2p(s, 0), DOWN)
arrow.set_color(interpolate_color(
GREY_A, GREY_C, random.random()
))
return arrow
for k in range(2, 50):
for n in range(1, k):
if np.gcd(n, k) != 1:
continue
s = n / k
arrows.add(get_arrow(s, k))
for k in range(50, 1000):
arrows.add(get_arrow(1 / k, k))
arrows.add(get_arrow(1 - 1 / k, k))
kw = {
"lag_ratio": 0.5,
"run_time": 5,
"rate_func": lambda t: t**4,
}
arrows.save_state()
for arrow in arrows:
arrow.stretch(0, 0)
arrow.set_stroke(width=0)
arrow.set_opacity(0)
self.play(Restore(arrows, **kw))
self.play(LaggedStartMap(
ApplyMethod, arrows,
lambda m: (m.scale, 0, {"about_edge": DOWN}),
**kw
))
self.remove(arrows)
self.wait()
def show_discrete_probabilities(self):
axes = self.axes
x_lines = VGroup()
dx = 0.01
for x in np.arange(0, 1, dx):
line = Line(
axes.c2p(x, 0),
axes.c2p(x + dx, 0),
)
line.set_stroke(BLUE, 3)
line.generate_target()
line.target.rotate(
90 * DEGREES,
about_point=line.get_start()
)
x_lines.add(line)
self.add(x_lines)
self.play(
FadeOut(self.x_line),
LaggedStartMap(
MoveToTarget, x_lines,
)
)
label = Integer(0)
label.set_height(0.5)
label.next_to(self.p_label[1], DOWN, LARGE_BUFF)
unit = TexMobject("\\%")
unit.match_height(label)
fix_percent(unit.family_members_with_points()[0])
always(unit.next_to, label, RIGHT, SMALL_BUFF)
arrow = Arrow()
arrow.max_stroke_width_to_length_ratio = 1
arrow.axes = axes
arrow.label = label
arrow.add_updater(lambda m: m.put_start_and_end_on(
m.label.get_bottom() + MED_SMALL_BUFF * DOWN,
m.axes.c2p(0.01 * m.label.get_value(), 0.03),
))
self.add(label, unit, arrow)
self.play(
ChangeDecimalToValue(label, 99),
run_time=5,
)
self.wait()
self.play(*map(FadeOut, [label, unit, arrow]))
# Show prior label
p_label = self.p_label
given_data = p_label[2:4]
prior_label = TexMobject("P(s)", tex_to_color_map={"s": YELLOW})
prior_label.match_height(p_label)
prior_label.move_to(p_label, DOWN, LARGE_BUFF)
p_label.save_state()
self.play(
given_data.scale, 0.5,
given_data.set_opacity, 0.5,
given_data.to_corner, UR,
Transform(p_label[:2], prior_label[:2]),
Transform(p_label[-1], prior_label[-1]),
)
self.wait()
# Zoom in on the y-values
new_ticks = VGroup()
new_labels = VGroup()
dy = 0.01
for y in np.arange(dy, 5 * dy, dy):
height = get_norm(axes.c2p(0, dy) - axes.c2p(0, 0))
tick = axes.y_axis.get_tick(y, SMALL_BUFF)
label = DecimalNumber(y)
label.match_height(axes.y_axis.numbers[0])
always(label.next_to, tick, LEFT, SMALL_BUFF)
new_ticks.add(tick)
new_labels.add(label)
for num in axes.y_axis.numbers:
height = num.get_height()
always(num.set_height, height, stretch=True)
bars = VGroup()
dx = 0.01
origin = axes.c2p(0, 0)
for x in np.arange(0, 1, dx):
rect = Rectangle(
width=get_norm(axes.c2p(dx, 0) - origin),
height=get_norm(axes.c2p(0, dy) - origin),
)
rect.x = x
rect.set_stroke(BLUE, 1)
rect.set_fill(BLUE, 0.5)
rect.move_to(axes.c2p(x, 0), DL)
bars.add(rect)
stretch_group = VGroup(
axes.y_axis,
bars,
new_ticks,
x_lines,
)
x_lines.set_height(
bars.get_height(),
about_edge=DOWN,
stretch=True,
)
self.play(
stretch_group.stretch, 25, 1, {"about_point": axes.c2p(0, 0)},
VFadeIn(bars),
VFadeIn(new_ticks),
VFadeIn(new_labels),
VFadeOut(x_lines),
run_time=4,
)
highlighted_bars = bars.copy()
highlighted_bars.set_color(YELLOW)
self.play(
LaggedStartMap(
FadeIn, highlighted_bars,
lag_ratio=0.5,
rate_func=there_and_back,
),
ShowCreationThenFadeAround(new_labels[0]),
run_time=3,
)
self.remove(highlighted_bars)
# Nmae as prior
prior_name = TextMobject("Prior", " distribution")
prior_name.set_height(0.6)
prior_name.next_to(prior_label, DOWN, LARGE_BUFF)
self.play(FadeInFromDown(prior_name))
self.wait()
# Show alternate distribution
bars.save_state()
for a, b in [(5, 2), (1, 6)]:
dist = scipy.stats.beta(a, b)
for bar, saved in zip(bars, bars.saved_state):
bar.target = saved.copy()
height = get_norm(axes.c2p(0.1 * dist.pdf(bar.x)) - axes.c2p(0, 0))
bar.target.set_height(height, about_edge=DOWN, stretch=True)
self.play(LaggedStartMap(MoveToTarget, bars, lag_ratio=0.00))
self.wait()
self.play(Restore(bars))
self.wait()
uniform_name = TextMobject("Uniform")
uniform_name.match_height(prior_name)
uniform_name.move_to(prior_name, DL)
uniform_name.shift(RIGHT)
uniform_name.set_y(bars.get_top()[1] + MED_SMALL_BUFF, DOWN)
self.play(
prior_name[0].next_to, uniform_name, RIGHT, MED_SMALL_BUFF, DOWN,
FadeOutAndShift(prior_name[1], RIGHT),
FadeInFrom(uniform_name, LEFT)
)
self.wait()
self.bars = bars
self.uniform_label = VGroup(uniform_name, prior_name[0])
def show_bayes_formula(self):
uniform_label = self.uniform_label
p_label = self.p_label
bars = self.bars
prior_label = VGroup(
p_label[0].deepcopy(),
p_label[1].deepcopy(),
p_label[4].deepcopy(),
)
eq = TexMobject("=")
likelihood_label = TexMobject(
"P(", "\\text{data}", "|", "s", ")",
)
likelihood_label.set_color_by_tex("data", GREEN)
likelihood_label.set_color_by_tex("s", YELLOW)
over = Line(LEFT, RIGHT)
p_data_label = TextMobject("P(", "\\text{data}", ")")
p_data_label.set_color_by_tex("data", GREEN)
for mob in [eq, likelihood_label, over, p_data_label]:
mob.scale(1.5)
mob.set_opacity(0.1)
eq.move_to(prior_label, LEFT)
over.set_width(
prior_label.get_width() +
likelihood_label.get_width() +
MED_SMALL_BUFF
)
over.next_to(eq, RIGHT, MED_SMALL_BUFF)
p_data_label.next_to(over, DOWN, MED_SMALL_BUFF)
likelihood_label.next_to(over, UP, MED_SMALL_BUFF, RIGHT)
self.play(
p_label.restore,
p_label.next_to, eq, LEFT, MED_SMALL_BUFF,
prior_label.next_to, over, UP, MED_SMALL_BUFF, LEFT,
FadeIn(eq),
FadeIn(likelihood_label),
FadeIn(over),
FadeIn(p_data_label),
FadeOut(uniform_label),
)
# Show new distribution
post_bars = bars.copy()
total_prob = 0
for bar, p in zip(post_bars, np.arange(0, 1, 0.01)):
prob = scipy.stats.binom(50, p).pmf(48)
bar.stretch(prob, 1, about_edge=DOWN)
total_prob += 0.01 * prob
post_bars.stretch(1 / total_prob, 1, about_edge=DOWN)
post_bars.stretch(0.25, 1, about_edge=DOWN) # Lie to fit on screen...
post_bars.set_color(MAROON_D)
post_bars.set_fill(opacity=0.8)
brace = Brace(p_label, DOWN)
post_word = brace.get_text("Posterior")
post_word.scale(1.25, about_edge=UP)
post_word.set_color(MAROON_D)
self.play(
ReplacementTransform(
bars.copy().set_opacity(0),
post_bars,
),
GrowFromCenter(brace),
FadeInFrom(post_word, 0.25 * UP)
)
self.wait()
self.play(
eq.set_opacity, 1,
likelihood_label.set_opacity, 1,
)
self.wait()
data = get_check_count_label(48, 2)
data.scale(1.5)
data.next_to(likelihood_label, DOWN, buff=2, aligned_edge=LEFT)
data_arrow = Arrow(
likelihood_label[1].get_bottom(),
data.get_top()
)
data_arrow.set_color(GREEN)
self.play(
GrowArrow(data_arrow),
GrowFromPoint(data, data_arrow.get_start()),
)
self.wait()
self.play(FadeOut(data_arrow))
self.play(
over.set_opacity, 1,
p_data_label.set_opacity, 1,
)
self.wait()
self.play(
FadeOut(brace),
FadeOut(post_word),
FadeOut(post_bars),
FadeOut(data),
p_label.set_opacity, 0.1,
eq.set_opacity, 0.1,
likelihood_label.set_opacity, 0.1,
over.set_opacity, 0.1,
p_data_label.set_opacity, 0.1,
)
self.bayes = VGroup(
p_label, eq,
prior_label, likelihood_label,
over, p_data_label
)
self.data = data
def parallel_universes(self):
bars = self.bars
cols = VGroup()
squares = VGroup()
sample_colors = color_gradient(
[GREEN_C, GREEN_D, GREEN_E],
100
)
for bar in bars:
n_rows = 12
col = VGroup()
for x in range(n_rows):
square = Rectangle(
width=bar.get_width(),
height=bar.get_height() / n_rows,
)
square.set_stroke(width=0)
square.set_fill(opacity=1)
square.set_color(random.choice(sample_colors))
col.add(square)
squares.add(square)
col.arrange(DOWN, buff=0)
col.move_to(bar)
cols.add(col)
squares.shuffle()
self.play(
LaggedStartMap(
VFadeInThenOut, squares,
lag_ratio=0.005,
run_time=3
)
)
self.remove(squares)
squares.set_opacity(1)
self.wait()
example_col = cols[95]
self.play(
bars.set_opacity, 0.25,
FadeIn(example_col, lag_ratio=0.1),
)
self.wait()
dist = scipy.stats.binom(50, 0.95)
for x in range(12):
square = random.choice(example_col).copy()
square.set_fill(opacity=0)
square.set_stroke(YELLOW, 2)
self.add(square)
nc = dist.ppf(random.random())
data = get_check_count_label(nc, 50 - nc)
data.next_to(example_col, UP)
self.add(square, data)
self.wait(0.5)
self.remove(square, data)
self.wait()
self.data.set_opacity(1)
self.play(
FadeIn(self.data),
FadeOut(example_col),
self.bayes[3].set_opacity, 1,
)
self.wait()
def update_from_data(self):
bars = self.bars
data = self.data
bayes = self.bayes
new_bars = bars.copy()
new_bars.set_stroke(opacity=1)
new_bars.set_fill(opacity=0.8)
for bar, p in zip(new_bars, np.arange(0, 1, 0.01)):
dist = scipy.stats.binom(50, p)
scalar = dist.pmf(48)
bar.stretch(scalar, 1, about_edge=DOWN)
self.play(
ReplacementTransform(
bars.copy().set_opacity(0),
new_bars
),
bars.set_fill, {"opacity": 0.1},
bars.set_stroke, {"opacity": 0.1},
run_time=2,
)
# Show example bar
bar95 = VGroup(
bars[95].copy(),
new_bars[95].copy()
)
bar95.save_state()
bar95.generate_target()
bar95.target.scale(2)
bar95.target.next_to(bar95, UP, LARGE_BUFF)
bar95.target.set_stroke(BLUE, 3)
ex_label = TexMobject("s", "=", "0.95")
ex_label.set_color(YELLOW)
ex_label.next_to(bar95.target, DOWN, submobject_to_align=ex_label[-1])
highlight = SurroundingRectangle(bar95, buff=0)
highlight.set_stroke(YELLOW, 2)
self.play(FadeIn(highlight))
self.play(
MoveToTarget(bar95),
FadeInFromDown(ex_label),
data.shift, LEFT,
)
self.wait()
side_brace = Brace(bar95[1], RIGHT, buff=SMALL_BUFF)
side_label = side_brace.get_text("0.26", buff=SMALL_BUFF)
self.play(
GrowFromCenter(side_brace),
FadeIn(side_label)
)
self.wait()
self.play(
FadeOut(side_brace),
FadeOut(side_label),
FadeOut(ex_label),
)
self.play(
bar95.restore,
bar95.set_opacity, 0,
)
for bar in bars[94:80:-1]:
highlight.move_to(bar)
self.wait(0.5)
self.play(FadeOut(highlight))
self.wait()
# Emphasize formula terms
tops = VGroup()
for bar, new_bar in zip(bars, new_bars):
top = Line(bar.get_corner(UL), bar.get_corner(UR))
top.set_stroke(YELLOW, 2)
top.generate_target()
top.target.move_to(new_bar, UP)
tops.add(top)
rect = SurroundingRectangle(bayes[2])
rect.set_stroke(YELLOW, 1)
rect.target = SurroundingRectangle(bayes[3])
rect.target.match_style(rect)
self.play(
ShowCreation(rect),
ShowCreation(tops),
)
self.wait()
self.play(
LaggedStartMap(
MoveToTarget, tops,
run_time=2,
lag_ratio=0.02,
),
MoveToTarget(rect),
)
self.play(FadeOut(tops))
self.wait()
# Show alternate priors
axes = self.axes
bar_groups = VGroup()
for bar, new_bar in zip(bars, new_bars):
bar_groups.add(VGroup(bar, new_bar))
bar_groups.save_state()
for a, b in [(5, 2), (7, 1)]:
dist = scipy.stats.beta(a, b)
for bar, saved in zip(bar_groups, bar_groups.saved_state):
bar.target = saved.copy()
height = get_norm(axes.c2p(0.1 * dist.pdf(bar[0].x)) - axes.c2p(0, 0))
height = max(height, 1e-6)
bar.target.set_height(height, about_edge=DOWN, stretch=True)
self.play(LaggedStartMap(MoveToTarget, bar_groups, lag_ratio=0))
self.wait()
self.play(Restore(bar_groups))
self.wait()
# Rescale
ex_p_label = TexMobject(
"P(s = 0.95 | 00000000) = ",
tex_to_color_map={
"s = 0.95": YELLOW,
"00000000": WHITE,
}
)
ex_p_label.scale(1.5)
ex_p_label.next_to(bars, UP, LARGE_BUFF)
ex_p_label.align_to(bayes, LEFT)
template = ex_p_label.get_part_by_tex("00000000")
template.set_opacity(0)
highlight = SurroundingRectangle(new_bars[95], buff=0)
highlight.set_stroke(YELLOW, 1)
self.remove(data)
self.play(
FadeIn(ex_p_label),
VFadeOut(data[0]),
data[1:].move_to, template,
FadeIn(highlight)
)
self.wait()
numer = new_bars[95].copy()
numer.set_stroke(YELLOW, 1)
denom = new_bars[80:].copy()
h_line = Line(LEFT, RIGHT)
h_line.set_width(3)
h_line.set_stroke(width=2)
h_line.next_to(ex_p_label, RIGHT)
self.play(
numer.next_to, h_line, UP,
denom.next_to, h_line, DOWN,
ShowCreation(h_line),
)
self.wait()
self.play(
denom.space_out_submobjects,
rate_func=there_and_back
)
self.play(
bayes[4].set_opacity, 1,
bayes[5].set_opacity, 1,
FadeOut(rect),
)
self.wait()
# Rescale
self.play(
FadeOut(highlight),
FadeOut(ex_p_label),
FadeOut(data),
FadeOut(h_line),
FadeOut(numer),
FadeOut(denom),
bayes.set_opacity, 1,
)
new_bars.unlock_shader_data()
self.remove(new_bars, *new_bars)
self.play(
new_bars.set_height, 5, {"about_edge": DOWN, "stretch": True},
new_bars.set_color, MAROON_D,
)
self.wait()
class UniverseOf95Percent(WhatsTheModel):
CONFIG = {"s": 0.95}
def construct(self):
self.introduce_buyer_and_seller()
for m, v in [(self.seller, RIGHT), (self.buyer, LEFT)]:
m.shift(v)
m.label.shift(v)
pis = VGroup(self.seller, self.buyer)
label = get_prob_positive_experience_label(True, True)
label[-1].set_value(self.s)
label.set_height(1)
label.next_to(pis, UP, LARGE_BUFF)
self.add(label)
for x in range(4):
self.play(*self.experience_animations(
self.seller, self.buyer, arc=30 * DEGREES, p=self.s
))
self.embed()
class UniverseOf50Percent(UniverseOf95Percent):
CONFIG = {"s": 0.5}
class OpenAndCloseAsideOnPdfs(Scene):
def construct(self):
labels = VGroup(
TextMobject("$\\langle$", "Aside on", " pdfs", "$\\rangle$"),
TextMobject("$\\langle$/", "Aside on", " pdfs", "$\\rangle$"),
)
labels.set_width(FRAME_WIDTH / 2)
for label in labels:
label.set_color_by_tex("pdfs", YELLOW)
self.play(FadeInFromDown(labels[0]))
self.wait()
self.play(Transform(*labels))
self.wait()
class BayesRuleWithPdf(ShowLimitToPdf):
def construct(self):
# Axes
axes = self.get_axes()
sf = 1.5
axes.y_axis.stretch(sf, 1, about_point=axes.c2p(0, 0))
for number in axes.y_axis.numbers:
number.stretch(1 / sf, 1)
self.add(axes)
# Formula
bayes = self.get_formula()
post = bayes[:5]
eq = bayes[5]
prior = bayes[6:9]
likelihood = bayes[9:14]
over = bayes[14]
p_data = bayes[15:]
self.play(FadeInFromDown(bayes))
self.wait()
# Prior
prior_graph = get_beta_graph(axes, 0, 0)
prior_graph_top = Line(
prior_graph.get_corner(UL),
prior_graph.get_corner(UR),
)
prior_graph_top.set_stroke(YELLOW, 3)
bayes.save_state()
bayes.set_opacity(0.2)
prior.set_opacity(1)
self.play(
Restore(bayes, rate_func=reverse_smooth),
FadeIn(prior_graph),
ShowCreation(prior_graph_top),
)
self.play(FadeOut(prior_graph_top))
self.wait()
# Scale Down
nh = 1
nt = 2
scaled_graph = axes.get_graph(
lambda x: scipy.stats.binom(3, x).pmf(1) + 1e-6
)
scaled_graph.set_stroke(GREEN)
scaled_region = get_region_under_curve(axes, scaled_graph, 0, 1)
def to_uniform(p, axes=axes):
return axes.c2p(
axes.x_axis.p2n(p),
int(axes.y_axis.p2n(p) != 0),
)
scaled_region.set_fill(opacity=0.75)
scaled_region.save_state()
scaled_region.apply_function(to_uniform)
self.play(
Restore(scaled_region),
UpdateFromAlphaFunc(
scaled_region,
lambda m, a: m.set_opacity(a * 0.75),
),
likelihood.set_opacity, 1,
)
self.wait()
# Rescale
new_graph = get_beta_graph(axes, nh, nt)
self.play(
ApplyMethod(
scaled_region.set_height, new_graph.get_height(),
{"about_edge": DOWN, "stretch": True},
run_time=2,
),
over.set_opacity, 1,
p_data.set_opacity, 1,
)
self.wait()
self.play(
post.set_opacity, 1,
eq.set_opacity, 1,
)
self.wait()
# Use lower case
new_bayes = self.get_formula(lowercase=True)
new_bayes.replace(bayes, dim_to_match=0)
rects = VGroup(
SurroundingRectangle(new_bayes[0][0]),
SurroundingRectangle(new_bayes[6][0]),
)
rects.set_stroke(YELLOW, 3)
self.remove(bayes)
bayes = self.get_formula()
bayes.unlock_triangulation()
self.add(bayes)
self.play(Transform(bayes, new_bayes))
self.play(ShowCreationThenFadeOut(rects))
def get_formula(self, lowercase=False):
p_sym = "p" if lowercase else "P"
bayes = TexMobject(
p_sym + "({s} \\,|\\, \\text{data})", "=",
"{" + p_sym + "({s})",
"P(\\text{data} \\,|\\, {s})",
"\\over",
"P(\\text{data})",
tex_to_color_map={
"{s}": YELLOW,
"\\text{data}": GREEN,
}
)
bayes.set_height(1.5)
bayes.to_edge(UP)
return bayes
class TalkThroughCoinExample(ShowBayesianUpdating):
def construct(self):
# Setup
axes = self.get_axes()
x_label = TexMobject("x")
x_label.next_to(axes.x_axis.get_end(), UR, MED_SMALL_BUFF)
axes.add(x_label)
p_label, prob, prob_box = self.get_probability_label()
prob_box_x = x_label.copy().move_to(prob_box)
self.add(axes)
self.add(p_label)
self.add(prob_box)
self.wait()
q_marks = prob_box[1]
prob_box.remove(q_marks)
self.play(
FadeOut(q_marks),
TransformFromCopy(x_label, prob_box_x)
)
prob_box.add(prob_box_x)
# Setup coins
bool_values = (np.random.random(100) < self.true_p)
bool_values[:5] = [True, False, True, True, False]
coins = self.get_coins(bool_values)
coins.next_to(axes.y_axis, RIGHT, MED_LARGE_BUFF)
coins.to_edge(UP)
# Random coin
rows = VGroup()
for x in range(5):
row = self.get_coins(np.random.random(10) < self.true_p)
row.arrange(RIGHT, buff=MED_LARGE_BUFF)
row.set_width(6)
row.move_to(UP)
rows.add(row)
last_row = VMobject()
for row in rows:
self.play(
FadeOutAndShift(last_row, DOWN),
FadeIn(row, lag_ratio=0.1)
)
last_row = row
self.play(FadeOutAndShift(last_row, DOWN))
# Uniform pdf
region = get_beta_graph(axes, 0, 0)
graph = Line(
region.get_corner(UL),
region.get_corner(UR),
)
func_label = TexMobject("f(x) =", "1")
func_label.next_to(graph, UP)
self.play(
FadeIn(func_label, lag_ratio=0.1),
ShowCreation(graph),
)
self.add(region, graph)
self.play(FadeIn(region))
self.wait()
# First flip
coin = coins[0]
arrow = Vector(0.5 * UP)
arrow.next_to(coin, DOWN, SMALL_BUFF)
data_label = TextMobject("New data")
data_label.set_height(0.25)
data_label.next_to(arrow, DOWN)
data_label.shift(0.5 * RIGHT)
self.play(
FadeInFrom(coin, DOWN),
GrowArrow(arrow),
Write(data_label, run_time=1)
)
self.wait()
# Show Bayes rule
bayes = TexMobject(
"p({x} | \\text{data})", "=",
"p({x})",
"{P(\\text{data} | {x})",
"\\over",
"P(\\text{data})",
tex_to_color_map={
"{x}": WHITE,
"\\text{data}": GREEN,
}
)
bayes.next_to(func_label, UP, LARGE_BUFF, LEFT)
likelihood = bayes[9:14]
p_data = bayes[15:]
likelihood_rect = SurroundingRectangle(likelihood, buff=0.05)
likelihood_rect.save_state()
p_data_rect = SurroundingRectangle(p_data, buff=0.05)
likelihood_x_label = TexMobject("x")
likelihood_x_label.next_to(likelihood_rect, UP)
self.play(FadeInFromDown(bayes))
self.wait()
self.play(ShowCreation(likelihood_rect))
self.wait()
self.play(TransformFromCopy(likelihood[-2], likelihood_x_label))
self.wait()
# Scale by x
times_x = TexMobject("\\cdot \\, x")
times_x.next_to(func_label, RIGHT, buff=0.2)
new_graph = axes.get_graph(lambda x: x)
sub_region = get_region_under_curve(axes, new_graph, 0, 1)
self.play(
Write(times_x),
Transform(graph, new_graph),
)
self.play(
region.set_opacity, 0.5,
FadeIn(sub_region),
)
self.wait()
# Show example scalings
low_x = 0.1
high_x = 0.9
lines = VGroup()
for x in [low_x, high_x]:
lines.add(Line(axes.c2p(x, 0), axes.c2p(x, 1)))
lines.set_stroke(YELLOW, 3)
for x, line in zip([low_x, high_x], lines):
self.play(FadeIn(line))
self.play(line.scale, x, {"about_edge": DOWN})
self.wait()
self.play(FadeOut(lines))
# Renormalize
self.play(
FadeOut(likelihood_x_label),
ReplacementTransform(likelihood_rect, p_data_rect),
)
self.wait()
one = func_label[1]
two = TexMobject("2")
two.move_to(one, LEFT)
self.play(
FadeOut(region),
sub_region.stretch, 2, 1, {"about_edge": DOWN},
sub_region.set_color, BLUE,
graph.stretch, 2, 1, {"about_edge": DOWN},
FadeInFromDown(two),
FadeOutAndShift(one, UP),
)
region = sub_region
func_label = VGroup(func_label[0], two, times_x)
self.add(func_label)
self.play(func_label.shift, 0.5 * UP)
self.wait()
const = TexMobject("C")
const.scale(0.9)
const.move_to(two, DR)
const.shift(0.07 * RIGHT)
self.play(
FadeOutAndShift(two, UP),
FadeInFrom(const, DOWN)
)
self.remove(func_label)
func_label = VGroup(func_label[0], const, times_x)
self.add(func_label)
self.play(FadeOut(p_data_rect))
self.wait()
# Show tails
coin = coins[1]
self.play(
arrow.next_to, coin, DOWN, SMALL_BUFF,
MaintainPositionRelativeTo(data_label, arrow),
FadeInFromDown(coin),
)
self.wait()
to_prior_arrow = Arrow(
func_label[0][3],
bayes[6],
max_tip_length_to_length_ratio=0.15,
stroke_width=3,
)
to_prior_arrow.set_color(RED)
self.play(Indicate(func_label, scale_factor=1.2, color=RED))
self.play(ShowCreation(to_prior_arrow))
self.wait()
self.play(FadeOut(to_prior_arrow))
# Scale by (1 - x)
eq_1mx = TexMobject("(1 - x)")
dot = TexMobject("\\cdot")
rhs_part = VGroup(dot, eq_1mx)
rhs_part.arrange(RIGHT, buff=0.2)
rhs_part.move_to(func_label, RIGHT)
l_1mx = eq_1mx.copy()
likelihood_rect.restore()
l_1mx.next_to(likelihood_rect, UP, SMALL_BUFF)
self.play(
ShowCreation(likelihood_rect),
FadeInFrom(l_1mx, 0.5 * DOWN),
)
self.wait()
self.play(ShowCreationThenFadeOut(Underline(p_label)))
self.play(Indicate(coins[1]))
self.wait()
self.play(
TransformFromCopy(l_1mx, eq_1mx),
FadeInFrom(dot, RIGHT),
func_label.next_to, dot, LEFT, 0.2,
)
scaled_graph = axes.get_graph(lambda x: 2 * x * (1 - x))
scaled_region = get_region_under_curve(axes, scaled_graph, 0, 1)
self.play(Transform(graph, scaled_graph))
self.play(FadeIn(scaled_region))
self.wait()
# Renormalize
self.remove(likelihood_rect)
self.play(
TransformFromCopy(likelihood_rect, p_data_rect),
FadeOut(l_1mx)
)
new_graph = get_beta_graph(axes, 1, 1)
group = VGroup(graph, scaled_region)
self.play(
group.set_height,
new_graph.get_height(), {"about_edge": DOWN, "stretch": True},
group.set_color, BLUE,
FadeOut(region),
)
region = scaled_region
self.play(FadeOut(p_data_rect))
self.wait()
self.play(ShowCreationThenFadeAround(const))
# Repeat
exp1 = Integer(1)
exp1.set_height(0.2)
exp1.move_to(func_label[2].get_corner(UR), DL)
exp1.shift(0.02 * DOWN + 0.07 * RIGHT)
exp2 = exp1.copy()
exp2.move_to(eq_1mx.get_corner(UR), DL)
exp2.shift(0.1 * RIGHT)
exp2.align_to(exp1, DOWN)
shift_vect = UP + 0.5 * LEFT
VGroup(exp1, exp2).shift(shift_vect)
self.play(
FadeInFrom(exp1, DOWN),
FadeInFrom(exp2, DOWN),
VGroup(func_label, dot, eq_1mx).shift, shift_vect,
bayes.scale, 0.5,
bayes.next_to, p_label, DOWN, LARGE_BUFF, {"aligned_edge": RIGHT},
)
nh = 1
nt = 1
for coin, is_heads in zip(coins[2:10], bool_values[2:10]):
self.play(
arrow.next_to, coin, DOWN, SMALL_BUFF,
MaintainPositionRelativeTo(data_label, arrow),
FadeInFrom(coin, DOWN),
)
if is_heads:
nh += 1
old_exp = exp1
else:
nt += 1
old_exp = exp2
new_exp = old_exp.copy()
new_exp.increment_value(1)
dist = scipy.stats.beta(nh + 1, nt + 1)
new_graph = axes.get_graph(dist.pdf)
new_region = get_region_under_curve(axes, new_graph, 0, 1)
new_region.match_style(region)
self.play(
FadeOut(graph),
FadeOut(region),
FadeIn(new_graph),
FadeIn(new_region),
FadeOutAndShift(old_exp, MED_SMALL_BUFF * UP),
FadeInFrom(new_exp, MED_SMALL_BUFF * DOWN),
)
graph = new_graph
region = new_region
self.remove(new_exp)
self.add(old_exp)
old_exp.increment_value()
self.wait()
if coin is coins[4]:
area_label = TextMobject("Area = 1")
area_label.move_to(axes.c2p(0.6, 0.8))
self.play(GrowFromPoint(
area_label, const.get_center()
))
class PDefectEqualsQmark(Scene):
def construct(self):
label = TexMobject(
"P(\\text{Defect}) = ???",
tex_to_color_map={
"\\text{Defect}": RED,
}
)
self.play(FadeInFrom(label, DOWN))
self.wait()
class UpdateOnceWithBinomial(TalkThroughCoinExample):
def construct(self):
# Fair bit of copy-pasting from above. If there's
# time, refactor this properly
# Setup
axes = self.get_axes()
x_label = TexMobject("x")
x_label.next_to(axes.x_axis.get_end(), UR, MED_SMALL_BUFF)
axes.add(x_label)
p_label, prob, prob_box = self.get_probability_label()
prob_box_x = x_label.copy().move_to(prob_box)
q_marks = prob_box[1]
prob_box.remove(q_marks)
prob_box.add(prob_box_x)
self.add(axes)
self.add(p_label)
self.add(prob_box)
# Coins
bool_values = (np.random.random(100) < self.true_p)
bool_values[:5] = [True, False, True, True, False]
coins = self.get_coins(bool_values)
coins.next_to(axes.y_axis, RIGHT, MED_LARGE_BUFF)
coins.to_edge(UP)
self.add(coins[:10])
# Uniform pdf
region = get_beta_graph(axes, 0, 0)
graph = axes.get_graph(
lambda x: 1,
min_samples=30,
)
self.add(region, graph)
# Show Bayes rule
bayes = TexMobject(
"p({x} | \\text{data})", "=",
"p({x})",
"{P(\\text{data} | {x})",
"\\over",
"P(\\text{data})",
tex_to_color_map={
"{x}": WHITE,
"\\text{data}": GREEN,
}
)
bayes.move_to(axes.c2p(0, 2.5))
bayes.align_to(coins, LEFT)
likelihood = bayes[9:14]
# likelihood_rect = SurroundingRectangle(likelihood, buff=0.05)
self.add(bayes)
# All data at once
brace = Brace(coins[:10], DOWN)
all_data_label = brace.get_text("One update from all data")
self.wait()
self.play(
GrowFromCenter(brace),
FadeInFrom(all_data_label, 0.2 * UP),
)
self.wait()
# Binomial formula
nh = sum(bool_values[:10])
nt = sum(~bool_values[:10])
likelihood_brace = Brace(likelihood, UP)
t2c = {
str(nh): BLUE,
str(nt): RED,
}
binom_formula = TexMobject(
"{10 \\choose ", str(nh), "}",
"x^{", str(nh), "}",
"(1-x)^{" + str(nt) + "}",
tex_to_color_map=t2c,
)
binom_formula[0][-1].set_color(BLUE)
binom_formula[1].set_color(WHITE)
binom_formula.set_width(likelihood_brace.get_width() + 0.5)
binom_formula.next_to(likelihood_brace, UP)
self.play(
TransformFromCopy(brace, likelihood_brace),
FadeOut(all_data_label),
FadeIn(binom_formula)
)
self.wait()
# New plot
rhs = TexMobject(
"C \\cdot",
"x^{", str(nh), "}",
"(1-x)^{", str(nt), "}",
tex_to_color_map=t2c
)
rhs.next_to(bayes[:5], DOWN, LARGE_BUFF, aligned_edge=LEFT)
eq = TexMobject("=")
eq.rotate(90 * DEGREES)
eq.next_to(bayes[:5], DOWN, buff=0.35)
dist = scipy.stats.beta(nh + 1, nt + 1)
new_graph = axes.get_graph(dist.pdf)
new_graph.shift(1e-6 * UP)
new_graph.set_stroke(WHITE, 1, opacity=0.5)
new_region = get_region_under_curve(axes, new_graph, 0, 1)
new_region.match_style(region)
new_region.set_opacity(0.75)
self.add(new_region, new_graph, bayes)
region.unlock_triangulation()
self.play(
FadeOut(graph),
FadeOut(region),
FadeIn(new_graph),
FadeIn(new_region),
run_time=1,
)
self.play(
Write(eq),
FadeInFrom(rhs, UP)
)
self.wait()