def measure():
def noisy_mapping(mapping):
def _(x):
y = mapping(x)
y += np.random.normal(0, 2, size=y.size)
return y
return _
funcs = {}
funcs['sin'] = noisy_mapping(np.sin)
funcs['log'] = noisy_mapping(np.log)
funcs['pow1.5'] = noisy_mapping(lambda x: x**(3 / 2))
X = np.arange(0.1, 10, step=0.001)
import pennpaper as pp
from collections import defaultdict
metrics = defaultdict(list)
for i in range(90):
for name, f in funcs.items():
m = pp.Metric(name)
m.add_arrays(X, f(X))
metrics[m.name].append(m)
metrics = [sum(v) for v in metrics.values()]
pp.plot_group(metrics)
pp.plot_group(metrics, smoothen=True)
def test_all_empty_no_plot(tmpdir):
files_before = len(os.listdir(tmpdir))
m1 = Metric('empty')
m2 = Metric('empty2')
plot_group([m1, m2], folder=tmpdir)
assert len(os.listdir(tmpdir)) == files_before
def test_some_empty_plots(tmpdir):
files_before = len(os.listdir(tmpdir))
m1 = Metric('empty')
m2 = Metric('empty2')
m2.add_record(1, 1)
plot_group([m1, m2], folder=tmpdir)
assert len(os.listdir(tmpdir)) > files_before
def many_plots():
def one_series(shift):
metrics = [Metric('x', 'y') for x in range(10)]
for m in metrics:
for i in range(40000):
m.add_record(i,
random.random() + shift + 0.4 * math.sin(i / 200))
m._sort()
return sum(metrics)
many = {f"random line {i}": one_series(i) for i in range(5)}
plot_group(many, 'temp', name="many_series")
def test_plot_group(tmpdir):
m = Metric('x', 'y')
m.add_ys(1, [1, 2, 3])
m.add_record(2, 4)
m.add_record(2, 4)
m.add_record(2, 4)
m.add_ys(4, [5, 6, 12])
m2 = Metric('x', 'y')
m2.add_ys(1, [4, 3, 5])
m2.add_record(2, 5)
m2.add_record(2, 9)
m2.add_record(2, 9)
m2.add_ys(4, [5, 6, -1])
files_before = len(os.listdir(tmpdir))
plot_group([m, m2], folder=tmpdir, name="bla3")
files_after = len(os.listdir(tmpdir))
assert files_after > files_before
pow2 = noisy_mapping(lambda x: x ** 2)
"""
lets record the pairs (x, f(x)) in a metric and make a plot:
"""
from pennpaper import Metric, plot_group, plot
m1 = Metric("pow2")
for x in X:
m1.add_record(x, pow2(x))
plot(m1)
m2 = Metric("pow2_2")
for x in X:
m2.add_record(x, pow2(x))
plot_group([m1, m2])
# Actually, m1 and m2 are metrics of the same process.
# What if we create a new metric tracking the mean and stddev of this process?
m3 = m1 + m2
plot(m3)
# the plot is too noisy to understand. We can smoothen it!
plot(m3, smoothen=True)
obs, rew, done, _ = env.step(action)
total_reward += rew
history.append((obs, action, rew))
return history, total_reward
def apply_grad(G, a, agent, lr):
prob_1 = sigmoid(agent.param)
prob = prob_1 if a == 1 else 1 - prob_1
update = lr * G * agent.grad(a) / (prob + 1e-10)
agent.param += update
m1 = sum(reinforce(2**-13) for _ in range(100))
m2 = sum(reinforce(2**-12) for _ in range(100))
m3 = sum(reinforce(2**-14) for _ in range(100))
m4 = sum(reinforce(2**-15) for _ in range(100))
m5 = sum(reinforce(2**-11) for _ in range(100))
# m6 = sum( reinforce(2 ** -10) for _ in range(100) )
plot_group([
m1,
m2,
m3,
m4,
m5,
],
smoothen=True,
stdev_factor=0.3,
name="reinforce_coridor")
import numpy as np
from pennpaper import Metric, plot_group
xs = np.arange(0.1, 5, step=0.01)
uni_noise = lambda x: np.random.uniform(size=x.shape) + x
funcs = {}
funcs['uniform'] = lambda x: np.random.uniform(size=x.shape) + x
funcs['weibull'] = lambda x: np.random.weibull(a=1, size=x.shape) + x
funcs['beta'] = lambda x: np.random.beta(a=1, b=1, size=x.shape) + x
metrics = []
for name, f in funcs.items():
m = Metric(name=name)
for i in range(100):
m.add_arrays(uni_noise(xs), f(xs), new_sample=True)
metrics.append(m)
plot_group(metrics)
plot_group(metrics, name='true', smoothen=False)
