def Run(self, params):
# Get filenames for inputs.
deps = self.Dependencies(params)
fig = plt.figure(figsize=[20, 10])
for dep in (dep for dep in deps if dep.task == 'sample'):
dep_outputs = jobs.Outputs(dep)
assert len(dep_outputs) == 1
dep_output = dep_outputs[0]
with h5py.File(dep_output) as hf:
samples = np.array(hf['/samples'])
min_x = np.min(samples)
max_x = np.max(samples)
plt.hist(samples, bins=100, density=True)
for dep in (dep for dep in deps if dep.task == 'approximation'):
dep_outputs = jobs.Outputs(dep)
assert len(dep_outputs) == 1
dep_output = dep_outputs[0]
with h5py.File(dep_output) as hf:
mode = np.array(hf['/mode'])
covariance = np.array(hf['/covariance'])
xs = np.linspace(min_x, max_x, num=100, endpoint=True)
plt.plot(xs,
scipy.stats.norm.pdf(xs, loc=mode[0], scale=np.sqrt(covariance[0, 0])),
label='approximation')
plt.legend()
plt.savefig(self.Outputs(params)[0],
bbox_inches='tight',
pad_inches=0)
def Run(self, params):
# Get filenames for inputs.
deps = self.Dependencies(params)
fig = plt.figure(figsize=[20, 10])
min_x_all = np.inf
max_x_all = -np.inf
for dep in (dep for dep in deps if dep.task == 'sample'):
dep_outputs = jobs.Outputs(dep)
assert len(dep_outputs) == 1
dep_output = dep_outputs[0]
with h5py.File(dep_output) as hf:
samples = np.array(hf['/samples'])
print(dep.params)
print(len(samples))
print(samples)
label = nice_alg_name(dep.params)
label += str(dep.params['alg']['params'])
if 'proposal' in dep.params:
label += str(dep.params['proposal'])
kde = scipy.stats.gaussian_kde(samples[len(samples) // 2:])
min_x = np.min(samples)
max_x = np.max(samples)
xs = np.linspace(min_x, max_x, num=100, endpoint=True)
plt.plot(xs, kde(xs), label=label)
min_x_all = min(min_x_all, min_x)
max_x_all = max(max_x_all, max_x)
for dep in (dep for dep in deps if dep.task == 'approximation'):
dep_outputs = jobs.Outputs(dep)
assert len(dep_outputs) == 1
dep_output = dep_outputs[0]
with h5py.File(dep_output) as hf:
mode = np.array(hf['/mode'])
covariance = np.array(hf['/covariance'])
xs = np.linspace(min_x_all, max_x_all, num=100, endpoint=True)
plt.plot(xs,
scipy.stats.norm.pdf(xs, loc=mode[0], scale=np.sqrt(covariance[0, 0])),
'--',
linewidth=5,
alpha=0.5,
label='approximation')
plt.legend()
plt.savefig(self.Outputs(params)[0],
bbox_inches='tight',
pad_inches=0)
def Run(self, params):
# Get filename for samples.
deps = self.Dependencies(params)
dep_outputs = [jobs.Outputs(j)[0] for j in deps]
datas = []
for output in dep_outputs:
with h5py.File(output) as hf:
data = {
'iact_iterations': np.array(hf['/iact_iterations'])[0],
'iact_likelihood_evaluations': np.array(hf['/iact_likelihood_evaluations'])[0],
'iact_seconds': np.array(hf['/iact_seconds'])[0],
}
datas.append(data)
pickle.dump(
{
'data': datas,
'params': params,
},
open(self.Outputs(params)[0], 'wb')
)
def Run(self, params):
print('running AcceptancePcn: %s' % jobs.params_to_string(params))
# Get filenames for inputs.
deps = self.Dependencies(params)
df = pandas.DataFrame(columns=[
'alg',
'N',
'rho',
'iact_seconds',
])
for dep in deps:
dep_outputs = jobs.Outputs(dep)
assert len(dep_outputs) == 1
dep_output = dep_outputs[0]
with h5py.File(dep_output) as hf:
samples = np.array(hf['/samples'])
acceptances = sum(samples[1:] != samples[:-1])
rate = acceptances / (len(samples) - 1)
label = nice_alg_name(dep.params)
df = df.append({
'alg': label,
'N': dep.params['dataset']['N'],
'rho': dep.params['proposal']['params']['rho'],
'acceptance_rate': rate,
}, ignore_index=True)
print(df)
hue_order = [
'MH',
'SMH-1',
'SMH-2',
'FlyMC',
]
plt.figure(figsize=[6.4, 3.2])
fig, subplots = plt.subplots(1, len(self.rhos))
for i in range(len(self.rhos)):
ax = subplots[i]
rho = self.rhos[i]
sns.pointplot(
ax=ax,
x='N',
y='acceptance_rate',
data=df[df['rho'] == rho],
hue='alg',
hue_order=hue_order,
markers=[markers[k] for k in hue_order],
linestyles=[linestyles[k] for k in hue_order],
palette=palette,
)
current_axes = ax.axis()
ax.axis([current_axes[0], current_axes[1], 0.0, 1.0])
if i == 0:
legend = ax.legend()
legend.set_title('')
else:
ax.get_legend().remove()
ax.set(ylabel='')
ax.yaxis.set_ticks([])
ax.yaxis.set_ticklabels([])
ticklabels = ax.xaxis.get_ticklabels()
num_ticks = len(ticklabels)
ax.xaxis.set_ticks([0, num_ticks-1])
ax.xaxis.set_ticklabels([ticklabels[0], ticklabels[-1]])
for i, l in enumerate(ax.xaxis.get_ticklabels()):
if i != 0 and i != len(ax.xaxis.get_ticklabels()) - 1:
l.set_visible(False)
ax.set(xlabel=r'$n$')
ax.set(ylabel=i == 0 and 'Acceptance Rate' or '')
ax.set_title(r'$\rho=%0.1f$' % rho)
plt.savefig(self.Outputs(params)[0],
bbox_inches='tight',
pad_inches=0)
def Run(self, params):
print('running EssPcn: %s' % jobs.params_to_string(params))
# Get filenames for inputs.
deps = self.Dependencies(params)
df = pandas.DataFrame(columns=[
'alg',
'N',
'rho',
'ess_seconds',
])
for dep in deps:
dep_outputs = jobs.Outputs(dep)
assert len(dep_outputs) == 1
dep_output = dep_outputs[0]
label = nice_alg_name(dep.params)
pickle_data = pickle.load(open(dep_output, 'rb'))
print('pickle_data:')
print(pickle_data)
# TODO rename this 'run' and restructure the pickled data
for run in pickle_data['data']:
df = df.append({
'alg': label,
'N': dep.params['dataset']['N'],
'rho': dep.params['proposal']['params']['rho'],
'ess_seconds': 1.0 / run['iact_seconds'],
}, ignore_index=True)
print(df)
hue_order = [
'MH',
'SMH-1',
'SMH-2',
'FlyMC',
]
plt.figure(figsize=[6.4, 3.2])
fig, subplots = plt.subplots(1, len(self.rhos))
for i in range(len(self.rhos)):
ax = subplots[i]
rho = self.rhos[i]
sns.pointplot(
ax=ax,
x='N',
y='ess_seconds',
data=df[df['rho'] == rho],
hue='alg',
hue_order=hue_order,
markers=[markers[k] for k in hue_order],
linestyles=[linestyles[k] for k in hue_order],
palette=palette,
)
if i == 0:
legend = ax.legend()
legend.set_title('')
ax.set(ylabel='First moment ESS per second')
else:
ax.get_legend().remove()
ax.set(ylabel='')
ax.yaxis.set_ticklabels([])
ax.yaxis.set_ticklabels([], minor=True)
ax.set_title(r'$\rho = %0.1f$' % rho)
ax.set(xlabel=r'$n$')
ax.set(yscale='log')
ticklabels = ax.xaxis.get_ticklabels()
num_ticks = len(ticklabels)
ax.xaxis.set_ticks([0, num_ticks - 1])
ax.xaxis.set_ticklabels([ticklabels[0], ticklabels[-1]])
for j, l in enumerate(ax.xaxis.get_ticklabels()):
if j != 0 and j != len(ax.xaxis.get_ticklabels()) - 1:
l.set_visible(False)
# Square up axes.
min_y = np.inf
max_y = -np.inf
for ax in subplots:
axis = ax.axis()
min_y = min(min_y, axis[2])
max_y = max(max_y, axis[3])
for ax in subplots:
axis = ax.axis()
ax.axis([axis[0], axis[1], min_y, max_y])
for i, ax in enumerate(subplots):
if i != 0:
ax.yaxis.set_ticklabels([])
ax.yaxis.set_ticklabels([], minor=True)
plt.savefig(self.Outputs(params)[0],
bbox_inches='tight',
pad_inches=0)
def Run(self, params):
print('running EssRandomWalk: %s' % jobs.params_to_string(params))
# Get filenames for inputs.
deps = self.Dependencies(params)
df = pandas.DataFrame(columns=[
'alg',
'N',
'ess_seconds',
])
for dep in deps:
dep_outputs = jobs.Outputs(dep)
assert len(dep_outputs) == 1
dep_output = dep_outputs[0]
label = nice_alg_name(dep.params)
pickle_data = pickle.load(open(dep_output, 'rb'))
# TODO rename this 'run' and restructure the pickled data
for run in pickle_data['data']:
df = df.append({
'alg': label,
'N': dep.params['dataset']['N'],
'ess_seconds': 1.0 / run['iact_seconds'],
}, ignore_index=True)
print(df)
hue_order = [
'MH',
'SMH-1',
'SMH-2',
'FlyMC',
]
if params['model_type'] == 'lr':
hue_order.append('Zig-Zag')
plt.figure(figsize=[6.4, 3.2])
ax = sns.pointplot(
x='N',
y='ess_seconds',
data=df,
hue='alg',
hue_order=hue_order,
markers=[markers[k] for k in hue_order],
linestyles=[linestyles[k] for k in hue_order],
palette=palette,
)
ax.set(ylabel=r'Effective sample size per second')
legend = ax.legend()
legend.set_title('')
ax.set(xlabel=r'$n$')
ax.set(yscale='log')
min_n = df['N'].min()
max_n = df['N'].max()
new_ticklabels = ax.xaxis.get_ticklabels()
new_ticklabels[-2] = ''
new_ticklabels[-4] = ''
ax.xaxis.set_ticklabels(new_ticklabels)
plt.savefig(self.Outputs(params)[0],
bbox_inches='tight',
pad_inches=0)
def Run(self, params):
print('running LikelihoodsPerIterationPlot: %s' % jobs.params_to_string(params))
# Get filenames for inputs.
deps = self.Dependencies(params)
df = pandas.DataFrame(columns=[
'alg',
'N',
'likelihood_evaluations',
'iterations',
])
for dep in deps:
dep_outputs = jobs.Outputs(dep)
assert len(dep_outputs) == 1
dep_output = dep_outputs[0]
label = nice_alg_name(dep.params)
with h5py.File(dep_output) as hf:
df = df.append({
'alg': label,
'N': dep.params['dataset']['N'],
'iterations': np.array(hf['/samples']).shape[0],
'likelihood_evaluations': np.array(hf['/likelihood_evaluations'])[0],
}, ignore_index=True)
df['likelihoods_per_iteration'] = df['likelihood_evaluations'] / df['iterations']
print(df)
hue_order = [
'MH',
'SMH-1',
'SMH-2',
]
poster = False
if poster:
context = {
'text.usetex': True,
'text.latex.preamble': '\\RequirePackage{cmbright}\n\\RequirePackage[default]{lato}',
'axes.facecolor': 'ffffff',
'axes.edgecolor': '000000',
'axes.labelcolor': 'ffffff',
'xtick.color': 'ffffff',
'ytick.color': 'ffffff',
'legend.edgecolor': '000000',
'legend.facecolor': 'ffffff',
'text.color': '000000',
'savefig.facecolor': '002147',
}
else:
context = {}
with plt.rc_context(context):
fig = plt.figure(figsize=[6.4, 3.2])
ax = sns.pointplot(
x='N',
y='likelihoods_per_iteration',
data=df,
hue='alg',
hue_order=hue_order,
markers=[markers[k] for k in hue_order],
linestyles=[linestyles[k] for k in hue_order],
palette=palette,
)
ax.set(ylabel=r'Likelihood evaluations per iteration')
legend = ax.legend()
legend.set_title('')
ax.set(xlabel=r'$n$')
ax.set(yscale='log')
min_n = df['N'].min()
max_n = df['N'].max()
new_ticklabels = ax.xaxis.get_ticklabels()
new_ticklabels[-2] = ''
new_ticklabels[-4] = ''
ax.xaxis.set_ticklabels(new_ticklabels)
plt.savefig(self.Outputs(params)[0],
bbox_inches='tight',
pad_inches=0)