def plot_simulated_histogram_variety_test(parameters: np.array) -> None:
params = {"review_prior": np.ones(5), "tendency_to_rate": 0.05}
simulator = simulator_class.DoubleRhoSimulator(params)
simulator.simulation_parameters = {"rho": parameters}
with tqdm_joblib(tqdm(desc="Simulations", total=parameters.shape[0])) as progress_bar:
simulations = Parallel(n_jobs=mp.cpu_count())(
delayed(simulator.simulate_review_histogram)(i) for i in range(parameters.shape[0])
)
simulations = np.array(simulations)
for i in range(parameters.shape[0]):
plt.figure()
plt.bar(
[1, 2, 3, 4, 5],
simulations[i, :],
width=0.5,
color=sns.xkcd_rgb["grey"],
label=r"$\rho_{-}=$" + f"{parameters[i, 0]}, " + r"$\rho_{+}=$" + f"{parameters[i, 1]}",
)
ax = plt.gca()
ax.spines["top"].set_visible(False)
ax.spines["right"].set_visible(False)
plt.xlabel("Rating")
plt.ylabel("Number of ratings")
plt.legend(loc="upper left", fontsize=20)
def generate_and_save_simulations(
num_simulations: int, review_prior: np.array, tendency_to_rate: float, simulation_type: str
) -> None:
params = {"review_prior": review_prior, "tendency_to_rate": tendency_to_rate, "simulation_type": simulation_type}
simulator = simulator_class.DoubleRhoSimulator(params)
simulator.simulate(num_simulations=num_simulations)
simulator.save_simulations(ARTIFACT_PATH)
def plot_simulated_vs_actual_histogram_test(
observed_histograms: np.array,
posterior_samples: np.array,
products_to_test: np.array,
plot_histograms: bool = False,
return_raw_simulations: bool = True,
) -> np.ndarray:
print(posterior_samples.shape)
products_to_test = products_to_test.astype("int")
simulated_histograms = np.zeros((posterior_samples.shape[0], len(products_to_test), 5))
# Get the total number of reviews of the products we want to test
# We will simulate as many reviews for each products as exist in their observed histograms
# total_reviews = np.sum(observed_histograms[products_to_test, :], axis=1)
params = {"review_prior": np.ones(5), "tendency_to_rate": 0.05, "simulation_type": "histogram"}
simulator = simulator_class.DoubleRhoSimulator(params)
# Take posterior samples of the products we want to test
# We will simulate distributions using these posterior samples as parameters
parameters = np.swapaxes(posterior_samples[:, products_to_test, :], 0, 1).reshape((-1, 2))
# We need to expand total reviews to be same number as the number of simulations to be run
# total_reviews = np.tile(total_reviews[:, None], (1, posterior_samples.shape[0])).flatten()
simulator.simulation_parameters = {"rho": parameters}
with tqdm_joblib(tqdm(desc="Simulations", total=parameters.shape[0])) as progress_bar:
simulations = Parallel(n_jobs=mp.cpu_count())(
delayed(simulator.simulate_review_histogram)(i) for i in range(parameters.shape[0])
)
simulations = np.array(simulations)
simulated_histograms[:, :, :] = simulations.reshape((-1, len(products_to_test), 5), order="F")
simulated_histograms /= np.sum(simulated_histograms, axis=-1)[:, :, None]
if plot_histograms:
for i in range(len(products_to_test)):
plt.figure()
plt.plot(
np.arange(5) + 1,
observed_histograms[i, :] / np.sum(observed_histograms[i, :]),
linewidth=4.0,
color="black",
)
# Get the HPDs of the simulated histograms
hpd = arviz.hdi(simulated_histograms[:, i, :], hdi_prob=0.95)
plt.fill_between(np.arange(5) + 1, hpd[:, 0], hpd[:, 1], color="black", alpha=0.4)
plt.ylim([0, 1])
if return_raw_simulations:
return simulations
else:
return simulated_histograms
def load_simulator(
self, dirname: Path, simulator_type: str = "double_rho", simulation_type: str = "timeseries"
) -> None:
# The parameters used to initialize the simulator object don't matter here
# as they will be overridden by those of the loaded simulator
params = {"review_prior": np.ones(5), "tendency_to_rate": 0.05, "simulation_type": simulation_type}
self.simulator_type = simulator_type
self.simulation_type = simulation_type
if self.simulator_type == "single_rho":
simulator = simulator_class.SingleRhoSimulator(params)
elif self.simulator_type == "double_rho":
simulator = simulator_class.DoubleRhoSimulator(params)
elif self.simulator_type == "herding":
# Add the additional parameters that the herding simulator needs to initialize
# Again the values aren't important as they will be overridden by those of the loaded simulator
params.update({"previous_rating_measure": "mean", "min_reviews_for_herding": 5})
simulator = simulator_class.HerdingSimulator(params)
else:
raise ValueError(
f"simulator_type has to be one of single_rho, double rho or herding, found {self.simulator_type}"
)
simulator.load_simulator(dirname)
self.simulator = simulator
def plot_test_parameter_recovery(
parameters: np.array,
num_posterior_samples: int,
simulator_type: str,
simulation_type: str,
plot_posteriors: bool = False,
get_stats: bool = False,
param_posterior_prob_band: Optional[float] = None,
) -> np.ndarray:
# Simulate review histograms using provided parameters
params = {"review_prior": np.ones(5), "tendency_to_rate": 0.05, "simulation_type": simulation_type}
simulator = simulator_class.DoubleRhoSimulator(params)
simulator.simulation_parameters = {"rho": parameters}
with tqdm_joblib(tqdm(desc="Simulations", total=parameters.shape[0])) as progress_bar:
simulations = Parallel(n_jobs=mp.cpu_count())(
delayed(simulator.simulate_review_histogram)(i) for i in range(parameters.shape[0])
)
simulations = np.array(simulations)
# The parameter prior doesn't matter here as it will be overridden by that of the loaded inference object
parameter_prior = sbi.utils.BoxUniform(
low=torch.tensor([0.0, 0.0]).type(torch.FloatTensor), high=torch.tensor([4.0, 4.0]).type(torch.FloatTensor)
)
inferrer = inference_class.HistogramInference(parameter_prior=parameter_prior)
inferrer.load_simulator(dirname=ARTIFACT_PATH, simulator_type=simulator_type, simulation_type=simulation_type)
inferrer.load_inference(dirname=ARTIFACT_PATH)
posterior_samples = inferrer.get_posterior_samples(simulations, num_samples=num_posterior_samples)
# Plot the posterior samples inferred for the simulated data
# We will plot upto 4 plots in one row of the panel
if plot_posteriors:
if len(parameters) <= 4:
fig, ax = plt.subplots(1, len(parameters), squeeze=False)
else:
fig, ax = plt.subplots((len(parameters) + 1) // 4, 4, squeeze=False)
row_index = 0
for i in range(len(parameters)):
if len(parameters) > 4:
row_index = i // 4
ax[row_index, i % 4].hist(
posterior_samples[:, i, 0], color="black", alpha=0.5, bins=10, label=r"$\rho_{-}$"
)
ax[row_index, i % 4].axvline(x=parameters[i, 0], linewidth=3.0, color="black", linestyle="--")
ax[row_index, i % 4].hist(posterior_samples[:, i, 1], color="red", alpha=0.5, bins=10, label=r"$\rho_{+}$")
ax[row_index, i % 4].axvline(x=parameters[i, 1], linewidth=3.0, color="red", linestyle="--")
ax[row_index, i % 4].set_xlim([0, 4])
ax[row_index, i % 4].set_xticks([0, 1, 2, 3, 4])
ax[row_index, i % 4].legend()
# add a big axis, hide frame
fig.add_subplot(111, frameon=False)
# hide tick and tick label of the big axis
plt.tick_params(labelcolor="none", top=False, bottom=False, left=False, right=False)
plt.xlabel(r"$\rho_{-}, \rho_{+}$")
plt.ylabel("Number of samples")
# If asked, print how many of the provided parameters are recovered by the inference engine
# i.e, how often do the supplied parameters lie within the 95% HPD of the posterior
if get_stats:
f = open(ARTIFACT_PATH / "stats_parameter_recovery.txt", "w")
assert (
posterior_samples.shape == (num_posterior_samples,) + parameters.shape
), f"""
Expected shape {(num_posterior_samples,) + parameters.shape} for array of posterior samples,
but got {posterior_samples.shape} instead
"""
# First get the HPD of each recovered posterior distribution
hpd = np.array([arviz.hdi(posterior_samples[:, i, :], hdi_prob=0.95) for i in range(parameters.shape[0])])
assert hpd.shape == parameters.shape + (2,), f"Found shape {hpd.shape} for hpd"
# See how many of the supplied rho_- and rho_+ are contained in these HPDs
contained_rho_0 = [
True if (parameters[i, 0] < hpd[i, 0, 1] and parameters[i, 0] > hpd[i, 0, 0]) else False
for i in range(parameters.shape[0])
]
contained_rho_1 = [
True if (parameters[i, 1] < hpd[i, 1, 1] and parameters[i, 1] > hpd[i, 1, 0]) else False
for i in range(parameters.shape[0])
]
print(
f"""
rho- is recovered {np.sum(contained_rho_0)} times out of {parameters.shape[0]}
= {100*(np.sum(contained_rho_0) / parameters.shape[0]):0.2f}%"
""",
file=f,
)
print(
f"""
rho+ is recovered {np.sum(contained_rho_1)} times out of {parameters.shape[0]}
= {100*(np.sum(contained_rho_1) / parameters.shape[0]):0.2f}%"
""",
file=f,
)
print("=======================================================", file=f)
# Now get the probability that the posterior distribution puts in a band/region around
# the passed parameter values. For good parameter recovery, this number should be high
assert (
param_posterior_prob_band is not None
), f"""
Posterior probability band around parameter values need to be passed if stats are needed
"""
param_band_low = parameters - param_posterior_prob_band
param_band_high = parameters + param_posterior_prob_band
rho_0_probs = (posterior_samples[:, :, 0] >= param_band_low[None, :, 0]) * (
posterior_samples[:, :, 0] <= param_band_high[None, :, 0]
)
rho_0_probs = np.mean(rho_0_probs, axis=0)
rho_1_probs = (posterior_samples[:, :, 1] >= param_band_low[None, :, 1]) * (
posterior_samples[:, :, 1] <= param_band_high[None, :, 1]
)
rho_1_probs = np.mean(rho_1_probs, axis=0)
print(
f"""
In {100*np.mean(rho_0_probs>=0.5):0.2f}% of cases, the inferred posterior places more than 50% probability
in a band of {2*param_posterior_prob_band} around the true value of rho-
""",
file=f,
)
print(
f"""
In {100*np.mean(rho_1_probs>=0.5):0.2f}% of cases, the inferred posterior places more than 50% probability
in a band of {2*param_posterior_prob_band} around the true value of rho+
""",
file=f,
)
f.close()
# Finally, plot the distribution of the posterior probability the inference engine places in a
# band around the true value of rho- and rho+
plt.figure()
plt.hist(rho_0_probs, alpha=0.5, label=r"$\rho_{-}$")
plt.hist(rho_1_probs, alpha=0.5, label=r"$\rho_{+}$")
plt.legend()
plt.title(
f"Posterior probability placed by inference engine in a band of {2*param_posterior_prob_band}"
+ f"\n around the true value of the parameters ({parameters.shape[0]} trials)",
fontsize=24.0,
)
return posterior_samples