def main(verbose=False,
beta=0.05,
prior_probability=0.5,
epsilon=0.05,
num_features=8,
feature_cardinality=5,
num_examples=100,
visualization_interval=100,
biased_feature_proportion=0.2,
biased_feature_effect_length=10**100,
directory="/tmp/adpredictor/",
extension="png"):
# Initialize globals
np.random.seed(1)
logging.basicConfig(level=logging.DEBUG if verbose else logging.INFO)
# Construct settings
simulation = SimulationRunner.Simulation(
predictor_config=AdPredictor.Config(
beta=beta,
prior_probability=prior_probability,
epsilon=epsilon,
num_features=num_features),
feature_cardinality=feature_cardinality,
num_examples=num_examples,
directory=directory,
biased_feature_proportion=biased_feature_proportion,
biased_feature_effect_length=biased_feature_effect_length,
visualization_interval=visualization_interval,
extension=extension)
# Train and output graphs
simulation_runner = SimulationRunner(simulation)
simulation_runner.run()
def libsvm_file_process(config):
adpredictor = AdPredictor(AdPredictor.Config(config.beta, config.epsilon), config.feature_num)
train_pred, train_label = [], []
for train_file in config.train_files:
for features, label in FileSampler(train_file).generate_samples():
prob = adpredictor.train(features, label, True)
train_pred.append(prob)
train_label.append(1 if label > 0 else 0)
adpredictor.save_model(config.model_file)
test_pred, test_label = [], []
for test_file in config.test_files:
for features, label in FileSampler(test_file).generate_samples():
prob = adpredictor.predict(features)
test_pred.append(prob)
test_label.append(1 if label > 0 else 0)
return train_pred, train_label, test_pred, test_label, config
def libsvm_file_process(config):
adpredictor = AdPredictor(AdPredictor.Config(config.beta, config.epsilon),
config.feature_num)
train_pred, train_label = [], []
for train_file in config.train_files:
for features, label in FileSampler(train_file).generate_samples():
prob = adpredictor.train(features, label, True)
train_pred.append(prob)
train_label.append(1 if label > 0 else 0)
adpredictor.save_model(config.model_file)
test_pred, test_label = [], []
for test_file in config.test_files:
for features, label in FileSampler(test_file).generate_samples():
prob = adpredictor.predict(features)
test_pred.append(prob)
test_label.append(1 if label > 0 else 0)
return train_pred, train_label, test_pred, test_label, config
def __init__(self, simulation):
self._simulation = simulation
self._predictor = AdPredictor(simulation.predictor_config)
self._sampler = Sampler(simulation)
class SimulationRunner(object):
Simulation = namedtuple('Simulation', [
'predictor_config', 'feature_cardinality', 'num_examples',
'biased_feature_proportion', 'directory', 'extension',
'visualization_interval', 'biased_feature_effect_length'
])
COLORS = brewer2mpl.get_map('Set2', 'qualitative', 8).mpl_colors
def __init__(self, simulation):
self._simulation = simulation
self._predictor = AdPredictor(simulation.predictor_config)
self._sampler = Sampler(simulation)
def _current_weights_by_feature(self):
by_feature = lambda kv: kv[0].feature
by_feature_value = lambda kv: (kv[0].feature, kv[0].value)
weights = sorted(self._predictor.weights, key=by_feature_value)
for feature, group in itertools.groupby(weights, key=by_feature):
yield feature, [(f, w.mean, w.variance) for (f, w) in group]
def _plot_weights(self):
for color, (feature, weights) in itertools.izip(
itertools.cycle(self.COLORS),
self._current_weights_by_feature()):
_, means, variances = zip(*weights)
logging.debug("Feature %s, Weights: %s", feature, weights)
label = "F{}".format(feature) if feature != 0 else "Bias"
plt.scatter(means,
variances,
label=label,
color=color,
alpha=0.8,
s=40)
def _annotate_biased_weights(self):
for _, weights in self._current_weights_by_feature():
for (feature, mean, variance) in weights:
bias_weight = self._sampler.get_bias_for_feature(feature)
if bias_weight is not None:
plt.annotate('+' if bias_weight else '-', (mean, variance),
size=40)
def _visualize(self, num_examples):
plt.clf()
self._plot_weights()
self._annotate_biased_weights()
plt.title(u"(μ, σ²) after {} examples".format(num_examples))
plt.xlabel(u"μ")
plt.ylabel(u"σ²")
plt.legend(loc='best')
plt.xlim(-4, 4)
plt.ylim(-0.1, 1.1)
filename = "{:03d}.{}".format(num_examples, self._simulation.extension)
logger.info("Saving graph to %s", filename)
plt.savefig(os.path.join(self._simulation.directory, filename),
dpi=300)
def run(self):
samples = itertools.islice(self._sampler,
self._simulation.num_examples)
for iteration, (features, label) in enumerate(samples):
self._predictor.train(features, label)
if iteration % self._simulation.visualization_interval == 0:
self._visualize(iteration)
def _create_predictor(beta=0.05, prior=0.3, epsilon=0.01, num_features=10):
config = AdPredictor.Config(beta, prior, epsilon, num_features)
return AdPredictor(config)
def __init__(self, simulation):
self._simulation = simulation
self._predictor = AdPredictor(simulation.predictor_config)
self._sampler = Sampler(simulation)
class SimulationRunner(object):
Simulation = namedtuple(
'Simulation',
['predictor_config', 'feature_cardinality', 'num_examples',
'biased_feature_proportion', 'directory', 'extension',
'visualization_interval', 'biased_feature_effect_length'])
COLORS = brewer2mpl.get_map('Set2', 'qualitative', 8).mpl_colors
def __init__(self, simulation):
self._simulation = simulation
self._predictor = AdPredictor(simulation.predictor_config)
self._sampler = Sampler(simulation)
def _current_weights_by_feature(self):
by_feature = lambda kv: kv[0].feature
by_feature_value = lambda kv: (kv[0].feature, kv[0].value)
weights = sorted(self._predictor.weights, key=by_feature_value)
for feature, group in itertools.groupby(weights, key=by_feature):
yield feature, [(f, w.mean, w.variance) for (f, w) in group]
def _plot_weights(self):
for color, (feature, weights) in itertools.izip(
itertools.cycle(self.COLORS),
self._current_weights_by_feature()):
_, means, variances = zip(*weights)
logging.debug("Feature %s, Weights: %s", feature, weights)
label = "F{}".format(feature) if feature != 0 else "Bias"
plt.scatter(means, variances,
label=label, color=color, alpha=0.8, s=40)
def _annotate_biased_weights(self):
for _, weights in self._current_weights_by_feature():
for (feature, mean, variance) in weights:
bias_weight = self._sampler.get_bias_for_feature(feature)
if bias_weight is not None:
plt.annotate('+' if bias_weight else '-', (mean, variance),
size=40)
def _visualize(self, num_examples):
plt.clf()
self._plot_weights()
self._annotate_biased_weights()
plt.title(u"(μ, σ²) after {} examples".format(num_examples))
plt.xlabel(u"μ")
plt.ylabel(u"σ²")
plt.legend(loc='best')
plt.xlim(-4, 4)
plt.ylim(-0.1, 1.1)
filename = "{:03d}.{}".format(num_examples, self._simulation.extension)
logger.info("Saving graph to %s", filename)
plt.savefig(os.path.join(self._simulation.directory, filename),
dpi=300)
def run(self):
samples = itertools.islice(self._sampler,
self._simulation.num_examples)
for iteration, (features, label) in enumerate(samples):
self._predictor.train(features, label)
if iteration % self._simulation.visualization_interval == 0:
self._visualize(iteration)