def _one_step_lr(self):
"""Computes the best learning rate for G-Shapley algorithm."""
if self.directory is None:
address = None
else:
address = os.path.join(self.directory, 'weights')
best_acc = 0.0
for i in np.arange(1, 5, 0.5):
model = ShapNN(self.problem,
batch_size=1,
max_epochs=1,
learning_rate=10**(-i),
weight_decay=0.,
validation_fraction=0,
optimizer='sgd',
warm_start=False,
address=address,
hidden_units=self.hidden_units)
accs = []
for _ in range(10):
model.fit(np.zeros((0, self.X.shape[-1])), self.y)
model.fit(self.X, self.y)
accs.append(model.score(self.X_test, self.y_test))
if np.mean(accs) - np.std(accs) > best_acc:
best_acc = np.mean(accs) - np.std(accs)
learning_rate = 10**(-i)
return learning_rate
def _g_shap(self, iterations, err=None, learning_rate=None, sources=None):
"""Method for running G-Shapley algorithm.
Args:
iterations: Number of iterations of the algorithm.
err: Stopping error criteria
learning_rate: Learning rate used for the algorithm. If None
calculates the best learning rate.
sources: If values are for sources of data points rather than
individual points. In the format of an assignment array
or dict.
"""
if sources is None:
sources = {i:np.array([i]) for i in range(len(self.X))}
elif not isinstance(sources, dict):
sources = {i:np.where(sources==i)[0] for i in set(sources)}
address = None
if self.directory is not None:
address = os.path.join(self.directory, 'weights')
if learning_rate is None:
try:
learning_rate = self.g_shap_lr
except AttributeError:
self.g_shap_lr = self._one_step_lr()
learning_rate = self.g_shap_lr
model = ShapNN(self.problem, batch_size=1, max_epochs=1,
learning_rate=learning_rate, weight_decay=0.,
validation_fraction=0, optimizer='sgd',
address=address, hidden_units=self.hidden_units)
for iteration in range(iterations):
model.fit(np.zeros((0, self.X.shape[-1])), self.y)
if 10 * (iteration+1) / iterations % 1 == 0:
print('{} out of {} G-Shapley iterations'.format(
iteration + 1, iterations))
marginal_contribs = np.zeros(len(sources.keys()))
model.fit(self.X, self.y, self.X_test, self.y_test,
sources=sources, metric=self.metric,
max_epochs=1, batch_size=1)
val_result = model.history['metrics']
marginal_contribs[1:] += val_result[0][1:]
marginal_contribs[1:] -= val_result[0][:-1]
individual_contribs = np.zeros(len(self.X))
for i, index in enumerate(model.history['idxs'][0]):
individual_contribs[sources[index]] += marginal_contribs[i]
individual_contribs[sources[index]] /= len(sources[index])
self.mem_g = np.concatenate(
[self.mem_g, np.reshape(individual_contribs, (1,-1))])
self.idxs_g = np.concatenate(
[self.idxs_g, np.reshape(model.history['idxs'][0], (1,-1))])
