def _solve_corr(self, HH, HT):
"""Solve a linear system from correlation matrices.
"""
if self.accelerator == "GPU":
Beta = self.magma_solver.gpu_solve(HH, HT, self.alpha)
else:
Beta = cpu_solve(HH, HT, sym_pos=True)
return Beta
def _solve_corr(self, HH, HT):
"""Solve a linear system from correlation matrices.
"""
if self.accelerator == "GPU":
Beta = self.magma_solver.gpu_solve(HH, HT, self.alpha)
else:
Beta = cpu_solve(HH, HT, sym_pos=True)
return Beta
def _calculate_beta_cpu(self, num_neurons, num_batches):
# We first calculate H.transpose * H and H.transpose * Targets
# We never actually have H in memory
# Since H=hidden_layer_output and is num_samples x num_neurons
# H.transpose * H is of size num_neurons x num_neurons
HTH = np.zeros((num_neurons, num_neurons))
# Since H.transpose is num_neurons x num_samples and output matrix
# is num_samples x num_output_dimensions that gives us this shape
HTT = np.zeros((num_neurons, self.num_output_dimensions))
# Adds to the matrix diagonal
# Adding a small amount to the matrix diagonal improves numerical stability
# Huang, G.-B., Zhou, H., Ding, X., Zhang, R.: Extreme learning machine for regression and
HTH.ravel()[::num_neurons + 1] += self.alpha
# Divide the data up into the batches
# This part is extremely parallelizable
for data_batch, output_batch in zip(np.array_split(self.data, num_batches),
np.array_split(self.targets, num_batches)):
batch_hidden_layer_output = self.calculate_neuron_outputs(np.asarray(data_batch))
# H.T * H is of size num_neurons by num_neurons - this is cheap to store
# H.T * T is of size num_neurons by num_output_dimensions - also cheap to store
# We never actually store all of H at once
# This is a piece of H.T * H
HTH += np.dot(batch_hidden_layer_output.T, batch_hidden_layer_output)
# This is a piece of H.T * T
HTT += np.dot(batch_hidden_layer_output.T, output_batch)
# Solve for beta using scipy.linalg.cpu_solve
# H * Beta = Target
# H.T * H * Beta = H.T * T
# (H.T * H) * Beta = (H.T * T), solve for beta by doing (H.T * H)^-1 * (H.T * T)
beta_matrix = cpu_solve(HTH, HTT, sym_pos=True)
return HTH, HTT, beta_matrix