def loss_l2_penalization_on_grad(H,
aest,
yij,
varyj,
nj,
j1,
j2,
mask,
l=1,
wtype=1):
grad = o.grad_for_masked_data(H, mask)
return f(H, aest, yij, varyj, nj, j1, j2, wtype) + l * np.sum(grad**2)
def grad_loss_l2_penalization_on_grad(H,
aest,
yij,
varyj,
nj,
j1,
j2,
mask,
l=1,
wtype=1):
grad = o.grad_for_masked_data(H, mask)
div = o.div(grad)[mask]
return gradf(H, aest, yij, varyj, nj, j1, j2, wtype) - 2 * l * div
def welch_squared_loss_l2pen(H,
aest,
log2frq,
log2pwr,
mask,
l=1,
consider_fBm=False):
loss, grad_loss = welch_squared_loss(H,
aest,
log2frq,
log2pwr,
compute_energy=True,
compute_grad=True,
consider_fBm=consider_fBm)
grad = o.grad_for_masked_data(H, mask)
div = o.div(grad)[mask]
return loss + l * np.dot(grad.ravel(),
grad.ravel()), grad_loss - 2 * l * div
def welch_squared_loss_l2pen(H, aest,log2frq, log2pwr, mask, l=1, consider_fBm=False):
loss, grad_loss = welch_squared_loss(H, aest, log2frq, log2pwr, compute_energy=True, compute_grad=True, consider_fBm=consider_fBm)
grad = o.grad_for_masked_data(H,mask)
div = o.div(grad)[mask]
return loss + l * np.dot(grad.ravel(), grad.ravel()), grad_loss - 2 * l * div
def grad_loss_l2_penalization_on_grad(H, aest, yij, varyj, nj, j1, j2, mask, l=1, wtype=1):
grad = o.grad_for_masked_data(H,mask)
div = o.div(grad)[mask]
return gradf(H, aest, yij, varyj, nj, j1, j2,
wtype) - 2 * l * div
def loss_l2_penalization_on_grad(H, aest, yij, varyj, nj, j1, j2, mask, l=1, wtype=1):
grad = o.grad_for_masked_data(H,mask)
return f(H, aest, yij, varyj, nj, j1, j2, wtype) + l * np.sum(grad**2)
