def save_unwhitened(th, lam0, lam0_delta, parser, K_chol, data_dir, split_type):
# grab betas, hit em w/ K_chol, and save
to_save = th.copy()
betas = parser.get(to_save, 'betas')
parser.set(to_save, 'betas', np.dot(K_chol, betas.T).T)
qfb.save_basis_fit(to_save, lam0, lam0_delta, parser,
data_dir=BASIS_DIR, split_type=SPLIT_TYPE)
def minimize_chunk(fun, jac, x0, method, chunk_size=4):
""" minimize function that saves every few iterations """
num_chunks = int(MAX_LBFGS_ITER / float(chunk_size)) + 1
for chunk_i in range(num_chunks):
print "optimizing chunk %d of %d (curr_ll = %2.5g)"%(chunk_i, num_chunks, fun(x0))
res = minimize(fun = fun, jac = jac, x0 = x0, method = method,
options = {'maxiter': chunk_size, 'disp': True})
x0 = res.x
qfb.save_basis_fit(res.x, lam0, lam0_delta, parser, data_dir="")
return res
def minimize_chunk(fun, jac, x0, method, chunk_size=250):
""" minimize function that saves every few iterations """
num_chunks = int(MAX_LBFGS_ITER / float(chunk_size)) + 1
for chunk_i in range(num_chunks):
print "optimizing chunk %d of %d (curr_ll = %2.5g)"%(chunk_i, num_chunks, fun(x0))
res = minimize(fun = fun, jac = jac, x0 = x0, method = method,
options = {'maxiter': chunk_size, 'disp': True})
x0 = res.x
qfb.save_basis_fit(res.x, lam0, lam0_delta, parser, data_dir="")
return res
def save_unwhitened(th, lam0, lam0_delta, parser, K_chol, data_dir,
split_type):
# grab betas, hit em w/ K_chol, and save
to_save = th.copy()
betas = parser.get(to_save, 'betas')
parser.set(to_save, 'betas', np.dot(K_chol, betas.T).T)
qfb.save_basis_fit(to_save,
lam0,
lam0_delta,
parser,
data_dir=BASIS_DIR,
split_type=SPLIT_TYPE)
print " %d, loss = %2.4g, grad = %2.4g " % \
(i, loss_val, np.sqrt(np.dot(g,g)))
obj_vals.append(loss_val)
## train with rms_prop
def full_loss_grad(th, idx=None):
return loss_grad(th, idx) + prior_loss_grad(th, idx)
minibatch_minimize(grad=full_loss_grad,
x=th,
callback=callback,
num_epochs=10,
batch_size=24,
step_size=RMSPROP_STEP,
mass=RMSPROP_MOMENTUM)
qfb.save_basis_fit(min_x, lam0, lam0_delta, parser, data_dir="")
## tighten it up a bit
res = minimize_chunk(fun=lambda th: loss_fun(th) + prior_loss(th),
jac=lambda th: loss_grad(th) + prior_loss_grad(th),
x0=min_x,
method='L-BFGS-B')
## save result
qfb.save_basis_fit(res.x, lam0, lam0_delta, parser, data_dir="")
th_mle = res.x
ll_mle = loss_fun(th_mle) + prior_loss(th_mle)
## plot random profiles
#plot a handful of random directions
egrid = np.linspace(-1.5, 1.5, 30)
loss_val = loss_fun(x) + prior_loss(x)
if loss_val < min_val:
min_x = x
min_val = loss_val
print " %d, loss = %2.4g, grad = %2.4g " % \
(i, loss_val, np.sqrt(np.dot(g,g)))
obj_vals.append(loss_val)
## train with rms_prop
out = rms_prop(grad = lambda th: loss_grad(th) + prior_loss_grad(th),
x = th,
callback = callback,
num_iters = MAX_RMSPROP_ITER,
step_size = RMSPROP_STEP,
gamma = RMSPROP_MOMENTUM)
qfb.save_basis_fit(min_x, lam0, lam0_delta, parser, data_dir="")
## tighten it up a bit
res = minimize_chunk(fun = lambda th: loss_fun(th) + prior_loss(th),
jac = lambda th: loss_grad(th) + prior_loss_grad(th),
x0 = min_x,
method = 'L-BFGS-B')
## save result
qfb.save_basis_fit(res.x, lam0, lam0_delta, parser, data_dir="")
th_mle = res.x
ll_mle = loss_fun(th_mle) + prior_loss(th_mle)
## plot random profiles
#plot a handful of random directions
egrid = np.linspace(-1.5, 1.5, 30)