def run():
(train_images, train_labels),\
(tests_images, tests_labels) = load_data_subset(N_train, N_tests)
parser, pred_fun, nllfun, frac_err = make_nn_funs(layer_sizes)
N_param = len(parser.vect)
print "Running experiment..."
results = defaultdict(list)
for i in xrange(N_samples):
x_init_scale = np.full(N_param, init_scale)
def indexed_loss_fun(w, i_iter):
rs = RandomState((seed, i, i_iter))
idxs = rs.randint(N_train, size=batch_size)
nll = nllfun(w, train_images[idxs], train_labels[idxs]) * N_train
nlp = neg_log_prior(w)
return nll + nlp
gradfun = grad(indexed_loss_fun)
def callback(x, t, v, entropy):
results[("entropy", i)].append(entropy / N_train)
results[("v_norm", i)].append(norm(v) / np.sqrt(N_param))
results[("minibatch_likelihood",
i)].append(-indexed_loss_fun(x, t))
results[("log_prior_per_dpt",
i)].append(-neg_log_prior(x) / N_train)
if t % thin != 0 and t != N_iter and t != 0: return
results[('iterations', i)].append(t)
results[("train_likelihood",
i)].append(-nllfun(x, train_images, train_labels))
results[("tests_likelihood",
i)].append(-nllfun(x, tests_images, tests_labels))
results[("tests_error",
i)].append(frac_err(x, tests_images, tests_labels))
print "Iteration {0:5} Train likelihood {1:2.4f} Test likelihood {2:2.4f}" \
" Test Err {3:2.4f}".format(t, results[("train_likelihood", i)][-1],
results[("tests_likelihood", i)][-1],
results[("tests_error", i)][-1])
rs = RandomState((seed, i))
entropic_descent_deterministic(gradfun,
callback=callback,
x_scale=x_init_scale,
epsilon=epsilon,
gamma=gamma,
alpha=alpha,
annealing_schedule=annealing_schedule,
rs=rs)
return results
def run():
x_init_scale = np.full(D, init_scale)
# annealing_schedule = np.linspace(0,1,N_iter)
annealing_schedule = np.concatenate((np.zeros(N_iter / 3), np.linspace(0, 1, N_iter / 3), np.ones(N_iter / 3)))
print "Running experiment..."
results = defaultdict(list)
for i in xrange(N_samples):
def callback(x, t, v, entropy):
# results[("x", i)].append(x.copy()) # Replace this with a loop over kwargs?
results[("entropy", i)].append(entropy)
# results[("velocity", i)].append(v)
results[("likelihood", i)].append(-nllfun(x))
rs = RandomState((seed, i))
x, entropy = entropic_descent_deterministic(
gradfun,
callback=callback,
x_scale=x_init_scale,
epsilon=epsilon,
gamma=gamma,
alpha=alpha,
annealing_schedule=annealing_schedule,
rs=rs,
scale_calc_method="exact_hessian",
hessian=0.5,
)
return results
def run():
(train_images, train_labels),\
(tests_images, tests_labels) = load_data_subset(N_train, N_tests)
parser, pred_fun, nllfun, frac_err = make_nn_funs(layer_sizes)
N_param = len(parser.vect)
print "Running experiment..."
results = defaultdict(list)
for i in xrange(N_samples):
x_init_scale = np.full(N_param, init_scale)
def indexed_loss_fun(w, i_iter):
rs = RandomState((seed, i, i_iter))
idxs = rs.randint(N_train, size=batch_size)
nll = nllfun(w, train_images[idxs], train_labels[idxs]) * N_train
nlp = neg_log_prior(w)
return nll + nlp
gradfun = grad(indexed_loss_fun)
def callback(x, t, v, entropy):
results[("entropy", i)].append(entropy / N_train)
results[("v_norm", i)].append(norm(v) / np.sqrt(N_param))
results[("minibatch_likelihood", i)].append(-indexed_loss_fun(x, t))
results[("log_prior_per_dpt", i)].append(-neg_log_prior(x) / N_train)
if t % thin != 0 and t != N_iter and t != 0: return
results[('iterations', i)].append(t)
results[("train_likelihood", i)].append(-nllfun(x, train_images, train_labels))
results[("tests_likelihood", i)].append(-nllfun(x, tests_images, tests_labels))
results[("tests_error", i)].append(frac_err(x, tests_images, tests_labels))
print "Iteration {0:5} Train likelihood {1:2.4f} Test likelihood {2:2.4f}" \
" Test Err {3:2.4f}".format(t, results[("train_likelihood", i)][-1],
results[("tests_likelihood", i)][-1],
results[("tests_error", i)][-1])
rs = RandomState((seed, i))
entropic_descent_deterministic(gradfun, callback=callback, x_scale=x_init_scale,
epsilon=epsilon, gamma=gamma, alpha=alpha,
annealing_schedule=annealing_schedule, rs=rs)
return results
def run():
x_init_scale = np.full(D, init_scale)
# annealing_schedule = np.linspace(0,1,N_iter)
annealing_schedule = np.concatenate((np.zeros(N_iter/3),
np.linspace(0, 1, N_iter/3),
np.ones(N_iter/3)))
print "Running experiment..."
results = defaultdict(list)
for i in xrange(N_samples):
def callback(x, t, v, entropy):
#results[("x", i)].append(x.copy()) # Replace this with a loop over kwargs?
results[("entropy", i)].append(entropy)
#results[("velocity", i)].append(v)
results[("likelihood", i)].append(-nllfun(x))
rs = RandomState((seed, i))
x, entropy = entropic_descent_deterministic(
gradfun, callback=callback, x_scale=x_init_scale,
epsilon=epsilon, gamma=gamma, alpha=alpha,
annealing_schedule=annealing_schedule, rs=rs,
scale_calc_method='exact_hessian', hessian=0.5)
return results
