def test_ais_with_semirbm_dbn(self):
dbn = DBN(RBM(5, 5))
dbn.add_layer(SemiRBM(5, 5))
ais = Estimator(dbn)
ais.estimate_log_partition_function(100,
np.arange(0, 1, 1E-3),
layer=0)
ais.estimate_log_partition_function(10, np.arange(0, 1, 1E-3), layer=1)
dbn[0]._brf_logz = utils.logsumexp(dbn[0]._ulogprob_vis(
utils.binary_numbers(dbn[0].X.shape[0])))
dbn[1]._brf_logz = utils.logsumexp(dbn[1]._ulogprob_vis(
utils.binary_numbers(dbn[1].X.shape[0])))
samples = np.concatenate(
[dbn.sample(25, 100, 20),
np.matrix(np.random.rand(5, 25) > 0.5)], 1)
Y = utils.binary_numbers(dbn[0].Y.shape[0])
X = utils.binary_numbers(dbn[0].X.shape[0])
logRy = dbn[1]._ulogprob_vis(Y)
logQy = utils.logsumexp(dbn[0]._ulogprob(X, Y, all_pairs=True), 0)
log_sum = utils.logsumexp(
dbn[0]._clogprob_hid_vis(samples, Y, all_pairs=True) - logQy +
logRy, 1)
logPx = log_sum + dbn[0]._ulogprob_vis(samples) - dbn[1]._brf_logz
logPx_ = ais.estimate_log_probability(samples)[0]
self.assertTrue(np.abs(logPx_.mean() - logPx.mean()) < 0.1)
def test_ais_with_semirbm_sanity_check(self):
grbm = GaussianRBM(15, 50)
grbm.b = np.random.randn(grbm.b.shape[0], 1)
grbm.c = np.random.randn(grbm.c.shape[0], 1)
srbm = SemiRBM(50, 20)
srbm.W = srbm.W * 0.
srbm.c = srbm.c * 0.
srbm.L = grbm.W.T * grbm.W
srbm.b = grbm.W.T * grbm.b + grbm.c + 0.5 * np.matrix(np.diag(
srbm.L)).T
srbm.L = srbm.L - np.matrix(np.diag(np.diag(srbm.L)))
ais = Estimator(grbm)
ais.estimate_log_partition_function(num_ais_samples=100,
beta_weights=np.arange(0, 1, 1E-3))
ais = Estimator(srbm)
ais.estimate_log_partition_function(num_ais_samples=100,
beta_weights=np.arange(0, 1, 1E-2))
glogz = grbm._ais_logz + srbm.Y.shape[0] * np.log(2)
slogz = srbm._ais_logz + grbm.X.shape[0] * np.log(np.sqrt(2 * np.pi))
self.assertTrue(np.abs(glogz - slogz) < 1.)
def test_ais_with_semirbm(self): rbm = SemiRBM(5, 20) rbm.L = np.matrix(np.random.randn(5, 5)) rbm.L = np.triu(rbm.L) + np.triu(rbm.L).T - 2 * np.diag(np.diag(rbm.L)) rbm.num_lateral_updates = 5 rbm.sampling_method = SemiRBM.GIBBS ais = Estimator(rbm) ais_logz = ais.estimate_log_partition_function(100, np.arange(0, 1, 0.001)) brf_logz = np.log(np.sum(np.exp(rbm._ulogprob_vis(utils.binary_numbers(rbm.X.shape[0]))))) lower = np.log(np.exp(ais_logz) - 4 * np.sqrt(rbm._ais_var)) upper = np.log(np.exp(ais_logz) + 4 * np.sqrt(rbm._ais_var)) self.assertTrue(upper - lower < 1.) self.assertTrue(lower < brf_logz and brf_logz < upper)
def test_ais_with_semirbm(self):
rbm = SemiRBM(5, 20)
rbm.L = np.matrix(np.random.randn(5, 5))
rbm.L = np.triu(rbm.L) + np.triu(rbm.L).T - 2 * np.diag(np.diag(rbm.L))
rbm.num_lateral_updates = 5
rbm.sampling_method = SemiRBM.GIBBS
ais = Estimator(rbm)
ais_logz = ais.estimate_log_partition_function(100,
np.arange(0, 1, 0.001))
brf_logz = np.log(
np.sum(
np.exp(rbm._ulogprob_vis(utils.binary_numbers(
rbm.X.shape[0])))))
lower = np.log(np.exp(ais_logz) - 4 * np.sqrt(rbm._ais_var))
upper = np.log(np.exp(ais_logz) + 4 * np.sqrt(rbm._ais_var))
self.assertTrue(upper - lower < 1.)
self.assertTrue(lower < brf_logz and brf_logz < upper)
def test_ais_with_semirbm_sanity_check(self): grbm = GaussianRBM(15, 50) grbm.b = np.random.randn(grbm.b.shape[0], 1) grbm.c = np.random.randn(grbm.c.shape[0], 1) srbm = SemiRBM(50, 20) srbm.W = srbm.W * 0. srbm.c = srbm.c * 0. srbm.L = grbm.W.T * grbm.W srbm.b = grbm.W.T * grbm.b + grbm.c + 0.5 * np.matrix(np.diag(srbm.L)).T srbm.L = srbm.L - np.matrix(np.diag(np.diag(srbm.L))) ais = Estimator(grbm) ais.estimate_log_partition_function(num_ais_samples=100, beta_weights=np.arange(0, 1, 1E-3)) ais = Estimator(srbm) ais.estimate_log_partition_function(num_ais_samples=100, beta_weights=np.arange(0, 1, 1E-2)) glogz = grbm._ais_logz + srbm.Y.shape[0] * np.log(2) slogz = srbm._ais_logz + grbm.X.shape[0] * np.log(np.sqrt(2 * np.pi)) self.assertTrue(np.abs(glogz - slogz) < 1.)
def test_all_pairs(self):
srbm = SemiRBM(10, 10)
srbm.W = np.matrix(np.random.randn(srbm.X.shape[0], srbm.Y.shape[0]))
srbm.b = np.matrix(np.random.rand(srbm.X.shape[0], 1))
srbm.c = np.matrix(np.random.randn(srbm.Y.shape[0], 1))
examples_vis = np.matrix(np.random.rand(srbm.X.shape[0], 100) < 0.5)
examples_hid = np.matrix(np.random.rand(srbm.Y.shape[0], 100) < 0.5)
logprob1 = srbm._ulogprob(examples_vis, examples_hid)
logprob2 = np.diag(
srbm._ulogprob(examples_vis, examples_hid, all_pairs=True))
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = srbm._ulogprob(examples_vis[:, 1], examples_hid)
logprob2 = srbm._ulogprob(examples_vis, examples_hid,
all_pairs=True)[1, :]
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = srbm._ulogprob(examples_vis, examples_hid[:, 1])
logprob2 = srbm._ulogprob(examples_vis, examples_hid,
all_pairs=True)[:, 1].T
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = srbm._clogprob_hid_vis(examples_vis, examples_hid)
logprob2 = np.diag(
srbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True))
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = srbm._clogprob_hid_vis(examples_vis[:, 1], examples_hid)
logprob2 = srbm._clogprob_hid_vis(examples_vis,
examples_hid,
all_pairs=True)[1, :]
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = srbm._clogprob_hid_vis(examples_vis, examples_hid[:, 1])
logprob2 = srbm._clogprob_hid_vis(examples_vis,
examples_hid,
all_pairs=True)[:, 1].T
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
def test_probabilities(self):
srbm = SemiRBM(9, 12)
srbm.W = np.matrix(np.random.randn(srbm.X.shape[0], srbm.Y.shape[0]))
srbm.b = np.matrix(np.random.rand(srbm.X.shape[0], 1))
srbm.c = np.matrix(np.random.randn(srbm.Y.shape[0], 1))
srbm.L = np.matrix(np.random.randn(srbm.X.shape[0],
srbm.X.shape[0])) / 2.
srbm.L = np.triu(srbm.L) + np.triu(
srbm.L).T - 2. * np.diag(np.diag(srbm.L))
examples_vis = np.matrix(np.random.rand(srbm.X.shape[0], 100) < 0.5)
examples_hid = np.matrix(np.random.rand(srbm.Y.shape[0], 100) < 0.5)
states_vis = utils.binary_numbers(srbm.X.shape[0])
states_hid = utils.binary_numbers(srbm.Y.shape[0])
# check that conditional probabilities are normalized
logprobs = srbm._clogprob_hid_vis(examples_vis,
states_hid,
all_pairs=True)
self.assertTrue(np.all(utils.logsumexp(logprobs, 1) < 1E-10))
# test for consistency
logprobs1 = srbm._ulogprob(examples_vis, examples_hid, all_pairs=True)
logprobs3 = srbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True) \
+ srbm._ulogprob_vis(examples_vis).T
self.assertTrue(np.all(np.abs(logprobs1 - logprobs3) < 1E-10))
rbm = RBM(srbm.X.shape[0], srbm.Y.shape[0])
rbm.W = srbm.W
rbm.b = srbm.b
rbm.c = srbm.c
srbm.L *= 0
logprobs1 = rbm._ulogprob_vis(examples_vis)
logprobs2 = srbm._ulogprob_vis(examples_vis)
self.assertTrue(np.all(np.abs(logprobs1 - logprobs2) < 1E-10))
logprobs1 = rbm._clogprob_hid_vis(examples_vis, examples_hid)
logprobs2 = srbm._clogprob_hid_vis(examples_vis, examples_hid)
self.assertTrue(np.all(np.abs(logprobs1 - logprobs2) < 1E-10))
def main(argv):
# load preprocessed data samples
print 'loading data...\t',
data = load('./data/vanhateren.npz')
print '[DONE]'
print
# remove DC component (first component)
data_train = data['train'][1:, :]
data_test = data['test'][1:, :]
# create 1st layer
dbn = DBN(GaussianRBM(num_visibles=data_train.shape[0], num_hiddens=100))
# hyperparameters
dbn[0].learning_rate = 5E-3
dbn[0].weight_decay = 1E-2
dbn[0].momentum = 0.9
dbn[0].sigma = 0.65
dbn[0].cd_steps = 1
dbn[0].persistent = True
# train 1st layer
print 'training...\t',
dbn.train(data_train, num_epochs=100, batch_size=100)
print '[DONE]'
# evaluate 1st layer
print 'evaluating...\t',
logptf = dbn.estimate_log_partition_function(num_ais_samples=100, beta_weights=arange(0, 1, 1E-3))
loglik = dbn.estimate_log_likelihood(data_test)
print '[DONE]'
print
print 'estimated log-partf.:\t', logptf
print 'estimated log-loss:\t', -loglik / data_test.shape[0] / log(2)
print
# create 2nd layer
dbn.add_layer(SemiRBM(num_visibles=100, num_hiddens=100))
# initialize parameters
dbn[1].L = dbn[0].W.T * dbn[0].W
dbn[1].b = dbn[0].W.T * dbn[0].b + dbn[0].c + 0.5 * asmatrix(diag(dbn[1].L)).T
dbn[1].L = dbn[1].L - asmatrix(diag(diag(dbn[1].L)))
# hyperparameters
dbn[1].learning_rate = 5E-3
dbn[1].learning_rate_lateral = 5E-4
dbn[1].weight_decay = 5E-3
dbn[1].weight_decay_lateral = 5E-3
dbn[1].momentum = 0.9
dbn[1].momentum_lateral = 0.9
dbn[1].num_lateral_updates = 20
dbn[1].damping = 0.2
dbn[1].cd_steps = 1
dbn[1].persistent = True
# train 2nd layer
print 'training...\t',
dbn.train(data_train, num_epochs=100, batch_size=100)
print '[DONE]'
# evaluate 2nd layer
print 'evaluating...\t',
logptf = dbn.estimate_log_partition_function(num_ais_samples=100, beta_weights=arange(0, 1, 1E-3))
loglik = dbn.estimate_log_likelihood(data_test, num_samples=100)
print '[DONE]'
print
print 'estimated log-partf.:\t', logptf
print 'estimated log-loss:\t', -loglik / data_test.shape[0] / log(2)
print
# fine-tune with wake-sleep
dbn[0].learning_rate /= 4.
dbn[1].learning_rate /= 4.
print 'fine-tuning...\t',
dbn.train_wake_sleep(data_train, num_epochs=10, batch_size=10)
print '[DONE]'
# reevaluate
print 'evaluating...\t',
logptf = dbn.estimate_log_partition_function(num_ais_samples=100, beta_weights=arange(0, 1, 1E-3))
loglik = dbn.estimate_log_likelihood(data_test, num_samples=100)
print '[DONE]'
print
print 'estimated log-partf.:\t', logptf
print 'estimated log-loss:\t', -loglik / data_test.shape[0] / log(2)
return 0
def test_probabilities(self): srbm = SemiRBM(9, 12) srbm.W = np.matrix(np.random.randn(srbm.X.shape[0], srbm.Y.shape[0])) srbm.b = np.matrix(np.random.rand(srbm.X.shape[0], 1)) srbm.c = np.matrix(np.random.randn(srbm.Y.shape[0], 1)) srbm.L = np.matrix(np.random.randn(srbm.X.shape[0], srbm.X.shape[0])) / 2. srbm.L = np.triu(srbm.L) + np.triu(srbm.L).T - 2. * np.diag(np.diag(srbm.L)) examples_vis = np.matrix(np.random.rand(srbm.X.shape[0], 100) < 0.5) examples_hid = np.matrix(np.random.rand(srbm.Y.shape[0], 100) < 0.5) states_vis = utils.binary_numbers(srbm.X.shape[0]) states_hid = utils.binary_numbers(srbm.Y.shape[0]) # check that conditional probabilities are normalized logprobs = srbm._clogprob_hid_vis(examples_vis, states_hid, all_pairs=True) self.assertTrue(np.all(utils.logsumexp(logprobs, 1) < 1E-10)) # test for consistency logprobs1 = srbm._ulogprob(examples_vis, examples_hid, all_pairs=True) logprobs3 = srbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True) \ + srbm._ulogprob_vis(examples_vis).T self.assertTrue(np.all(np.abs(logprobs1 - logprobs3) < 1E-10)) rbm = RBM(srbm.X.shape[0], srbm.Y.shape[0]) rbm.W = srbm.W rbm.b = srbm.b rbm.c = srbm.c srbm.L *= 0 logprobs1 = rbm._ulogprob_vis(examples_vis) logprobs2 = srbm._ulogprob_vis(examples_vis) self.assertTrue(np.all(np.abs(logprobs1 - logprobs2) < 1E-10)) logprobs1 = rbm._clogprob_hid_vis(examples_vis, examples_hid) logprobs2 = srbm._clogprob_hid_vis(examples_vis, examples_hid) self.assertTrue(np.all(np.abs(logprobs1 - logprobs2) < 1E-10))
def test_all_pairs(self): srbm = SemiRBM(10, 10) srbm.W = np.matrix(np.random.randn(srbm.X.shape[0], srbm.Y.shape[0])) srbm.b = np.matrix(np.random.rand(srbm.X.shape[0], 1)) srbm.c = np.matrix(np.random.randn(srbm.Y.shape[0], 1)) examples_vis = np.matrix(np.random.rand(srbm.X.shape[0], 100) < 0.5) examples_hid = np.matrix(np.random.rand(srbm.Y.shape[0], 100) < 0.5) logprob1 = srbm._ulogprob(examples_vis, examples_hid) logprob2 = np.diag(srbm._ulogprob(examples_vis, examples_hid, all_pairs=True)) self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = srbm._ulogprob(examples_vis[:, 1], examples_hid) logprob2 = srbm._ulogprob(examples_vis, examples_hid, all_pairs=True)[1, :] self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = srbm._ulogprob(examples_vis, examples_hid[:, 1]) logprob2 = srbm._ulogprob(examples_vis, examples_hid, all_pairs=True)[:, 1].T self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = srbm._clogprob_hid_vis(examples_vis, examples_hid) logprob2 = np.diag(srbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True)) self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = srbm._clogprob_hid_vis(examples_vis[:, 1], examples_hid) logprob2 = srbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True)[1, :] self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = srbm._clogprob_hid_vis(examples_vis, examples_hid[:, 1]) logprob2 = srbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True)[:, 1].T self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
