def test_ais_with_dbn(self):
dbn = DBN(RBM(10, 10))
dbn.add_layer(RBM(10, 20))
hidden_states = utils.binary_numbers(dbn[0].Y.shape[0])
ais = Estimator(dbn)
ais_logz = ais.estimate_log_partition_function(
100, np.sqrt(np.arange(0, 1, 0.001)))
brf_logz = utils.logsumexp(dbn[1]._ulogprob_vis(hidden_states))
brf_probs = []
dbn_probs = []
dbn_bound = []
for i in range(50):
sample = np.matrix(np.random.rand(10, 1)) > 0.5
prob, bound = ais.estimate_log_probability(sample, 200)
brf_probs.append(
utils.logsumexp(
dbn[1]._ulogprob_vis(hidden_states) +
dbn[0]._clogprob_vis_hid(sample, hidden_states), 1) -
brf_logz)
dbn_probs.append(prob)
dbn_bound.append(bound)
self.assertTrue(np.mean(dbn_bound) < np.mean(dbn_probs))
self.assertTrue(np.abs(np.mean(dbn_probs) - np.mean(brf_probs)) < 0.1)
def test_ais_with_dbn_sanity_check(self):
dbn = DBN(RBM(5, 20))
dbn.add_layer(RBM(20, 5))
dbn.add_layer(RBM(5, 20))
dbn[0].W = np.matrix(np.random.randn(5, 20))
dbn[0].b = np.matrix(np.random.rand(5, 1) - 0.5)
dbn[0].c = np.matrix(np.random.rand(20, 1) - 0.5)
dbn[1].W = dbn[0].W.T
dbn[1].b = dbn[0].c
dbn[1].c = dbn[0].b
dbn[2].W = dbn[0].W
dbn[2].b = dbn[0].b
dbn[2].c = dbn[0].c
samples = dbn.sample(100)
ais = Estimator(dbn[0])
rbm_logz = ais.estimate_log_partition_function(
100, np.sqrt(np.arange(0, 1, 0.001)))
rbm_probs = ais.estimate_log_probability(samples)
ais = Estimator(dbn)
dbn_logz = ais.estimate_log_partition_function(
100, np.sqrt(np.arange(0, 1, 0.001)))
dbn_probs = ais.estimate_log_probability(samples)
self.assertTrue(abs(dbn_logz - rbm_logz) / rbm_logz < 0.02)
self.assertTrue((np.exp(dbn_probs) - np.exp(rbm_probs)).mean() < 0.02)
def test_ais_with_rbm(self): rbm = RBM(5, 20) rbm.W = np.matrix(np.random.randn(5, 20)) rbm.b = np.matrix(np.random.randn(5, 1)) rbm.c = np.matrix(np.random.randn(20, 1)) ais = Estimator(rbm) ais_logz = ais.estimate_log_partition_function(100, np.arange(0, 1, 0.001)) brf_logz = utils.logsumexp(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_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 main(argv):
num_visibles = 28 * 28
num_hiddens = [1000, 1000]
num_epochs = 50
batch_size = 100
# load data samples
data = load('../data/mnist.npz')['train'] / 255.
# train 1st layer
dbn = DBN(RBM(num_visibles, num_hiddens[0]))
dbn.train(data, num_epochs, batch_size)
# train 2nd layer
dbn.add_layer(RBM(num_hiddens[0], num_hiddens[1]))
dbn.train(data, num_epochs, batch_size, [1E-1, 1E-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_ais_with_rbm(self):
rbm = RBM(5, 20)
rbm.W = np.matrix(np.random.randn(5, 20))
rbm.b = np.matrix(np.random.randn(5, 1))
rbm.c = np.matrix(np.random.randn(20, 1))
ais = Estimator(rbm)
ais_logz = ais.estimate_log_partition_function(100,
np.arange(0, 1, 0.001))
brf_logz = utils.logsumexp(
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_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_probs(self): rbm = RBM(9, 11) rbm.W = np.matrix(np.random.randn(rbm.X.shape[0], rbm.Y.shape[0])) rbm.b = np.matrix(np.random.rand(rbm.X.shape[0], 1)) rbm.c = np.matrix(np.random.randn(rbm.Y.shape[0], 1)) examples_vis = np.matrix(np.random.rand(rbm.X.shape[0], 100) < 0.5) examples_hid = np.matrix(np.random.rand(rbm.Y.shape[0], 100) < 0.5) states_vis = utils.binary_numbers(rbm.X.shape[0]) states_hid = utils.binary_numbers(rbm.Y.shape[0]) # check that conditional probabilities are normalized logprobs = rbm._clogprob_vis_hid(states_vis, examples_hid, all_pairs=True) self.assertTrue(np.all(utils.logsumexp(logprobs, 0) < 1E-10)) logprobs = rbm._clogprob_hid_vis(examples_vis, states_hid, all_pairs=True) self.assertTrue(np.all(utils.logsumexp(logprobs, 1) < 1E-10)) # test for consistency logprobs1 = rbm._ulogprob(examples_vis, examples_hid, all_pairs=True) logprobs2 = rbm._clogprob_vis_hid(examples_vis, examples_hid, all_pairs=True) \ + rbm._ulogprob_hid(examples_hid) logprobs3 = rbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True) \ + rbm._ulogprob_vis(examples_vis).T self.assertTrue(np.all(np.abs(logprobs1 - logprobs2) < 1E-10)) self.assertTrue(np.all(np.abs(logprobs1 - logprobs3) < 1E-10))
def test_all_pairs(self): rbm = RBM(10, 20) rbm.W = np.matrix(np.random.randn(rbm.X.shape[0], rbm.Y.shape[0])) rbm.b = np.matrix(np.random.rand(rbm.X.shape[0], 1)) rbm.c = np.matrix(np.random.randn(rbm.Y.shape[0], 1)) examples_vis = np.matrix(np.random.rand(rbm.X.shape[0], 100) < 0.5) examples_hid = np.matrix(np.random.rand(rbm.Y.shape[0], 100) < 0.5) logprob1 = rbm._ulogprob(examples_vis, examples_hid) logprob2 = np.diag(rbm._ulogprob(examples_vis, examples_hid, all_pairs=True)) self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = rbm._ulogprob(examples_vis[:, 1], examples_hid) logprob2 = rbm._ulogprob(examples_vis, examples_hid, all_pairs=True)[1, :] self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = rbm._ulogprob(examples_vis, examples_hid[:, 1]) logprob2 = rbm._ulogprob(examples_vis, examples_hid, all_pairs=True)[:, 1].T self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = rbm._clogprob_vis_hid(examples_vis, examples_hid) logprob2 = np.diag(rbm._clogprob_vis_hid(examples_vis, examples_hid, all_pairs=True)) self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = rbm._clogprob_vis_hid(examples_vis[:, 1], examples_hid) logprob2 = rbm._clogprob_vis_hid(examples_vis, examples_hid, all_pairs=True)[1, :] self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = rbm._clogprob_vis_hid(examples_vis, examples_hid[:, 1]) logprob2 = rbm._clogprob_vis_hid(examples_vis, examples_hid, all_pairs=True)[:, 1].T self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = rbm._clogprob_hid_vis(examples_vis, examples_hid) logprob2 = np.diag(rbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True)) self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = rbm._clogprob_hid_vis(examples_vis[:, 1], examples_hid) logprob2 = rbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True)[1, :] self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10) logprob1 = rbm._clogprob_hid_vis(examples_vis, examples_hid[:, 1]) logprob2 = rbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True)[:, 1].T self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
def test_all_pairs(self):
rbm = RBM(10, 20)
rbm.W = np.matrix(np.random.randn(rbm.X.shape[0], rbm.Y.shape[0]))
rbm.b = np.matrix(np.random.rand(rbm.X.shape[0], 1))
rbm.c = np.matrix(np.random.randn(rbm.Y.shape[0], 1))
examples_vis = np.matrix(np.random.rand(rbm.X.shape[0], 100) < 0.5)
examples_hid = np.matrix(np.random.rand(rbm.Y.shape[0], 100) < 0.5)
logprob1 = rbm._ulogprob(examples_vis, examples_hid)
logprob2 = np.diag(
rbm._ulogprob(examples_vis, examples_hid, all_pairs=True))
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = rbm._ulogprob(examples_vis[:, 1], examples_hid)
logprob2 = rbm._ulogprob(examples_vis, examples_hid,
all_pairs=True)[1, :]
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = rbm._ulogprob(examples_vis, examples_hid[:, 1])
logprob2 = rbm._ulogprob(examples_vis, examples_hid,
all_pairs=True)[:, 1].T
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = rbm._clogprob_vis_hid(examples_vis, examples_hid)
logprob2 = np.diag(
rbm._clogprob_vis_hid(examples_vis, examples_hid, all_pairs=True))
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = rbm._clogprob_vis_hid(examples_vis[:, 1], examples_hid)
logprob2 = rbm._clogprob_vis_hid(examples_vis,
examples_hid,
all_pairs=True)[1, :]
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = rbm._clogprob_vis_hid(examples_vis, examples_hid[:, 1])
logprob2 = rbm._clogprob_vis_hid(examples_vis,
examples_hid,
all_pairs=True)[:, 1].T
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = rbm._clogprob_hid_vis(examples_vis, examples_hid)
logprob2 = np.diag(
rbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True))
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = rbm._clogprob_hid_vis(examples_vis[:, 1], examples_hid)
logprob2 = rbm._clogprob_hid_vis(examples_vis,
examples_hid,
all_pairs=True)[1, :]
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
logprob1 = rbm._clogprob_hid_vis(examples_vis, examples_hid[:, 1])
logprob2 = rbm._clogprob_hid_vis(examples_vis,
examples_hid,
all_pairs=True)[:, 1].T
self.assertTrue(np.abs(logprob1 - logprob2).sum() < 1E-10)
def test_probs(self):
rbm = RBM(9, 11)
rbm.W = np.matrix(np.random.randn(rbm.X.shape[0], rbm.Y.shape[0]))
rbm.b = np.matrix(np.random.rand(rbm.X.shape[0], 1))
rbm.c = np.matrix(np.random.randn(rbm.Y.shape[0], 1))
examples_vis = np.matrix(np.random.rand(rbm.X.shape[0], 100) < 0.5)
examples_hid = np.matrix(np.random.rand(rbm.Y.shape[0], 100) < 0.5)
states_vis = utils.binary_numbers(rbm.X.shape[0])
states_hid = utils.binary_numbers(rbm.Y.shape[0])
# check that conditional probabilities are normalized
logprobs = rbm._clogprob_vis_hid(states_vis,
examples_hid,
all_pairs=True)
self.assertTrue(np.all(utils.logsumexp(logprobs, 0) < 1E-10))
logprobs = rbm._clogprob_hid_vis(examples_vis,
states_hid,
all_pairs=True)
self.assertTrue(np.all(utils.logsumexp(logprobs, 1) < 1E-10))
# test for consistency
logprobs1 = rbm._ulogprob(examples_vis, examples_hid, all_pairs=True)
logprobs2 = rbm._clogprob_vis_hid(examples_vis, examples_hid, all_pairs=True) \
+ rbm._ulogprob_hid(examples_hid)
logprobs3 = rbm._clogprob_hid_vis(examples_vis, examples_hid, all_pairs=True) \
+ rbm._ulogprob_vis(examples_vis).T
self.assertTrue(np.all(np.abs(logprobs1 - logprobs2) < 1E-10))
self.assertTrue(np.all(np.abs(logprobs1 - logprobs3) < 1E-10))
