def setUp(self):
self.NUM_V = 6
self.NUM_H = 5
self.BATCH_SIZE = 20
self.V_BATCH_SHAPE = (self.BATCH_SIZE, 6)
self.H_BATCH_SHAPE = (self.BATCH_SIZE, 5)
self.W_STD = 0.01
self.GRAD_ERROR = 1e-3
self.sess = tf.Session()
self.rbm = RBM(num_v=self.NUM_V,
num_h=self.NUM_H,
W_std=self.W_STD,
session=self.sess)
self.h_batch = tf.constant(np.random.binomial(1,
0.5,
size=self.H_BATCH_SHAPE),
dtype=c.float_type)
self.v_batch = tf.constant(np.random.binomial(1,
0.5,
size=self.V_BATCH_SHAPE),
dtype=c.float_type)
h_probs = self.rbm.h_probs(self.v_batch)
self.sample_h = self.rbm.sample_h(h_probs)
v_probs = self.rbm.v_probs(self.h_batch)
self.sample_v = self.rbm.sample_v(v_probs)
self.free_energy = self.rbm.free_energy(self.v_batch)
self.sess.run(tf.global_variables_initializer())
test_data_batch = test_data.next_batch()
config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
session = tf.Session(config=config)
layer1 = SBM_Lower(session=session,
side=flags['img_shape'][0],
side_overlap=flags['layer_1_side_overlap'],
num_h=flags['layer_1_num_h'],
name=flags['layer_1_name'])
layer1.restore(flags['layer_1_ckpt'])
layer2 = RBM(session=session,
num_v=flags['layer_1_num_h'],
num_h=flags['layer_2_num_h'],
bottom=layer1,
name=flags['layer_2_name'])
layer2.restore(flags['layer_2_ckpt'])
table_path = os.path.join(flags['test_generalisation_plots_dir'], 'table')
layer2.test_generalisation(
test_data=test_data_batch,
output_table_path=table_path,
num_iterations=flags['test_generalisation_num_iterations'],
num_runs=flags['test_generalisation_num_runs'],
num_samples_to_average=flags['test_generalisation_num_samples_to_average'])
with open(table_path, 'rb') as f:
batch_size=flags['test_generalisation_batch_size'],
log_epochs=flags['data_log_epochs'],
name='TestData')
test_data_batch = test_data.next_batch()
x_test = tf.get_variable(name='x_test',
initializer=tf.random_normal(shape=(1, flags['q']),
dtype=c.float_type))
config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
session = tf.Session(config=config)
layer1 = RBM(session=session,
num_v=flags['img_shape'][0] * flags['img_shape'][1],
num_h=flags['layer_1_num_h'],
temperature=flags['temperature'],
name=flags['layer_1_name'])
layer1.restore(flags['layer_1_ckpt'])
layer2 = RBM(session=session,
num_v=flags['layer_1_num_h'],
num_h=flags['layer_2_num_h'],
temperature=flags['temperature'],
bottom=layer1,
name=flags['layer_2_name'])
layer2.restore(flags['layer_2_ckpt'])
kern = gplvm.SEKernel(session=session,
name='TrainingData')
test_data = Data(flags['test_data'],
shuffle_first=flags['shuffle_data'],
batch_size=flags['test_batch_size'],
log_epochs=flags['data_log_epochs'],
name='TestData')
optimizer = tf.train.AdamOptimizer(learning_rate=flags['learning_rate'])
config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
session = tf.Session(config=config)
layer1 = RBM(session=session,
num_v=flags['img_shape'][0] * flags['img_shape'][1],
num_h=flags['layer_1_num_h'],
lr=flags['learning_rate'],
W_std=flags['layer_1_W_std'],
loss_plotter=loss_plotter,
distr_plotter=distr_plotter,
name=flags['layer_1_name'])
layer1.train(optimizer=optimizer,
training_data=training_data,
test_data=test_data,
num_gibbs_steps=flags['num_gibbs_steps'],
pcd=flags['pcd'],
max_iterations=flags['num_iterations'],
eval_interval=flags['eval_interval'],
ckpt_interval=flags['ckpt_interval'],
ckpt_dir=flags['ckpt_dir'])
layer2 = RBM(session=session,
class TestRBM(unittest.TestCase):
def setUp(self):
self.NUM_V = 6
self.NUM_H = 5
self.BATCH_SIZE = 20
self.V_BATCH_SHAPE = (self.BATCH_SIZE, 6)
self.H_BATCH_SHAPE = (self.BATCH_SIZE, 5)
self.W_STD = 0.01
self.GRAD_ERROR = 1e-3
self.sess = tf.Session()
self.rbm = RBM(num_v=self.NUM_V,
num_h=self.NUM_H,
W_std=self.W_STD,
session=self.sess)
self.h_batch = tf.constant(np.random.binomial(1,
0.5,
size=self.H_BATCH_SHAPE),
dtype=c.float_type)
self.v_batch = tf.constant(np.random.binomial(1,
0.5,
size=self.V_BATCH_SHAPE),
dtype=c.float_type)
h_probs = self.rbm.h_probs(self.v_batch)
self.sample_h = self.rbm.sample_h(h_probs)
v_probs = self.rbm.v_probs(self.h_batch)
self.sample_v = self.rbm.sample_v(v_probs)
self.free_energy = self.rbm.free_energy(self.v_batch)
self.sess.run(tf.global_variables_initializer())
def tearDown(self):
tf.reset_default_graph()
self.sess.close()
def test_sample_h_shape(self):
self.assertEqual(
self.sess.run(self.sample_h).shape, self.H_BATCH_SHAPE)
def test_sample_h_values(self):
values = [0., 1.]
self.assertTrue(np.all(np.in1d(self.sess.run(self.sample_h), values)))
def test_sample_v_shape(self):
self.assertEqual(
self.sess.run(self.sample_v).shape, self.V_BATCH_SHAPE)
def test_sample_v_values(self):
values = [0., 1.]
self.assertTrue(np.all(np.in1d(self.sess.run(self.sample_v), values)))
def test_free_energy_shape(self):
free_energy_arr = self.sess.run(self.free_energy)
self.assertTrue(free_energy_arr.shape, [self.BATCH_SIZE, 1])
def test_free_energy_W_grad(self):
with self.sess as sess:
error = tf.test.compute_gradient_error(
self.rbm.W, self.rbm.W.shape.as_list(), self.free_energy,
self.free_energy.shape.as_list())
self.assertLess(error, self.GRAD_ERROR)
def test_free_energy_b_grad(self):
with self.sess as sess:
error = tf.test.compute_gradient_error(
self.rbm.b, self.rbm.b.shape.as_list(), self.free_energy,
self.free_energy.shape.as_list())
self.assertLess(error, self.GRAD_ERROR)
def test_free_energy_c_grad(self):
with self.sess as sess:
error = tf.test.compute_gradient_error(
self.rbm.c, self.rbm.c.shape.as_list(), self.free_energy,
self.free_energy.shape.as_list())
self.assertLess(error, self.GRAD_ERROR)
output_dir=flags['latent_samples_dir'])
training_data = Data(flags['training_data'],
shuffle_first=flags['shuffle_data'],
batch_size=flags['training_batch_size'],
log_epochs=flags['data_log_epochs'],
name='TrainingData')
Y = training_data.next_batch()
config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
session = tf.Session(config=config)
layer1 = RBM(session=session,
num_v=flags['img_shape'][0] * flags['img_shape'][1],
num_h=flags['layer_1_num_h'],
W_std=flags['layer_1_W_std'],
temperature=flags['temperature'],
name=flags['layer_1_name'])
layer1.init_variables()
layer2 = RBM(session=session,
num_v=flags['layer_1_num_h'],
num_h=flags['layer_2_num_h'],
W_std=flags['layer_2_W_std'],
temperature=flags['temperature'],
bottom=layer1,
name=flags['layer_2_name'])
layer2.init_variables()