def init_variables(self):
self.variables.update(
h_samples=theano.shared(
name='algo:rbm/matrix:hidden-samples',
value=asint(np.zeros((self.batch_size, self.n_hidden))),
),
)
def test_asint(self):
int2float_types = {
'float32': 'int32',
'float64': 'int64',
}
int_type = int2float_types[theano.config.floatX]
# Sparse matrix
sparse_matrix = csr_matrix((3, 4), dtype=np.int8)
self.assertIs(sparse_matrix, asint(sparse_matrix))
# Numpy array-like elements
x = np.array([1, 2, 3], dtype=int_type)
self.assertIs(x, asint(x))
x = np.array([1, 2, 3], dtype=np.int8)
self.assertIsNot(x, asint(x))
# Python list
x = [1, 2, 3]
self.assertEqual(asint(x).shape, (3, ))
# Theano variables
x = theano.tensor.fmatrix()
self.assertNotEqual(x.dtype, int_type)
self.assertEqual(asint(x).dtype, int_type)
def test_asint(self):
int2float_types = {
'float32': 'int32',
'float64': 'int64',
}
int_type = int2float_types[theano.config.floatX]
# Sparse matrix
sparse_matrix = csr_matrix((3, 4), dtype=np.int8)
self.assertIs(sparse_matrix, asint(sparse_matrix))
# Numpy array-like elements
x = np.array([1, 2, 3], dtype=int_type)
self.assertIs(x, asint(x))
x = np.array([1, 2, 3], dtype=np.int8)
self.assertIsNot(x, asint(x))
# Python list
x = [1, 2, 3]
self.assertEqual(asint(x).shape, (3,))
# Theano variables
x = theano.tensor.fmatrix()
self.assertNotEqual(x.dtype, int_type)
self.assertEqual(asint(x).dtype, int_type)
def categorical_hinge(expected, predicted, delta=1):
"""
Computes the multi-class hinge loss between
predictions and targets.
.. math::
hinge_{i}(t, o) = \\max_{j \\not = o_i} (0, t_j - t_{o_i} + \\delta)
Parameters
----------
expected : Theano 2D tensor or 1D tensor
Either a vector of int giving the correct class index
per data point or a 2D tensor of one-hot encoding of
the correct class in the same layout as predictions
(non-binary targets in [0, 1] do not work!).
predicted : Theano 2D tensor
Predictions in (0, 1), such as softmax output of
a neural network, with data points in rows and class
probabilities in columns.
delta : scalar
The hinge loss margin. Defaults to ``1``.
Returns
-------
Theano 1D tensor
An expression for the average multi-class hinge loss.
Notes
-----
This is an alternative to the categorical cross-entropy
loss for multi-class classification problems.
"""
n_classes = predicted.shape[1]
if expected.ndim == (predicted.ndim - 1):
expected = T.extra_ops.to_one_hot(asint(expected), n_classes)
if expected.ndim != predicted.ndim:
raise TypeError('Rank mismatch between expected and prediced values')
invalid_class_indeces = expected.nonzero()
valid_class_indeces = (1 - expected).nonzero()
new_shape = (-1, n_classes - 1)
rest = T.reshape(predicted[valid_class_indeces], new_shape)
rest = T.max(rest, axis=1)
corrects = predicted[invalid_class_indeces]
error = T.nnet.relu(rest - corrects + delta)
return error.mean()
def output(self, Q, input_state_1, input_state_2):
# Number of samples dependce on the state batch size.
# Each iteration we can try to predict direction from
# multiple different starting points at the same time.
n_states = input_state_1.shape[1]
# Output is a matrix that has n_samples * n_states rows
# and n_filters (which is Q.shape[1]) columns.
return Q[
# Numer of repetitions depends on the size of
# the state batch
T.extra_ops.repeat(T.arange(Q.shape[0]), n_states),
# Extract all channels
:,
# Each state is a coordinate (x and y)
# that point to some place on a grid.
asint(input_state_1.flatten()),
asint(input_state_2.flatten()),
]
def categorical_hinge(expected, predicted, delta=1):
"""
Computes the multi-class hinge loss between
predictions and targets.
.. math::
hinge_{i}(t, o) = \\max_{j \\not = o_i} (0, t_j - t_{o_i} + \\delta)
Parameters
----------
expected : Theano 2D tensor or 1D tensor
Either a vector of int giving the correct class index
per data point or a 2D tensor of one-hot encoding of
the correct class in the same layout as predictions
(non-binary targets in [0, 1] do not work!).
predicted : Theano 2D tensor
Predictions in (0, 1), such as softmax output of
a neural network, with data points in rows and class
probabilities in columns.
delta : scalar
The hinge loss margin. Defaults to ``1``.
Returns
-------
Theano 1D tensor
An expression for the average multi-class hinge loss.
Notes
-----
This is an alternative to the categorical cross-entropy
loss for multi-class classification problems.
"""
n_classes = predicted.shape[1]
if expected.ndim == (predicted.ndim - 1):
expected = T.extra_ops.to_one_hot(asint(expected), n_classes)
if expected.ndim != predicted.ndim:
raise TypeError('Rank mismatch between expected and prediced values')
invalid_class_indeces = expected.nonzero()
valid_class_indeces = (1 - expected).nonzero()
new_shape = (-1, n_classes - 1)
rest = T.reshape(predicted[valid_class_indeces], new_shape)
rest = T.max(rest, axis=1)
corrects = predicted[invalid_class_indeces]
error = T.nnet.relu(rest - corrects + delta)
return error.mean()
def init_methods(self):
def free_energy(visible_sample):
wx_b = T.dot(visible_sample, self.weight) + self.hidden_bias
visible_bias_term = T.dot(visible_sample, self.visible_bias)
hidden_term = T.log(asfloat(1) + T.exp(wx_b)).sum(axis=1)
return -visible_bias_term - hidden_term
def visible_to_hidden(visible_sample):
wx_b = T.dot(visible_sample, self.weight) + self.hidden_bias
return T.nnet.sigmoid(wx_b)
def hidden_to_visible(hidden_sample):
wx_b = T.dot(hidden_sample, self.weight.T) + self.visible_bias
return T.nnet.sigmoid(wx_b)
def sample_hidden_from_visible(visible_sample):
theano_random = self.theano_random
hidden_prob = visible_to_hidden(visible_sample)
hidden_sample = theano_random.binomial(n=1,
p=hidden_prob,
dtype=theano.config.floatX)
return hidden_sample
def sample_visible_from_hidden(hidden_sample):
theano_random = self.theano_random
visible_prob = hidden_to_visible(hidden_sample)
visible_sample = theano_random.binomial(n=1,
p=visible_prob,
dtype=theano.config.floatX)
return visible_sample
network_input = self.variables.network_input
n_samples = asfloat(network_input.shape[0])
theano_random = self.theano_random
weight = self.weight
h_bias = self.hidden_bias
v_bias = self.visible_bias
h_samples = self.variables.h_samples
step = asfloat(self.step)
sample_indeces = theano_random.random_integers(
low=0, high=n_samples - 1, size=(self.batch_size, ))
v_pos = ifelse(
T.eq(n_samples, self.batch_size),
network_input,
# In case if final batch has less number of
# samples then expected
network_input[sample_indeces])
h_pos = visible_to_hidden(v_pos)
v_neg = sample_visible_from_hidden(h_samples)
h_neg = visible_to_hidden(v_neg)
weight_update = v_pos.T.dot(h_pos) - v_neg.T.dot(h_neg)
h_bias_update = (h_pos - h_neg).mean(axis=0)
v_bias_update = (v_pos - v_neg).mean(axis=0)
# Stochastic pseudo-likelihood
feature_index_to_flip = theano_random.random_integers(
low=0,
high=self.n_visible - 1,
)
rounded_input = T.round(network_input)
rounded_input = network_input
rounded_input_flip = T.set_subtensor(
rounded_input[:, feature_index_to_flip],
1 - rounded_input[:, feature_index_to_flip])
error = T.mean(self.n_visible * T.log(
T.nnet.sigmoid(
free_energy(rounded_input_flip) - free_energy(rounded_input))))
self.methods.update(train_epoch=theano.function(
[network_input],
error,
name='algo:rbm/func:train-epoch',
updates=[
(weight, weight + step * weight_update / n_samples),
(h_bias, h_bias + step * h_bias_update),
(v_bias, v_bias + step * v_bias_update),
(h_samples, asint(theano_random.binomial(n=1, p=h_neg))),
]),
prediction_error=theano.function(
[network_input],
error,
name='algo:rbm/func:prediction-error',
),
visible_to_hidden=theano.function(
[network_input],
visible_to_hidden(network_input),
name='algo:rbm/func:visible-to-hidden',
),
hidden_to_visible=theano.function(
[network_input],
hidden_to_visible(network_input),
name='algo:rbm/func:hidden-to-visible',
),
gibbs_sampling=theano.function(
[network_input],
sample_visible_from_hidden(
sample_hidden_from_visible(network_input)),
name='algo:rbm/func:gibbs-sampling',
))
def output(self, input_value):
return self.weight[asint(input_value)]
def init_methods(self):
def free_energy(visible_sample):
wx_b = T.dot(visible_sample, self.weight) + self.hidden_bias
visible_bias_term = T.dot(visible_sample, self.visible_bias)
hidden_term = T.log(asfloat(1) + T.exp(wx_b)).sum(axis=1)
return -visible_bias_term - hidden_term
def visible_to_hidden(visible_sample):
wx_b = T.dot(visible_sample, self.weight) + self.hidden_bias
return T.nnet.sigmoid(wx_b)
def hidden_to_visible(hidden_sample):
wx_b = T.dot(hidden_sample, self.weight.T) + self.visible_bias
return T.nnet.sigmoid(wx_b)
def sample_hidden_from_visible(visible_sample):
theano_random = self.theano_random
hidden_prob = visible_to_hidden(visible_sample)
hidden_sample = theano_random.binomial(n=1, p=hidden_prob,
dtype=theano.config.floatX)
return hidden_sample
def sample_visible_from_hidden(hidden_sample):
theano_random = self.theano_random
visible_prob = hidden_to_visible(hidden_sample)
visible_sample = theano_random.binomial(n=1, p=visible_prob,
dtype=theano.config.floatX)
return visible_sample
network_input = self.variables.network_input
n_samples = asfloat(network_input.shape[0])
theano_random = self.theano_random
weight = self.weight
h_bias = self.hidden_bias
v_bias = self.visible_bias
h_samples = self.variables.h_samples
step = asfloat(self.step)
sample_indeces = theano_random.random_integers(
low=0, high=n_samples - 1,
size=(self.batch_size,)
)
v_pos = ifelse(
T.eq(n_samples, self.batch_size),
network_input,
# In case if final batch has less number of
# samples then expected
network_input[sample_indeces]
)
h_pos = visible_to_hidden(v_pos)
v_neg = sample_visible_from_hidden(h_samples)
h_neg = visible_to_hidden(v_neg)
weight_update = v_pos.T.dot(h_pos) - v_neg.T.dot(h_neg)
h_bias_update = (h_pos - h_neg).mean(axis=0)
v_bias_update = (v_pos - v_neg).mean(axis=0)
# Stochastic pseudo-likelihood
feature_index_to_flip = theano_random.random_integers(
low=0,
high=self.n_visible - 1,
)
rounded_input = T.round(network_input)
rounded_input = network_input
rounded_input_flip = T.set_subtensor(
rounded_input[:, feature_index_to_flip],
1 - rounded_input[:, feature_index_to_flip]
)
error = T.mean(
self.n_visible * T.log(T.nnet.sigmoid(
free_energy(rounded_input_flip) -
free_energy(rounded_input)
))
)
self.methods.update(
train_epoch=theano.function(
[network_input],
error,
name='algo:rbm/func:train-epoch',
updates=[
(weight, weight + step * weight_update / n_samples),
(h_bias, h_bias + step * h_bias_update),
(v_bias, v_bias + step * v_bias_update),
(h_samples, asint(theano_random.binomial(n=1, p=h_neg))),
]
),
prediction_error=theano.function(
[network_input], error,
name='algo:rbm/func:prediction-error',
),
visible_to_hidden=theano.function(
[network_input],
visible_to_hidden(network_input),
name='algo:rbm/func:visible-to-hidden',
),
hidden_to_visible=theano.function(
[network_input],
hidden_to_visible(network_input),
name='algo:rbm/func:hidden-to-visible',
),
gibbs_sampling=theano.function(
[network_input],
sample_visible_from_hidden(
sample_hidden_from_visible(network_input)
),
name='algo:rbm/func:gibbs-sampling',
)
)
def loss_function(expected, predicted):
epsilon = 1e-7
log_predicted = T.log(T.clip(predicted, epsilon, 1.0 - epsilon))
errors = log_predicted[T.arange(expected.size), asint(expected.flatten())]
return -T.mean(errors)
def init_variables(self):
self.init_layers()
self.variables.update(h_samples=theano.shared(
name='h_samples',
value=asint(np.zeros((self.batch_size, self.n_hidden))),
), )
