def __init__(self, D, hidden_layer_sizes, loss_fn='sigmoid_cross_entropy'):
self.hidden_layer_sizes = hidden_layer_sizes
# input batch of training data (batch_size x D):
self.X = tf.placeholder(tf.float32, shape=(None, D))
# create hidden layers:
self.hidden_layers = []
M1 = D
for i, M2 in enumerate(self.hidden_layer_sizes):
h = HiddenLayer(M1, M2, i)
self.hidden_layers.append(h)
M1 = M2
# hidden --> output layer parameters:
Wo, bo = init_weights_and_biases(M2, D)
self.Wo = tf.Variable(Wo, name='Wo') # (M2 x D)
self.bo = tf.Variable(bo, name='bo') # D
# collect all network parameters:
self.params = []
for h in self.hidden_layers:
self.params += h.params
self.params += [self.Wo, self.bo]
# print('self.params:', self.params)
# get output - our reconstruction:
logits = self.forward(self.X)
if loss_fn == 'sigmoid_cross_entropy':
# assuming inputs and outputs to be Bernoulli probabilities
self.output = tf.nn.sigmoid(logits)
# define the cost function:
self.cost = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
labels=self.X,
logits=logits
)
)
elif loss_fn == 'mse':
# assuming the difference (error) between inputs and outputs to be Gaussian
self.output = tf.nn.sigmoid(logits) # assuming output is in range [0, 1]
# self.output = logits
self.cost = tf.reduce_mean(
tf.losses.mean_squared_error(
labels=self.X,
predictions=self.output
)
)
# define session:
self.sess = tf.InteractiveSession()
def __init__(self, D, hidden_layer_sizes, loss_fn='sigmoid_cross-entropy'):
self.hidden_layer_sizes = hidden_layer_sizes
# input batch of training data (batch_size x D):
self.X = T.matrix('X')
# create hidden layers:
self.hidden_layers = []
M1 = D
for i, M2 in enumerate(self.hidden_layer_sizes):
h = HiddenLayer(M1, M2, i)
self.hidden_layers.append(h)
M1 = M2
# hidden --> output layer parameters:
Wo, bo = init_weights_and_biases(M2, D)
self.Wo = theano.shared(Wo, name='Wo') # (M2 x D)
self.bo = theano.shared(bo, name='bo') # D
# collect all network parameters:
self.params = []
for h in self.hidden_layers:
self.params += h.params
self.params += [self.Wo, self.bo]
# print('self.params:', self.params)
# get output - our reconstruction:
self.output = self.forward(self.X)
if loss_fn == 'sigmoid_cross-entropy':
# assuming inputs and outputs to be Bernoulli probabilities
# define the cost function:
self.output = T.nnet.sigmoid(self.output)
# self.cost = -T.mean(self.X*T.log(self.output) + (1 - self.X)*T.log(1 - self.output))
self.cost = T.mean(
T.nnet.binary_crossentropy(output=self.output, target=self.X))
elif loss_fn == 'mse':
# assuming the difference (error) between inputs and outputs to be Gaussian
self.output = T.nnet.sigmoid(
self.output) # assuming output is in range [0, 1]
self.cost = T.mean((self.X - self.output)**2)
self.predict = theano.function(inputs=[self.X], outputs=self.output)
def __init__(self, M1, M2, an_id, f=T.nnet.relu):
W0, b0 = init_weights_and_biases(M1, M2)
self.W = theano.shared(W0, name='W%s' % an_id)
self.b = theano.shared(b0, name='b%s' % an_id)
self.f = f
self.params = [self.W, self.b]
def __init__(self, M1, M2, an_id): W0, b0 = init_weights_and_biases(M1, M2) self.W = tf.Variable(W0, name='W%s'%an_id) self.b = tf.Variable(b0, name='b%s'%an_id) self.params = [self.W, self.b]
def __init__(self, M1, M2, an_id, f=tf.nn.relu):
W0, b0 = init_weights_and_biases(M1, M2)
self.W = tf.Variable(W0, name="W%s" % an_id)
self.b = tf.Variable(b0, name="b%s" % an_id)
self.f = f
self.params = [self.W, self.b]
def __init__(self, M1, M2, an_id, f=tf.nn.relu):
W0, b0 = init_weights_and_biases(M1, M2)
self.W = tf.Variable(W0, name='W%s' % an_id)
self.b = tf.Variable(b0, name='b%s' % an_id)
self.f = f