name='decoder_network')
decoder_p = decoder_network.apply(z)
# Define the cost
KL_term = -0.5 * (1 + encoder_lognu -T.exp(encoder_lognu) - encoder_mu ** 2).sum(axis=1)
reconstruction_term = (x * T.log(decoder_p) +
(1 - x) * T.log(1 - decoder_p)).sum(axis=1)
cost = (KL_term -reconstruction_term).mean()
cost.name = 'negative_log_likelihood'
# Initialize the parameters
encoder_network._push_initialization_config()
for layer in encoder_network.linear_transformations:
layer.weights_init = Uniform(
width=12. / (layer.input_dim + layer.output_dim))
encoder_network.initialize()
decoder_network._push_initialization_config()
for layer in decoder_network.linear_transformations:
layer.weights_init = Uniform(
width=12. / (layer.input_dim + layer.output_dim))
decoder_network.initialize()
mnist = BinarizedMNIST(("train",), sources=('features',))
train_data_stream = Flatten(DataStream.default_stream(
decoder_p = decoder_network3.apply(T.concatenate([h1_decode, h2_decode], axis=1))
# Define the cost
KL_term = -0.5 * ((1 + z1_lognu -T.exp(z1_lognu) - z1_mu ** 2).sum(axis=1)
+ (1 + z2_lognu -T.exp(z2_lognu) - z2_mu ** 2).sum(axis=1))
reconstruction_term = (x * T.log(decoder_p) +
(1 - x) * T.log(1 - decoder_p)).sum(axis=1)
cost = (KL_term -reconstruction_term).mean()
cost.name = 'negative_log_likelihood'
# Initialize the parameters
encoder_network1._push_initialization_config()
for layer in encoder_network1.linear_transformations:
layer.weights_init = Uniform(
width=12. / (layer.input_dim + layer.output_dim))
encoder_network1.initialize()
encoder_network2._push_initialization_config()
for layer in encoder_network2.linear_transformations:
layer.weights_init = Uniform(
width=12. / (layer.input_dim + layer.output_dim))
encoder_network2.initialize()
encoder_network3._push_initialization_config()
for layer in encoder_network3.linear_transformations:
layer.weights_init = Uniform(
width=12. / (layer.input_dim + layer.output_dim))
#y_hat =
#reconstruct1 = BinaryCrossEntropy().apply(x, mlp1.apply(x))
reconstruct2 = SquaredError().apply(y, y_hat)
regularizer = 0
#for mlp in [mlp1, mlp2, mlp3, mlp4]:
for child in mlp1.children:
for param in child.parameters:
regularizer = regularizer + (param ** 2).sum()
cost = reconstruct1 + 0.1*regularizer
cost.name = 'reconstruct_error'
#------------------------------------------------------------------
#for mlp in [mlp1, mlp2, mlp3, mlp4]:
mlp1._push_initialization_config()
for layer in mlp1.linear_transformations:
layer.weights_init = Uniform(width=12. / (layer.input_dim + layer.output_dim))
mlp1.initialize()
mnist = BinarizedMNIST(("train",), sources=('features',))
train_data_stream = Flatten(DataStream.default_stream(
mnist,
iteration_scheme=SequentialScheme(mnist.num_examples, batch_size=batch_size)))
monitor_data_stream = Flatten(DataStream.default_stream(
mnist,
iteration_scheme=SequentialScheme(mnist.num_examples, batch_size=batch_size)))
cg = ComputationGraph(cost)
