class DNCOutputLayer():
def __init__(self, layers_desc, name='linear_output'):
self.layers_desc = layers_desc
self.ffnn = FeedForwardNN(layers_desc, name)
'''
input:
layers_desc = Description of each layer (LayerDescription object type )
error_func = Error function, must has this input:output, expected_output
opt_func = Optimization function. Must have next input: error
'''
def run_output_layer(self, inputs):
if len(inputs) != 2:
raise ValueError(
'For DNC output layer inputs must be of length 2. Header and read vectors.'
)
output_vector = self.ffnn.run_feed_forward_nn(inputs[0])
output_layer = None
for layer in self.layers_desc:
if layer.__class__ is OutputLayerDescription:
output_layer = layer
break
if not output_layer:
raise Exception(
'Not OutputLayerDescription was found in layer description.')
return tf.add(
output_vector,
Utility._linear(inputs[1], output_layer.size,
output_layer.has_bias))
class InterfaceParameters():
def __init__(self,
key_w_layers_desc,
intensity_w_layers_desc,
erase_layers_desc,
add_layers_desc,
allocation_layers_desc,
write_gate_layers_desc,
free_gate_layers_desc,
key_r_layers_desc,
intensity_r_layers_desc,
read_mode_layers_desc,
name='interface_parameters'):
self.key_w_nn = FeedForwardNN(key_w_layers_desc, name)
self.intensity_w_nn = FeedForwardNN(intensity_w_layers_desc, name)
self.add_nn = FeedForwardNN(add_layers_desc, name)
self.erase_nn = FeedForwardNN(erase_layers_desc, name)
self.allocation_nn = FeedForwardNN(allocation_layers_desc, name)
self.write_gate_nn = FeedForwardNN(write_gate_layers_desc, name)
self.free_gate_nn = FeedForwardNN(free_gate_layers_desc, name)
self.key_r_nn = FeedForwardNN(key_r_layers_desc, name)
self.intensity_r_nn = FeedForwardNN(intensity_r_layers_desc, name)
self.read_mode_nn = FeedForwardNN(read_mode_layers_desc, name)
def run_generation(self, inputs):
key = self.key_w_nn.run_feed_forward_nn(inputs)
intensity = self.intensity_w_nn.run_feed_forward_nn(inputs)
add = self.add_nn.run_feed_forward_nn(inputs)
erase = self.erase_nn.run_feed_forward_nn(inputs)
allocation = self.allocation_nn.run_feed_forward_nn(inputs)
write_gate = self.write_gate_nn.run_feed_forward_nn(inputs)
free_gates = self.free_gate_nn.run_feed_forward_nn(inputs)
keys = self.key_r_nn.run_feed_forward_nn(inputs)
intensities = self.intensity_r_nn.run_feed_forward_nn(inputs)
read_mode = self.read_mode_nn.run_feed_forward_nn(inputs)
return key, intensity, add, erase, allocation, write_gate, free_gates, keys, intensities, read_mode
class FeedForwardController():
def __init__(self, layers_desc, name='ffnn_controller'):
self.ffnn = FeedForwardNN(layers_desc, name)
'''
input:
layers_desc = Description of each layer (LayerDescription object type )
error_func = Error function, must has this input:output, expected_output
opt_func = Optimization function. Must have next input: error
'''
def run_controller(self, inputs):
return self.ffnn.run_feed_forward_nn(inputs)
def __init__(self, layers_desc, name='ffnn_controller'):
self.ffnn = FeedForwardNN(layers_desc, name)
def __init__(self, layers_desc, name='linear_output'):
self.layers_desc = layers_desc
self.ffnn = FeedForwardNN(layers_desc, name)
'''
def __init__(self,
key_w_layers_desc,
intensity_w_layers_desc,
erase_layers_desc,
add_layers_desc,
allocation_layers_desc,
write_gate_layers_desc,
free_gate_layers_desc,
key_r_layers_desc,
intensity_r_layers_desc,
read_mode_layers_desc,
name='interface_parameters'):
self.key_w_nn = FeedForwardNN(key_w_layers_desc, name)
self.intensity_w_nn = FeedForwardNN(intensity_w_layers_desc, name)
self.add_nn = FeedForwardNN(add_layers_desc, name)
self.erase_nn = FeedForwardNN(erase_layers_desc, name)
self.allocation_nn = FeedForwardNN(allocation_layers_desc, name)
self.write_gate_nn = FeedForwardNN(write_gate_layers_desc, name)
self.free_gate_nn = FeedForwardNN(free_gate_layers_desc, name)
self.key_r_nn = FeedForwardNN(key_r_layers_desc, name)
self.intensity_r_nn = FeedForwardNN(intensity_r_layers_desc, name)
self.read_mode_nn = FeedForwardNN(read_mode_layers_desc, name)
def __init__(self, layers_desc, name='sigmoid_output'):
self.ffnn = FeedForwardNN(layers_desc, name)