def __init__(self,
incoming,
memory,
controller,
heads,
only_return_final=False,
**kwargs):
super(NTMLayer, self).__init__(incoming, **kwargs)
self.memory = memory
self.controller = controller
self.heads = heads
self.write_heads = WriteHeadCollection(heads=\
filter(lambda head: isinstance(head, WriteHead), heads))
self.read_heads = ReadHeadCollection(heads=\
filter(lambda head: isinstance(head, ReadHead), heads))
self.only_return_final = only_return_final
def __init__(self, incoming,
memory,
controller,
heads,
only_return_final=False,
**kwargs):
super(NTMLayer, self).__init__(incoming, **kwargs)
self.memory = memory
self.controller = controller
self.heads = heads
self.write_heads = WriteHeadCollection(heads=\
filter(lambda head: isinstance(head, WriteHead), heads))
self.read_heads = ReadHeadCollection(heads=\
filter(lambda head: isinstance(head, ReadHead), heads))
self.only_return_final = only_return_final
class NTMLayer(Layer):
r"""
A Neural Turing Machine layer.
Parameters
----------
incoming: a :class:`lasagne.layers.Layer` instance
The layer feeding into the Neural Turing Machine. This
layer must match the incoming layer in the controller.
memory: a :class:`Memory` instance
The memory of the NTM.
controller: a :class:`Controller` instance
The controller of the NTM.
heads: a list of :class:`Head` instances
The read and write heads of the NTM.
only_return_final: bool
If ``True``, only return the final sequential output (e.g.
for tasks where a single target value for the entire
sequence is desired). In this case, Theano makes an
optimization which saves memory.
"""
def __init__(self, incoming,
memory,
controller,
heads,
only_return_final=False,
**kwargs):
super(NTMLayer, self).__init__(incoming, **kwargs)
self.memory = memory
self.controller = controller
self.heads = heads
self.write_heads = WriteHeadCollection(heads=\
filter(lambda head: isinstance(head, WriteHead), heads))
self.read_heads = ReadHeadCollection(heads=\
filter(lambda head: isinstance(head, ReadHead), heads))
self.only_return_final = only_return_final
def get_output_shape_for(self, input_shapes):
if self.only_return_final:
return (input_shapes[0], self.controller.num_units)
else:
return (input_shapes[0], input_shapes[1], self.controller.num_units)
def get_params(self, **tags):
params = super(NTMLayer, self).get_params(**tags)
params += self.controller.get_params(**tags)
params += self.memory.get_params(**tags)
for head in self.heads:
params += head.get_params(**tags)
return params
def get_output_for(self, input, get_details=False, **kwargs):
input = input.dimshuffle(1, 0, 2)
def step(x_t, M_tm1, h_tm1, state_tm1, ww_tm1, wr_tm1, *params):
# Update the memory (using w_tm1 of the writing heads & M_tm1)
M_t = self.write_heads.write(h_tm1, ww_tm1, M_tm1)
# Get the read vector (using w_tm1 of the reading heads & M_t)
r_t = self.read_heads.read(wr_tm1, M_t)
# Apply the controller (using x_t, r_t & the requirements for the controller)
h_t, state_t = self.controller.step(x_t, r_t, h_tm1, state_tm1)
# Update the weights (using h_t, M_t & w_tm1)
ww_t = self.write_heads.get_weights(h_t, ww_tm1, M_t)
wr_t = self.read_heads.get_weights(h_t, wr_tm1, M_t)
return [M_t, h_t, state_t, ww_t, wr_t]
memory_init = T.tile(self.memory.memory_init, (input.shape[1], 1, 1))
memory_init = T.unbroadcast(memory_init, 0)
write_weights_init = T.tile(self.write_heads.weights_init, (input.shape[1], 1, 1))
write_weights_init = T.unbroadcast(write_weights_init, 0)
read_weights_init = T.tile(self.read_heads.weights_init, (input.shape[1], 1, 1))
read_weights_init = T.unbroadcast(read_weights_init, 0)
non_seqs = self.controller.get_params() + self.memory.get_params() + \
self.write_heads.get_params() + self.read_heads.get_params()
hids, _ = theano.scan(
fn=step,
sequences=input,
outputs_info=[memory_init] + self.controller.outputs_info(input.shape[1]) + \
[write_weights_init, read_weights_init],
non_sequences=non_seqs,
strict=True)
# dimshuffle back to (n_batch, n_time_steps, n_features)
if get_details:
hid_out = [
hids[0].dimshuffle(1, 0, 2, 3),
hids[1].dimshuffle(1, 0, 2),
hids[2].dimshuffle(1, 0, 2),
hids[3].dimshuffle(1, 0, 2, 3),
hids[4].dimshuffle(1, 0, 2, 3)]
else:
if self.only_return_final:
hid_out = hids[1][-1]
else:
hid_out = hids[1].dimshuffle(1, 0, 2)
return hid_out
class NTMLayer(Layer):
r"""
A Neural Turing Machine layer.
Parameters
----------
incoming: a :class:`lasagne.layers.Layer` instance
The layer feeding into the Neural Turing Machine. This
layer must match the incoming layer in the controller.
memory: a :class:`Memory` instance
The memory of the NTM.
controller: a :class:`Controller` instance
The controller of the NTM.
heads: a list of :class:`Head` instances
The read and write heads of the NTM.
only_return_final: bool
If ``True``, only return the final sequential output (e.g.
for tasks where a single target value for the entire
sequence is desired). In this case, Theano makes an
optimization which saves memory.
"""
def __init__(self,
incoming,
memory,
controller,
heads,
only_return_final=False,
**kwargs):
super(NTMLayer, self).__init__(incoming, **kwargs)
self.memory = memory
self.controller = controller
self.heads = heads
self.write_heads = WriteHeadCollection(heads=\
filter(lambda head: isinstance(head, WriteHead), heads))
self.read_heads = ReadHeadCollection(heads=\
filter(lambda head: isinstance(head, ReadHead), heads))
self.only_return_final = only_return_final
def get_output_shape_for(self, input_shapes):
if self.only_return_final:
return (input_shapes[0], self.controller.num_units)
else:
return (input_shapes[0], input_shapes[1],
self.controller.num_units)
def get_params(self, **tags):
params = super(NTMLayer, self).get_params(**tags)
params += self.controller.get_params(**tags)
params += self.memory.get_params(**tags)
for head in self.heads:
params += head.get_params(**tags)
return params
def get_output_for(self, input, get_details=False, **kwargs):
input = input.dimshuffle(1, 0, 2)
def step(x_t, M_tm1, h_tm1, state_tm1, ww_tm1, wr_tm1, *params):
# Update the memory (using w_tm1 of the writing heads & M_tm1)
M_t = self.write_heads.write(h_tm1, ww_tm1, M_tm1)
# Get the read vector (using w_tm1 of the reading heads & M_t)
r_t = self.read_heads.read(wr_tm1, M_t)
# Apply the controller (using x_t, r_t & the requirements for the controller)
h_t, state_t = self.controller.step(x_t, r_t, h_tm1, state_tm1)
# Update the weights (using h_t, M_t & w_tm1)
ww_t = self.write_heads.get_weights(h_t, ww_tm1, M_t)
wr_t = self.read_heads.get_weights(h_t, wr_tm1, M_t)
return [M_t, h_t, state_t, ww_t, wr_t]
memory_init = T.tile(self.memory.memory_init, (input.shape[1], 1, 1))
memory_init = T.unbroadcast(memory_init, 0)
write_weights_init = T.tile(self.write_heads.weights_init,
(input.shape[1], 1, 1))
write_weights_init = T.unbroadcast(write_weights_init, 0)
read_weights_init = T.tile(self.read_heads.weights_init,
(input.shape[1], 1, 1))
read_weights_init = T.unbroadcast(read_weights_init, 0)
non_seqs = self.controller.get_params() + self.memory.get_params() + \
self.write_heads.get_params() + self.read_heads.get_params()
hids, _ = theano.scan(
fn=step,
sequences=input,
outputs_info=[memory_init] + self.controller.outputs_info(input.shape[1]) + \
[write_weights_init, read_weights_init],
non_sequences=non_seqs,
strict=True)
# dimshuffle back to (n_batch, n_time_steps, n_features)
if get_details:
hid_out = [
hids[0].dimshuffle(1, 0, 2, 3), hids[1].dimshuffle(1, 0, 2),
hids[2].dimshuffle(1, 0, 2), hids[3].dimshuffle(1, 0, 2, 3),
hids[4].dimshuffle(1, 0, 2, 3)
]
else:
if self.only_return_final:
hid_out = hids[1][-1]
else:
hid_out = hids[1].dimshuffle(1, 0, 2)
return hid_out
