def build_recurrent_graph(self, n_steps=None, reverse=False, **kwargs):
self.nonseq_args = kwargs.pop('nonseq_args', None)
self.output_args = kwargs.pop('output_args', None)
self.context_args = kwargs.pop('context_args', None)
self.iterators = kwargs.pop('iterators', None)
self.nonseq_inputs = kwargs.pop('nonseq_inputs', None)
self.nNone = 0
inputs = self.inputs.values()
seqs = []
outputs = []
nonseqs = []
if self.nonseq_inputs is not None:
for i in self.nonseq_inputs:
nonseqs.append(inputs.pop(i))
self.input_args = OrderedDict()
self.recur_args = OrderedDict()
for name, node in self.nodes.items():
if hasattr(node, 'get_init_state'):
self.recur_args[name] = node
state = node.get_init_state()
outputs.append(state)
self.nrecur = len(self.recur_args)
# Substitutes initial hidden state into a context
if self.context_args is not None:
for i, (nname, node) in enumerate(self.output_args.items()):
for aname, arg in self.context_args.items():
if nname == aname:
outputs[i] = arg
if self.iterators is None:
seqs += inputs
elif self.iterators is not None:
seqs += inputs[len(self.iterators):]
outputs += inputs[:len(self.iterators)]
if self.output_args is not None:
self.nNone = len(self.output_args)
outputs = flatten(outputs + [None] * self.nNone)
if self.nonseq_args is not None:
for arg in self.nonseq_args:
nonseqs.append(arg)
self.nseqs = len(seqs)
self.noutputs = len(outputs)
self.nnonseqs = len(nonseqs)
result, updates = theano.scan(
fn=self.scan_fn,
sequences=seqs,
outputs_info=outputs,
non_sequences=nonseqs,
n_steps=n_steps,
go_backwards=reverse)
result = tolist(result)
if self.output_args is None and self.iterators is None:
return result
if len(updates) == 0:
return result[-self.nNone:]
for k, v in updates.iteritems():
k.default_update = v
return result[-self.nNone:], updates
def build_recurrent_graph(self, n_steps=None, reverse=False, **kwargs):
self.nonseq_args = kwargs.pop('nonseq_args', None)
self.output_args = kwargs.pop('output_args', None)
self.context_args = kwargs.pop('context_args', None)
self.iterators = kwargs.pop('iterators', None)
self.nonseq_inputs = kwargs.pop('nonseq_inputs', None)
self.nNone = 0
inputs = self.inputs.values()
seqs = []
outputs = []
nonseqs = []
if self.nonseq_inputs is not None:
for i in self.nonseq_inputs:
nonseqs.append(inputs.pop(i))
self.input_args = OrderedDict()
self.recur_args = OrderedDict()
for name, node in self.nodes.items():
if hasattr(node, 'get_init_state'):
self.recur_args[name] = node
state = node.get_init_state()
outputs.append(state)
self.nrecur = len(self.recur_args)
# Substitutes initial hidden state into a context
if self.context_args is not None:
for i, (nname, node) in enumerate(self.output_args.items()):
for aname, arg in self.context_args.items():
if nname == aname:
outputs[i] = arg
if self.iterators is None:
seqs += inputs
elif self.iterators is not None:
seqs += inputs[len(self.iterators):]
outputs += inputs[:len(self.iterators)]
if self.output_args is not None:
self.nNone = len(self.output_args)
outputs = flatten(outputs + [None] * self.nNone)
if self.nonseq_args is not None:
for arg in self.nonseq_args:
nonseqs.append(arg)
self.nseqs = len(seqs)
self.noutputs = len(outputs)
self.nnonseqs = len(nonseqs)
result, updates = theano.scan(
fn=self.scan_fn,
sequences=seqs,
outputs_info=outputs,
non_sequences=nonseqs,
n_steps=n_steps,
go_backwards=reverse)
result = tolist(result)
if self.output_args is None and self.iterators is None:
return result
if len(updates) == 0:
return result[-self.nNone:]
for k, v in updates.iteritems():
k.default_update = v
return result[-self.nNone:], updates
init_b=init_b)
output = FullyConnectedLayer(name='output',
parent=['h1', 'h2', 'h3'],
parent_dim=[200, 200, 200],
nout=205,
unit='softmax',
init_W=init_W,
init_b=init_b)
nodes = [h1, h2, h3, output]
for node in nodes:
node.initialize()
params = flatten([node.get_params().values() for node in nodes])
step_count = sharedX(0, name='step_count')
last_h = np.zeros((batch_size, 400), dtype=np.float32)
h1_tm1 = sharedX(last_h, name='h1_tm1')
h2_tm1 = sharedX(last_h, name='h2_tm1')
h3_tm1 = sharedX(last_h, name='h3_tm1')
update_list = [step_count, h1_tm1, h2_tm1, h3_tm1]
step_count = T.switch(T.le(step_count, reset_freq),
step_count + 1, 0)
s1_0 = T.switch(T.or_(T.cast(T.eq(step_count, 0), 'int32'),
T.cast(T.eq(T.sum(h1_tm1), 0.), 'int32')),
h1.get_init_state(), h1_tm1)
s2_0 = T.switch(T.or_(T.cast(T.eq(step_count, 0), 'int32'),
init_b=init_b)
output = FullyConnectedLayer(name='output',
parent=['h1', 'h2', 'h3'],
parent_dim=[200, 200, 200],
nout=frame_size,
unit='sigmoid',
init_W=init_W,
init_b=init_b)
nodes = [h1, h2, h3, output]
for node in nodes:
node.initialize()
params = flatten([node.get_params().values() for node in nodes])
s1_0 = h1.get_init_state(batch_size)
s2_0 = h2.get_init_state(batch_size)
s3_0 = h3.get_init_state(batch_size)
def inner_fn(x_t, s1_tm1, s2_tm1, s3_tm1):
s1_t = h1.fprop([[x_t], [s1_tm1]])
s2_t = h2.fprop([[s1_t], [s2_tm1]])
s3_t = h3.fprop([[s2_t], [s3_tm1]])
y_hat_t = output.fprop([s1_t, s2_t, s3_t])
return s1_t, s2_t, s3_t, y_hat_t
def get_params(self):
return flatten([node.get_params().values()
for node in self.nodes.values()])
def get_params(self):
return flatten([node.get_params().values()
for node in self.nodes.values()])
