def _get_default_variable_store():
store = get_collection(_VARSTORE_KEY)
if store:
return store[0]
# create a new store
store = variable_store()
add_to_collection(_VARSTORE_KEY, store)
return store
def get_variable_scope():
# get_collection returns a list
scope = get_collection(_VARSCOPE_KEY)
if scope:
# only 1 element in the list
return scope[0]
# create a new scope
scope = var_scope(False)
add_to_collection(_VARSCOPE_KEY, scope)
return scope
def variable(initial_value=None,
trainable=True,
name=None,
dtype=theano.config.floatX):
global _TRAINABLE_VARIABLES
global _ALL_VARIABLES
if initial_value is None:
raise ValueError("initial_value must not be None")
if callable(initial_value):
val = initial_value()
else:
val = initial_value
var = theano.shared(val, name=name, borrow=True)
if trainable:
add_to_collection(_TRAINABLE_VARIABLES_KEY, var)
add_to_collection(_GLOBAL_VARIABLES_KEY, var)
return var
def add_regularization_loss(loss):
add_to_collection(_REGULARIZATION_LOSSES_KEYS, loss)
def scan(fn,
sequences=None,
outputs_info=None,
non_sequences=None,
return_updates=False,
use_extension=False,
**kwargs):
if sequences is None:
sequences = []
if outputs_info is None:
outputs_info = []
if non_sequences is None:
non_sequences = []
# support nested structure for sequences, outputs_info and non_sequences
if use_extension:
if isinstance(outputs_info, dict):
raise ValueError("only support nested structure, not dict")
nest_sequences = sequences
nest_outputs_info = outputs_info
nest_non_sequences = non_sequences
# inputs to Theano's scan
sequences = utils.flatten(nest_sequences)
outputs_info = utils.flatten(nest_outputs_info)
non_sequences = utils.flatten(nest_non_sequences)
# input structure for fn
nest_rec_info = []
for item in nest_outputs_info:
if item is not None:
nest_rec_info.append(item)
rec_info = utils.flatten(nest_rec_info)
n_seq = len(sequences)
n_rec = len(rec_info)
for item in rec_info:
if item is not None:
continue
raise ValueError("None can only appear in the outer level of "
"outputs_info")
inner_fn = fn
def wrapper_fn(*args):
seqs = args[:n_seq]
recs = args[n_seq:n_seq + n_rec]
nonseq = args[n_seq + n_rec:]
nest_seqs = utils.pack_sequence_as(nest_sequences, seqs)
nest_recs = utils.pack_sequence_as(nest_rec_info, recs)
nest_nonseq = utils.pack_sequence_as(nest_non_sequences, nonseq)
newargs = list(nest_seqs) + list(nest_recs) + list(nest_nonseq)
nest_outs = inner_fn(*newargs)
return utils.flatten(nest_outs)
fn = wrapper_fn
outputs, updates = theano.scan(fn, sequences, outputs_info, non_sequences,
**kwargs)
if use_extension:
outputs = utils.pack_sequence_as(nest_outputs_info, outputs)
if "key" not in kwargs or not kwargs["key"]:
key = "training"
else:
key = kwargs["key"]
if return_updates:
return outputs, updates
add_to_collection(_SCAN_UPDATES_KEYS + "/" + key, updates)
return outputs
