def __init__(self,
layers,
softmaxoutblock=None,
innerdim=None,
attention=None,
inconcat=True,
outconcat=False,
dropout=False,
**kw):
super(SeqDecoder, self).__init__(**kw)
self.embedder = layers[0]
self.block = RecStack(*layers[1:])
self.outdim = innerdim
self.attention = attention
self.inconcat = inconcat
self.outconcat = outconcat
self._mask = False
self._attention = None
assert (isinstance(self.block, ReccableBlock))
if softmaxoutblock is None: # default softmax out block
sm = Softmax()
self.lin = Linear(indim=self.outdim,
dim=self.embedder.indim,
dropout=dropout)
self.softmaxoutblock = asblock(lambda x: sm(self.lin(x)))
elif softmaxoutblock is False:
self.softmaxoutblock = asblock(lambda x: x)
else:
self.softmaxoutblock = softmaxoutblock
def __init__(self, enc, dec, statetrans=None, **kw):
super(SeqEncDec, self).__init__(**kw)
self.enc = enc
self.dec = dec
if isinstance(statetrans, Block):
self.statetrans = asblock(lambda x, y: statetrans(x))
elif statetrans is True:
self.statetrans = asblock(lambda x, y: x)
else:
self.statetrans = statetrans
def __init__(self, enc, dec, statetrans=None, **kw):
super(SeqEncDec, self).__init__(**kw)
self.enc = enc
self.dec = dec
if isinstance(statetrans, Block):
self.statetrans = asblock(lambda x, y: statetrans(x))
elif statetrans is True:
self.statetrans = asblock(lambda x, y: x)
else:
self.statetrans = statetrans
def __init__(self,
lenc,
renc,
aggregator=asblock(lambda x: T.sum(x, axis=1)),
**kw):
self.agg = aggregator
super(SeqMatchScore, self).__init__(lenc, renc, **kw)
def innerf(encstates, encspec, decspec):
decspec = reduce(lambda x, y: list(x) + list(y), decspec, [])
encspec = reduce(lambda x, y: list(x) + list(y), encspec, [])
assert (len(decspec) == len(encspec))
ret = []
for i in range(len(encspec)):
if encspec[i][0] == "state" and decspec[i][0] == "state":
if decspec[i][1][0] != encspec[i][1][
0] or statetrans == "matdot":
t = MatDot(encspec[i][1][0], decspec[i][1][0])
else:
t = asblock(lambda x: x)
elif encspec[i][0] == decspec[i][0]:
t = None
else:
raise Exception()
ret.append(t)
assert (len(encstates) == len(ret))
out = []
for encstate, rete in zip(encstates, ret):
if rete is None:
out.append(None)
else:
out.append(rete(encstate))
return out
def setUp(self):
dim = 50
self.vocabsize = 2000
data = np.arange(0, self.vocabsize).astype("int32")
self.O = param((dim, self.vocabsize)).uniform()
self.W = VectorEmbed(indim=self.vocabsize, dim=50)
self.out = stack(self.W, asblock(lambda x: T.dot(self.O, x)),
Softmax())(Input(ndim=1, dtype="int32"))
def setUp(self):
dim=50
self.vocabsize=2000
data = np.arange(0, self.vocabsize).astype("int32")
self.O = param((dim, self.vocabsize)).uniform()
self.W = VectorEmbed(indim=self.vocabsize, dim=50)
self.out = stack(self.W,
asblock(lambda x: T.dot(self.O, x)),
Softmax())(Input(ndim=1, dtype="int32"))
def __init__(self, layers, softmaxoutblock=None, innerdim=None, attention=None, inconcat=False, outconcat=False, **kw): # limit says at most how many is produced
self.embedder = layers[0]
self.outdim = innerdim
self.inconcat = inconcat
self.outconcat = outconcat
self.attention = attention
super(SeqDecoder, self).__init__(*layers[1:], **kw) # puts layers into a ReccableBlock
self._mask = False
self._attention = None
assert(isinstance(self.block, ReccableBlock))
if softmaxoutblock is None: # default softmax out block
sm = Softmax()
self.lin = MatDot(indim=self.outdim, dim=self.embedder.indim)
self.softmaxoutblock = asblock(lambda x: sm(self.lin(x)))
else:
self.softmaxoutblock = softmaxoutblock
self.init_states = None
def __init__(self, lenc, renc,
aggregator=asblock(lambda x: T.sum(x, axis=1)), **kw):
self.agg = aggregator
super(SeqMatchScore, self).__init__(lenc, renc, **kw)
