def apply(self, x): # x: (batsize, seqlen)
emb = self.E[x] # (batsize, seqlen, inpembdim)
outs = T.tensordot(emb, self.W, 1) # (batsize, seqlen, outdim)
outsf = outs.reshape((outs.shape[0] * outs.shape[1], outs.shape[2])) # (batsize*seqlen, outdim)
outsfsm = Softmax()(outsf)
ret = outsfsm.reshape(outs.shape) # (batsize, seqlen, outdim)
return ret
def apply(self, x): # x: (batsize, seqlen)
emb = self.E[x] # (batsize, seqlen, inpembdim)
outs = T.tensordot(emb, self.W, 1) # (batsize, seqlen, outdim)
outsf = outs.reshape((outs.shape[0] * outs.shape[1],
outs.shape[2])) # (batsize*seqlen, outdim)
outsfsm = Softmax()(outsf)
ret = outsfsm.reshape(outs.shape) # (batsize, seqlen, outdim)
return ret
def __init__(self, memblock, memaddr, memattdim=100, indim=None, outnorm=Softmax(), **kw):
super(MemoryStack, self).__init__(**kw)
if not isinstance(memblock, MemoryBlock):
raise Exception("must provide a loaded memory block")
memdim = memblock.outdim
indim = memdim if indim is None else indim
self.exe = stack(memaddr(memblock, memdim=memdim, indim=indim, attdim=memattdim), outnorm)
def __init__(self, encdim, invocsize, outvocsize, innerdim, seqlen, **kw):
super(idx2seqStupid, self).__init__(**kw)
self.encdim = encdim
self.invocsize = invocsize
self.outvocsize = outvocsize
self.innerdim = innerdim
self.seqlen = seqlen
self.emb = VectorEmbed(indim=self.invocsize,
dim=self.encdim,
normalize=True)
self.aletter = stack(Lin(indim=self.encdim, dim=self.outvocsize),
Softmax())
self.bletter = stack(Lin(indim=self.encdim, dim=self.outvocsize),
Softmax())
self.cletter = stack(Lin(indim=self.encdim, dim=self.outvocsize),
Softmax())
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 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, numin, *dims, **kw):
super(Model, self).__init__(**kw)
self.layers = []
dims = list(dims)
dims = [numin] + dims
for i in range(1, len(dims)):
self.layers.append(Linear(indim=dims[i - 1], dim=dims[i]))
self.layers.append(Tanh())
self.layers[-1] = Softmax()
def apply(self, x):
emb = self.recembed(x) if self.scanemb else self.E[
x] # (batsize, seqlen, inpembdim)
outs = self.recout(emb) if self.scanout else T.tensordot(
emb, self.W, 1)
ret = self.recret(outs) if self.scansm else Softmax()(outs.reshape(
(outs.shape[0] * outs.shape[1],
outs.shape[2]))).reshape(outs.shape) # (batsize*seqlen, outdim)
return ret
def __init__(self, outlayers, **kw):
super(Vec2Idx, self).__init__(**kw)
if isinstance(outlayers, MemoryStack):
out = outlayers
else:
if not issequence(outlayers):
outlayers = [outlayers]
if type(outlayers[-1]) is not Softmax:
outlayers.append(Softmax())
out = stack(*outlayers)
self.out = out
def apply(
self,
criterion,
data,
mask=None
): # criterion: (batsize, indim), data: (batsize, seqlen, memdim)
o = self.getscores(criterion, data) # (batsize, seqlen)
o = Softmax()(o)
if mask is not None: # {0,1}^(batsize, seqlen)
o = mask * o
return o
def __init__(self, embedder, *layers, **kw):
""" layers must have an embedding layers first, final softmax layer is added automatically"""
assert ("smodim" in kw and "outdim" in kw)
self.embedder = embedder
smodim = kw["smodim"]
outdim = kw["outdim"]
del kw["smodim"]
del kw["outdim"]
super(SeqTransducer, self).__init__(**kw)
self.block = RecStack(*(layers +
(Lin(indim=smodim, dim=outdim), Softmax())))
def __init__(self, *layers, **kw):
""" first two layers must be embedding layers. Final softmax is added automatically"""
assert ("smodim" in kw and "outdim" in kw)
smodim = kw["smodim"]
outdim = kw["outdim"]
del kw["smodim"]
del kw["outdim"]
super(SeqTransDec, self).__init__(**kw)
self.inpemb = layers[0]
self.outemb = layers[1]
self.block = RecStack(*(layers[2:] +
(Lin(indim=smodim, dim=outdim), Softmax())))
def apply(self, x):
E = self.E
W = self.W
sm = Softmax()
def rec(x_t):
emb = E[x_t]
outs = T.dot(emb, W)
return sm(outs)
o, _ = T.scan(fn=rec, sequences=x.dimshuffle(1, 0), outputs_info=None)
return o.dimshuffle(1, 0, 2)
def apply(self, x, mask=None, weights=None):
ret = self.enc(x, mask=mask,
weights=weights) # (batsize, seqlen, lastdim)
outs = []
# apply mask (SeqEncoder should attach mask to outvar if all_outputs()
mask = ret.mask
for i in range(self.numouts):
selfweights = Softmax()(ret[:, :, i]) # (batsize, seqlen)
selfweights *= mask # apply mask
selfweights = selfweights / T.sum(selfweights, axis=1).dimshuffle(
0, "x") # renormalize
weightedstates = ret[:, :, self.numouts:] * selfweights.dimshuffle(
0, 1, "x")
out = T.sum(weightedstates, axis=1) # (batsize, lastdim)
outs.append(out)
if self.mode == "concat":
ret = T.concatenate(outs, axis=1)
elif self.mode == "seq":
outs = [out.dimshuffle(0, "x", 1) for out in outs]
ret = T.concatenate(outs, axis=1)
return ret
def apply(self, x): # x: (batsize, 4)
o, _ = T.scan(fn=self.rec,
sequences=[x[:, 0], x[:, 1], x[:, 2]],
non_sequences=[self.xes, self.yes, self.divmul],
outputs_info=None) # (batsize, outdim)
#axes = T.tile(x[:, 0], (self.xes.shape[0], 1)).T
#ayes = T.tile(x[:, 1], (self.xes.shape[0], 1)).T
#adivs = T.tile(x[:, 2], (self.xes.shape[0], 1)).T
#bxes = T.tile(self.xes, (x.shape[0], 1))
#byes = T.tile(self.yes, (x.shape[0], 1))
#o = self.rec(axes, ayes, adivs, bxes, byes, self.divmul)
ret = Softmax()(o)
return ret
def init(self):
#MEMORY: encodes how entity is written + custom entity embeddings
wencpg = WordEncoderPlusGlove(numchars=self.numchars,
numwords=self.numwords,
encdim=self.wordencdim,
embdim=self.wordembdim,
embtrainfrac=0.0,
glovepath=self.glovepath)
self.memenco = SeqEncoder(
wencpg,
GRU(dim=self.wordembdim + self.wordencdim,
innerdim=self.encinnerdim))
entemb = VectorEmbed(indim=self.outdim, dim=self.entembdim)
self.mempayload = ConcatBlock(entemb, self.memenco)
self.memblock = MemoryBlock(self.mempayload,
self.memdata,
indim=self.outdim,
outdim=self.encinnerdim + self.entembdim)
#ENCODER: uses the same language encoder as memory
#wencpg2 = WordEncoderPlusGlove(numchars=self.numchars, numwords=self.numwords, encdim=self.wordencdim, embdim=self.wordembdim, embtrainfrac=0.0, glovepath=glovepath)
self.enc = RecStack(
wencpg,
GRU(dim=self.wordembdim + self.wordencdim,
innerdim=self.encinnerdim))
#ATTENTION
attgen = LinearGateAttentionGenerator(indim=self.encinnerdim +
self.decinnerdim,
innerdim=self.attdim)
attcon = WeightedSumAttCon()
#DECODER
#entemb2 = VectorEmbed(indim=self.outdim, dim=self.entembdim)
self.softmaxoutblock = stack(
self.memaddr(self.memblock,
indim=self.decinnerdim + self.encinnerdim,
memdim=self.memblock.outdim,
attdim=self.attdim), Softmax())
self.dec = SeqDecoder([
self.memblock,
GRU(dim=self.entembdim + self.encinnerdim,
innerdim=self.decinnerdim)
],
outconcat=True,
inconcat=False,
attention=Attention(attgen, attcon),
innerdim=self.decinnerdim + self.encinnerdim,
softmaxoutblock=self.softmaxoutblock)
def init(self):
#memory
wencpg = WordEncoderPlusGlove(numchars=self.numchars,
numwords=self.numwords,
encdim=self.wordencdim,
embdim=self.wordembdim,
embtrainfrac=0.0,
glovepath=self.glovepath)
self.memenco = SeqEncoder(
wencpg,
GRU(dim=self.wordembdim + self.wordencdim,
innerdim=self.encinnerdim))
entemb = VectorEmbed(indim=self.outdim, dim=self.entembdim)
self.mempayload = ConcatBlock(entemb, self.memenco)
self.memblock = MemoryBlock(self.mempayload,
self.memdata,
indim=self.outdim,
outdim=self.encinnerdim + self.entembdim)
#encoder
wencpg2 = WordEncoderPlusGlove(numchars=self.numchars,
numwords=self.numwords,
encdim=self.wordencdim,
embdim=self.wordembdim,
embtrainfrac=0.0,
glovepath=self.glovepath)
self.enc = SeqEncoder(
wencpg2,
GRU(dim=self.wordembdim + self.wordencdim,
innerdim=self.encinnerdim))
#decoder
entemb2 = VectorEmbed(indim=self.outdim, dim=self.entembdim)
self.softmaxoutblock = stack(
self.memaddr(self.memblock,
indim=self.decinnerdim,
memdim=self.memblock.outdim,
attdim=self.attdim), Softmax())
self.dec = SeqDecoder(
[
entemb2, #self.memblock,
GRU(
dim=entemb.dim + self.encinnerdim,
innerdim=self.decinnerdim
), # GRU(dim=self.memblock.outdim + self.encinnerdim, innerdim=self.decinnerdim),
],
inconcat=True,
innerdim=self.decinnerdim,
softmaxoutblock=self.softmaxoutblock)
def apply(self, x, mask=None, weights=None):
ret = self.enc(x, mask=mask,
weights=weights) # (batsize, seqlen, lastdim)
outs = []
# apply mask (SeqEncoder should attach mask to outvar if all_outputs()
mask = mask if mask is not None else ret.mask if hasattr(
ret, "mask") else None
if self.bidir:
mid = ret.shape[2] / 2
ret1 = ret[:, :, :mid]
ret2 = ret[:, :, mid:]
ret = ret1
for i in range(self.numouts):
selfweights = ret[:, :, i] # (batsize, seqlen)
if self.bidir:
selfweights += ret2[:, :, i]
selfweights = Softmax()(selfweights)
if mask is not None:
selfweights *= mask # apply mask
selfweights = selfweights / T.sum(selfweights, axis=1).dimshuffle(
0, "x") # renormalize
weightedstates = ret[:, :, self.numouts:] * selfweights.dimshuffle(
0, 1, "x")
if self.bidir:
weightedstates2 = ret2[:, :,
self.numouts:] * selfweights.dimshuffle(
0, 1, "x")
weightedstates = T.concatenate(
[weightedstates, weightedstates2], axis=2)
out = T.sum(weightedstates, axis=1) # (batsize, lastdim)
outs.append(out)
if self.mode == "concat":
ret = T.concatenate(outs, axis=1)
elif self.mode == "seq":
outs = [out.dimshuffle(0, "x", 1) for out in outs]
ret = T.concatenate(outs, axis=1)
return ret
def __init__(self,
indim=400,
embdim=50,
inpemb=None,
innerdim=100,
outdim=50,
rnu=GRU,
**kw):
if inpemb is None:
emb = VectorEmbed(indim=indim, dim=embdim)
else:
emb = inpemb
embdim = emb.outdim
if not issequence(innerdim):
innerdim = [innerdim]
innerdim = [embdim] + innerdim
rnn, _ = MakeRNU.fromdims(innerdim, rnu=rnu)
smo = Lin(indim=innerdim[-1], dim=outdim)
super(SimpleSeqTrans, self).__init__(emb, *(rnn + [smo, Softmax()]),
**kw)
def apply(self, inp):
enco = self.enc(inp)
ret = Softmax()(self.out(enco))
return ret
def apply(self, x): # (batsize,)
m = self.W_m[x]
s = self.W_s[x]
z = RVal(seed=self.seed).normal(m.shape) * s + m
o = T.dot(z, self.O)
return Softmax()(o)
def apply(self, inpseq): # (batsize, amwords, amchars+1)
inpenc = self.phraseencoder(inpseq) # (batsize, encdim)
#scores = T.dot(inpenc, self.W) # (batsize, memdim)
#scores = T.nnet.sigmoid(T.dot(scores, self.memblock.innervar.T)) # (batsize, memsize)
#return Softmax()(scores) #
return Softmax()(self.mema(inpenc))
def apply(self, inptensor):
emb = self.W(inptensor)
out = T.dot(emb, self.O)
out.output_as("out")
probs = Softmax()(out)
return probs
def __init__(self, distance, normalizer=Softmax(), **kw):
super(AttGen, self).__init__(**kw)
self.dist = distance
self.normalizer = normalizer
def apply(self, seqs):
return Softmax()(self.outlin(self.enc(seqs)))
