def __init__(self,
inpvocsize=400,
inpembdim=None,
outvocsize=100,
outembdim=None,
encdim=100,
decdim=100,
attdim=100,
bidir=False,
rnu=GRU,
statetrans=None,
vecout=False,
inconcat=True,
outconcat=False,
**kw):
encinnerdim = [encdim] if not issequence(encdim) else encdim
decinnerdim = [decdim] if not issequence(decdim) else decdim
self.enclayers, lastencinnerdim = \
self.getenclayers(inpembdim, inpvocsize, encinnerdim, bidir, rnu)
self.declayers = \
self.getdeclayers(outembdim, outvocsize, lastencinnerdim,
decinnerdim, rnu, inconcat)
# attention
lastdecinnerdim = decinnerdim[-1]
argdecinnerdim = lastdecinnerdim if outconcat is False else lastencinnerdim + lastdecinnerdim
attgen = LinearGateAttentionGenerator(indim=lastencinnerdim +
lastdecinnerdim,
attdim=attdim)
attcon = WeightedSumAttCon()
if statetrans is True:
if lastencinnerdim != lastdecinnerdim: # state shape mismatch
statetrans = MatDot(lastencinnerdim, lastdecinnerdim)
elif statetrans == "matdot":
statetrans = MatDot(lastencinnerdim, lastdecinnerdim)
super(SimpleSeqEncDecAtt, self).__init__(self.enclayers,
self.declayers,
attgen,
attcon,
argdecinnerdim,
statetrans=statetrans,
vecout=vecout,
inconcat=inconcat,
outconcat=outconcat,
**kw)
def remake_encoder(self,
inpvocsize=None,
inpembdim=None,
inpemb=None,
innerdim=None,
bidir=False,
maskid=-1,
dropout_in=False,
dropout_h=False,
rnu=GRU,
sepatt=False):
innerdim = ([innerdim] if not issequence(innerdim) else
innerdim) if innerdim is not None else self.encinnerdim
if sepatt:
enc = self._getencoder_sepatt(indim=inpvocsize,
inpembdim=inpembdim,
inpemb=inpemb,
innerdim=innerdim,
bidir=bidir,
maskid=maskid,
dropout_in=dropout_in,
dropout_h=dropout_h,
rnu=rnu)
else:
enc = self._getencoder(indim=inpvocsize,
inpembdim=inpembdim,
inpemb=inpemb,
innerdim=innerdim,
bidir=bidir,
maskid=maskid,
dropout_in=dropout_in,
dropout_h=dropout_h,
rnu=rnu)
self.statetrans = self._build_state_trans(self.statetrans_setting)
self.enc = enc
def __init__(self,
inpvocsize=400,
inpembdim=None,
outvocsize=100,
outembdim=None,
encdim=100,
decdim=100,
attdim=100,
bidir=False,
rnu=GRU,
statetrans=None,
vecout=False,
inconcat=True,
outconcat=False,
**kw):
encinnerdim = [encdim] if not issequence(encdim) else encdim
decinnerdim = [decdim] if not issequence(decdim) else decdim
self.enclayers, lastencinnerdim = \
self.getenclayers(inpembdim, inpvocsize, encinnerdim, bidir, rnu)
self.declayers = \
self.getdeclayers(outembdim, outvocsize, lastencinnerdim,
decinnerdim, rnu, inconcat)
# attention
lastdecinnerdim = decinnerdim[-1]
argdecinnerdim = lastdecinnerdim if outconcat is False else lastencinnerdim + lastdecinnerdim
attgen = LinearGateAttentionGenerator(indim=lastencinnerdim + lastdecinnerdim,
attdim=attdim)
attcon = WeightedSumAttCon()
if statetrans is True:
if lastencinnerdim != lastdecinnerdim: # state shape mismatch
statetrans = MatDot(lastencinnerdim, lastdecinnerdim)
elif statetrans == "matdot":
statetrans = MatDot(lastencinnerdim, lastdecinnerdim)
super(SimpleSeqEncDecAtt, self).__init__(self.enclayers, self.declayers,
attgen, attcon, argdecinnerdim, statetrans=statetrans, vecout=vecout,
inconcat=inconcat, outconcat=outconcat, **kw)
def get_init_info(self, inpseq, batsize, maskseq=None): # TODO: must evaluate enc here, in place, without any side effects
"""
VERY DIFFERENT FROM THE PURELY SYMBOLIC GET_INIT_INFO IN REAL REC BLOCKS !!!
This one is used in decoder/prediction
"""
enco, allenco, encmask = self.enc.predict(inpseq, mask=maskseq)
if self.statetrans is not None:
topstate = self.statetrans.predict(enco, allenco) # this gives unused input warning in theano - it's normal
initstates = [topstate]
else:
initstates = batsize
return self.dec.get_init_info(Val(allenco),
[Val(x) for x in initstates]
if issequence(initstates)
else initstates,
encmask=Val(encmask))
def get_init_info(
self,
inpseq,
batsize,
maskseq=None
): # TODO: must evaluate enc here, in place, without any side effects
"""
VERY DIFFERENT FROM THE PURELY SYMBOLIC GET_INIT_INFO IN REAL REC BLOCKS !!!
This one is used in decoder/prediction
"""
enco, allenco, encmask = self.enc.predict(inpseq, mask=maskseq)
if self.statetrans is not None:
topstate = self.statetrans.predict(
enco, allenco
) # this gives unused input warning in theano - it's normal
initstates = [topstate]
else:
initstates = batsize
return self.dec.get_init_info(
Val(allenco),
[Val(x)
for x in initstates] if issequence(initstates) else initstates,
encmask=Val(encmask))
def __init__(self,
inpvocsize=400,
inpembdim=None,
inpemb=None,
outvocsize=100,
outembdim=None,
outemb=None,
encdim=100,
decdim=100,
bidir=False,
rnu=GRU,
statetrans=None,
vecout=None,
inconcat=True,
outconcat=False,
maskid=-1,
dropout=False,
attdist=CosineDistance(),
sepatt=False,
encoder=None,
decoder=None,
attention=None,
**kw):
self.encinnerdim = [encdim] if not issequence(encdim) else encdim
self.decinnerdim = [decdim] if not issequence(decdim) else decdim
self.dropout = dropout
# encoder
if encoder is None:
if sepatt:
enc = self._getencoder_sepatt(indim=inpvocsize,
inpembdim=inpembdim,
inpemb=inpemb,
innerdim=self.encinnerdim,
bidir=bidir,
maskid=maskid,
dropout_in=dropout,
dropout_h=dropout,
rnu=rnu)
else:
enc = self._getencoder(indim=inpvocsize,
inpembdim=inpembdim,
inpemb=inpemb,
innerdim=self.encinnerdim,
bidir=bidir,
maskid=maskid,
dropout_in=dropout,
dropout_h=dropout,
rnu=rnu)
else:
enc = encoder
if attention is None:
attention = self._getattention(attdist, sepatt=sepatt)
self.lastencinnerdim = enc.outdim
if decoder is None:
dec = self._getdecoder(outvocsize=outvocsize,
outembdim=outembdim,
outemb=outemb,
maskid=maskid,
attention=attention,
lastencinnerdim=self.lastencinnerdim,
decinnerdim=self.decinnerdim,
inconcat=inconcat,
outconcat=outconcat,
softmaxout=vecout,
dropout=dropout,
rnu=rnu)
else:
dec = decoder
self.lastdecinnerdim = self.decinnerdim[-1]
self.statetrans_setting = statetrans
statetrans = self._build_state_trans(self.statetrans_setting)
super(SimpleSeqEncDecAtt, self).__init__(enc,
dec,
statetrans=statetrans,
**kw)
