def test_output_shape(self):
batsize = 100
wordembdim = 50
wordencdim = 20
innerdim = 40
datanuments = 77
vocnumwords = 100
numchars = 10
wseqlen = 3
cseqlen = 5
m = FBBasicCompositeEncoder(
wordembdim=wordembdim,
wordencdim=wordencdim,
innerdim=innerdim,
outdim=datanuments,
numchars=numchars,
numwords=vocnumwords,
glovepath="../../../data/glove/miniglove.%dd.txt",
)
worddata = np.random.randint(0, vocnumwords, (batsize, wseqlen, 1))
chardata = np.random.randint(0, numchars, (batsize, wseqlen, cseqlen))
data = np.concatenate([worddata, chardata], axis=2)
predshape = m.predict(data).shape
self.assertEqual(predshape, (batsize, datanuments))
def test_output_shape(self):
batsize = 100
wordembdim = 50
wordencdim = 20
innerdim = 40
datanuments = 77
vocnumwords = 100
numchars = 10
wseqlen = 3
cseqlen = 5
m = FBBasicCompositeEncoder(
wordembdim=wordembdim,
wordencdim=wordencdim,
innerdim=innerdim,
outdim=datanuments,
numchars=numchars,
numwords=vocnumwords,
glovepath="../../../data/glove/miniglove.%dd.txt",
)
worddata = np.random.randint(0, vocnumwords, (batsize, wseqlen, 1))
chardata = np.random.randint(0, numchars, (batsize, wseqlen, cseqlen))
data = np.concatenate([worddata, chardata], axis=2)
predshape = m.predict(data).shape
self.assertEqual(predshape, (batsize, datanuments))
def run(
epochs=100,
lr=0.5,
wreg=0.0001,
numbats=100,
fblexpath="../../data/freebase/labelsrevlex.map.sample",
glovepath="../../data/glove/glove.6B.50d.txt",
fbentdicp="../../data/freebase/entdic.all.map",
numwords=10,
numchars=30,
wordembdim=50,
wordencdim=100,
innerdim=300,
wordoffset=1,
validinter=3,
gradnorm=1.0,
validsplit=100,
):
tt = ticktock("fblextransrun")
traindata, golddata, vocnuments, vocnumwords, datanuments = \
loadlexdata(glovepath, fbentdicp, fblexpath, wordoffset, numwords, numchars)
tt.tock("made data").tick()
# define model
m = FBBasicCompositeEncoder(
wordembdim=wordembdim,
wordencdim=wordencdim,
innerdim=innerdim,
outdim=datanuments,
numchars=128, # ASCII
numwords=vocnumwords,
)
#wenc = WordEncoderPlusGlove(numchars=numchars, numwords=vocnumwords, encdim=wordencdim, embdim=wordembdim)
# train model TODO
tt.tick("training")
m.train([traindata], golddata).adagrad(lr=lr).grad_total_norm(gradnorm).cross_entropy()\
.autovalidate(splits=validsplit, random=True).validinter(validinter).accuracy()\
.train(numbats, epochs)
#embed()
tt.tock("trained").tick("predicting")
print m.predict(traindata).shape
tt.tock("predicted sample")
def run(
epochs=100,
lr=0.5,
wreg=0.0001,
numbats=100,
fblexpath="../../data/freebase/labelsrevlex.map.sample",
glovepath="../../data/glove/glove.6B.50d.txt",
fbentdicp="../../data/freebase/entdic.all.map",
numwords=10,
numchars=30,
wordembdim=50,
wordencdim=100,
innerdim=300,
wordoffset=1,
validinter=3,
gradnorm=1.0,
validsplit=100,
):
tt = ticktock("fblextransrun")
traindata, golddata, vocnuments, vocnumwords, datanuments = \
loadlexdata(glovepath, fbentdicp, fblexpath, wordoffset, numwords, numchars)
tt.tock("made data").tick()
# define model
m = FBBasicCompositeEncoder(
wordembdim=wordembdim,
wordencdim=wordencdim,
innerdim=innerdim,
outdim=datanuments,
numchars=128, # ASCII
numwords=vocnumwords,
)
#wenc = WordEncoderPlusGlove(numchars=numchars, numwords=vocnumwords, encdim=wordencdim, embdim=wordembdim)
# train model TODO
tt.tick("training")
m.train([traindata], golddata).adagrad(lr=lr).grad_total_norm(gradnorm).cross_entropy()\
.autovalidate(splits=validsplit, random=True).validinter(validinter).accuracy()\
.train(numbats, epochs)
#embed()
tt.tock("trained").tick("predicting")
print m.predict(traindata).shape
tt.tock("predicted sample")