def create_model(self):
tdat_input = Input(shape=(self.tdatlen,))
smlnode_input = Input(shape=(self.smllen,))
smledge_input = Input(shape=(self.smllen, self.smllen))
tdel = Embedding(output_dim=self.embdims, input_dim=self.tdatvocabsize, mask_zero=False)
tde = tdel(tdat_input)
se = tdel(smlnode_input)
tenc = CuDNNGRU(self.recdims, return_state=True, return_sequences=True)
tencout, tstate_h = tenc(tde)
wrknodes = se
for k in range(self.config['asthops']):
astwork = OurCustomGraphLayer()([wrknodes, smledge_input])
astwork = concatenate([astwork, wrknodes]) # combine the new node vectors with the previous iteration
astwork = Dense(self.embdims)(astwork) # use a dense layer to squash back to proper dimension
wrknodes = astwork
context = concatenate([tencout, astwork], axis=1)
out = Flatten()(context)
out1 = Dense(self.comvocabsize, activation="softmax")(out)
model = Model(inputs=[tdat_input, smlnode_input, smledge_input], outputs=out1)
if self.config['multigpu']:
model = keras.utils.multi_gpu_model(model, gpus=2)
model.compile(loss='categorical_crossentropy', optimizer=Adam(lr=0.001, clipnorm=20.), metrics=['accuracy'])
return self.config, model
def create_model(self):
tdat_input = Input(shape=(self.tdatlen, ))
sdat_input = Input(shape=(self.sdatlen, self.config['stdatlen']))
com_input = Input(shape=(self.comlen, ))
smlnode_input = Input(shape=(self.smllen, ))
smledge_input = Input(shape=(self.smllen, self.smllen))
tdel = Embedding(output_dim=self.embdims,
input_dim=self.tdatvocabsize,
mask_zero=False)
tde = tdel(tdat_input)
# Adding file context information to the graph2seq model
# note on embedding spaces below
# in this model, the embedding is shared between tdats, sdats, and ast nodes
se = tdel(smlnode_input)
tenc = CuDNNGRU(self.recdims, return_state=True, return_sequences=True)
tencout, tstate_h = tenc(tde)
de = Embedding(output_dim=self.embdims,
input_dim=self.comvocabsize,
mask_zero=False)(com_input)
dec = CuDNNGRU(self.recdims, return_sequences=True)
decout = dec(de, initial_state=tstate_h)
tattn = dot([decout, tencout], axes=[2, 2])
tattn = Activation('softmax')(tattn)
tcontext = dot([tattn, tencout], axes=[2, 1])
semb = TimeDistributed(tdel)
sde = semb(sdat_input)
senc = TimeDistributed(CuDNNGRU(int(self.recdims)))
senc = senc(sde)
sattn = dot([decout, senc], axes=[2, 2])
sattn = Activation('softmax')(sattn)
scontext = dot([sattn, senc], axes=[2, 1])
wrknodes = se
for k in range(self.config['asthops']):
astwork = OurCustomGraphLayer()([wrknodes, smledge_input])
astwork = concatenate([
astwork, wrknodes
]) # combine the new node vectors with the previous iteration
astwork = Dense(self.embdims)(
astwork
) # use a dense layer to squash back to proper dimension
wrknodes = astwork
#astwork = CuDNNGRU(self.recdims, return_sequences=True)(astwork, initial_state=tstate_h)
# attend decoder words to nodes in ast
aattn = dot([decout, astwork], axes=[2, 2])
aattn = Activation('softmax')(aattn)
acontext = dot([aattn, astwork], axes=[2, 1])
context = concatenate([scontext, tcontext, decout, acontext])
out = TimeDistributed(Dense(self.tdddims, activation="relu"))(context)
out = Flatten()(out)
out1 = Dense(self.comvocabsize, activation="softmax")(out)
model = Model(inputs=[
tdat_input, sdat_input, com_input, smlnode_input, smledge_input
],
outputs=out1)
if self.config['multigpu']:
model = keras.utils.multi_gpu_model(model, gpus=2)
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=0.001, clipnorm=20.),
metrics=['accuracy'])
return self.config, model