input_length=MAX_LEN,
trainable=False)
ant_prem = antonym_embed(premise)
ant_hypo = antonym_embed(hypothesis)
if args.timedist:
antonym_translate = TimeDistributed(
Dense(ANT_SENT_HIDDEN_SIZE, activation=ACTIVATION))
ant_prem = antonym_translate(ant_prem)
ant_hypo = antonym_translate(ant_hypo)
if args.align_op_we is not None:
alignment = _align(prem, hypo, normalize=True)
alignment_aggr_1 = _aggregate(alignment, args.align_op_we, axis=1)
alignment_aggr_2 = _aggregate(alignment, args.align_op_we, axis=2)
prem_reps.append(alignment_aggr_1)
#prem_reps.append(add(prem,-alignment_aggr_1))
prem_reps.append(multiply(prem, alignment_aggr_1))
hypo_reps.append(alignment_aggr_2)
#hypo_reps.append(add(hypo, -alignment_aggr_2))
hypo_reps.append(multiply(hypo, alignment_aggr_2))
if args.align_op_ae is not None:
alignment = _align(ant_prem, ant_hypo, normalize=True)
alignment_aggr_1 = _aggregate(alignment, args.align_op_ae, axis=1)
alignment_aggr_2 = _aggregate(alignment, args.align_op_ae, axis=2)
prem_reps.append(multiply(prem, alignment_aggr_1))
hypo_reps.append(multiply(hypo, alignment_aggr_2))
# read in embedding and translate
print("> fetching word embedding")
embedding_matrix = get_embedding_matrix(args.embedding, VOCAB, EMBED_HIDDEN_SIZE, tokenizer)
embed = Embedding(VOCAB, EMBED_HIDDEN_SIZE, weights=[embedding_matrix], input_length=MAX_LEN, trainable=False)
prem = embed(premise)
hypo = embed(hypothesis)
translate_1 = TimeDistributed(Dense(SENT_HIDDEN_SIZE, activation=ACTIVATION))
translate_2 = TimeDistributed(Dense(200, activation=ACTIVATION))
t_prem = translate_1(prem)
t_hypo = translate_1(hypo)
t_prem = _aggregate(t_prem, "SUM", axis=1)
t_hypo = _aggregate(t_hypo, "SUM", axis=1)
a_prem = translate_2(prem)
a_hypo = translate_2(hypo)
align = Align(normalize=True)([a_prem, a_hypo])
print(align.get_shape())
conv1 = Conv1D(32, kernel_size=3, activation='relu')(align)
conv2 = Conv1D(64, 3, activation='relu')(conv1)
pool = MaxPooling1D(pool_size=2)(conv2)
flattened = Flatten()(pool)
# assert len(reps) > 0, "no sentence representations, hence no output of the translation layer"
prem = translate(prem)
hypo = translate(hypo)
alignment = _align(prem, hypo, normalize=True)
prem_c = _softalign(prem, alignment, transpose=True)
hypo_c = _softalign(hypo, alignment)
prem = concatenate([prem, hypo_c], axis=-1)
hypo = concatenate([hypo, prem_c], axis=-1)
translate2 = TimeDistributed(Dense(SENT_HIDDEN_SIZE, activation=ACTIVATION))
prem = translate2(prem)
hypo = translate2(hypo)
prem = _aggregate(prem, "SUM", axis=1)
hypo = _aggregate(hypo, "SUM", axis=1)
joint = concatenate([prem, hypo])
joint = Dropout(DP)(joint)
for i in range(3):
joint = Dense(2 * SENT_HIDDEN_SIZE,
activation=ACTIVATION,
kernel_regularizer=l2(L2))(joint)
joint = Dropout(DP)(joint)
joint = BatchNormalization()(joint)
pred = Dense(3, activation='softmax')(joint)
model = Model(inputs=[premise, hypothesis], outputs=pred)
model.compile(optimizer=OPTIMIZER,
input_length=MAX_LEN,
trainable=False)
ant_prem = antonym_embed(premise)
ant_hypo = antonym_embed(hypothesis)
if args.timedist:
antonym_translate = TimeDistributed(
Dense(ANT_SENT_HIDDEN_SIZE, activation=ACTIVATION))
ant_prem = antonym_translate(ant_prem)
ant_hypo = antonym_translate(ant_hypo)
reps = [] # sentence representations
if args.agg_we is not None:
prem_aggr = _aggregate(prem, args.agg_we, axis=1)
hypo_aggr = _aggregate(hypo, args.agg_we, axis=1)
reps.append(prem_aggr)
reps.append(hypo_aggr)
if args.align_op_we is not None:
alignment = _align(prem, hypo, normalize=True)
alignment_aggr_1 = _aggregate(alignment, args.align_op_we, axis=1)
alignment_aggr_2 = _aggregate(alignment, args.align_op_we, axis=2)
reps.append(alignment_aggr_1)
reps.append(alignment_aggr_2)
if args.agg_ae is not None:
ant_prem_aggr = _aggregate(ant_prem, args.agg_ae, axis=1)
ant_hypo_aggr = _aggregate(ant_hypo, args.agg_ae, axis=1)
reps.append(ant_prem_aggr)