def model():
input1 = Input((50, 200, 3))
input2 = Input((mxlen, ))
cnn_features = conv_net(input1)
embedding_layer = prepare.embedding_layer(prepare.tokenizer.word_index,
prepare.get_embeddings_index(),
mxlen)
embedding = embedding_layer(input2)
bi_lstm = Bidirectional(
LSTM(lstm_unit, implementation=2, return_sequences=False))
lstm_encode = bi_lstm(embedding)
shapes = cnn_features.shape
w, h = shapes[1], shapes[2]
features = []
for k1 in range(w):
for k2 in range(h):
def get_feature(t):
return t[:, k1, k2, :]
get_feature_layer = Lambda(get_feature)
features.append(get_feature_layer(cnn_features))
relations = []
concat = Concatenate()
for feature1 in features:
for feature2 in features:
relations.append(concat([feature1, feature2, lstm_encode]))
g_MLP = get_MLP(4, get_dense(4))
f_MLP = get_MLP(2, get_dense(2))
mid_relations = []
for r in relations:
mid_relations.append(g_MLP(r))
combined_relation = Add()(mid_relations)
rn = dropout_dense(combined_relation)
rn = dropout_dense(rn)
pred = Dense(1, activation='sigmoid')(rn)
model = Model(inputs=[input1, input2], outputs=pred)
optimizer = Adam(lr=3e-5)
model.compile(optimizer=optimizer,
loss='binary_crossentropy',
metrics=['accuracy'])
return model
model = bn_layer(32, 3)(inputs)
model = MaxPooling2D((2, 2), 2)(model)
model = bn_layer(32, 3)(model)
model = MaxPooling2D((2, 2), 2)(model)
model = bn_layer(32, 3)(model)
model = MaxPooling2D((2, 2), 2)(model)
model = bn_layer(32, 3)(model)
model = MaxPooling2D((2, 2), 2)(model)
model = bn_layer(64, 3)(model)
return model
input1 = Input((50, 200, 3))
input2 = Input((mxlen,))
cnn_features = conv_net(input1)
embedding_layer = prepare.embedding_layer(prepare.tokenizer.word_index, prepare.get_embeddings_index(), mxlen)
embedding = embedding_layer(input2)
bi_lstm = Bidirectional(LSTM(lstm_unit, implementation=2, return_sequences=False))
lstm_encode = bi_lstm(embedding)
shapes = cnn_features.shape
w, h = shapes[1], shapes[2]
def slice_1(t):
return t[:, 0, :, :]
def slice_2(t):
return t[:, 1:, :, :]
def slice_3(t):
return t[:, 0, :]