def f2(input): cnn_fea = [] tmp = input for i in range(nb_layers): tmp = basic_conv2d(nb_filter=nb_filter, nb_row=nb_row, nb_col=nb_col, padding=padding, bn=bn, dropout_rate=dropout_rate)(tmp) tmp = BatchNormalization()(tmp) cnn_fea.append(tmp) # allocate weights to 3 cnn features new_features = [] for fea in cnn_fea: tmp1 = iLayer()(fea) tmp2 = Reshape(([nb_filter * 16 * 8]))(tmp1) tmp3 = Dense(units=1024, activation='tanh')(tmp2) tmp3 = Dropout(rate=0.5)(tmp3) new_features.append(tmp3) # cnn_fea_flatten = Reshape(([nb_layers * nb_filter * map_height * map_width]))(cnn_fea) # conatenate the 3 cnn layer features res = Concatenate(axis=1)(new_features) return res
def f3(data_in): # main_output1 = Reshape((64 * 3 *6* map_height, map_width))(data_in) weighted_data_in = [] for i in range(len(data_in)): tmp = iLayer()(data_in[i]) weighted_data_in.append(tmp) main_output1 = Concatenate(axis=1)(weighted_data_in) main_output2 = LSTM(units=nb_flow * map_height * map_width, dropout=0.1)(main_output1) main_output2 = Activation('tanh')(main_output2) data_out = Reshape((nb_flow, map_height, map_width))(main_output2) return data_out
def f2(input): cnn_fea = [] tmp = input for i in range(nb_layers): tmp = basic_conv2d(nb_filter=nb_filter, nb_row=nb_row, nb_col=nb_col, padding=padding, bn=bn)(tmp) cnn_fea.append(tmp) # allocate weights to 3 cnn features new_features = [] for fea in cnn_fea: new_features.append(iLayer()(fea)) # conatenate the 3 cnn layer features res = Concatenate(axis=1)(new_features) return res