Fig. 6
ACEP architecture consists of four module R1-R4. The R1-R3 are used to process different sequence information, and the R4 is used to fuse the feature tensors generated by the first three regions. In module R1, the Conv layer and the LSTM layer extract sequence features, and the attention layer predicts scores for different parts in sequences. In module R3, two fully connected layers are used to map 20-dimensional AAC vectors to 64-dimensional feature tensors. In module R4, we use the CVCA layer and the attention mechanism to fuse F1- F3 into Fmeg, then Fmeg is passed to a Sigmoid function to predict results