LSTM units impact

[akhaldi]

Hello Mr. Ahmad

• I noticed in most history commits that you are using a consistent 100 units in all LSTM layers. Have you experimented with varying these, such as starting with 20 then moving to 100 or even 200, or the reverse? I'm curious about the impact on performance and efficiency with these configurations. I'm thinking starting with 20 might act as a preliminary filter for simpler features, reducing initial computational load, while in the other hand ending with 20 could focus the network on key aspects, potentially improving generalization.

• about adding Dropout layer after each LSTM to reduce overfitting, have tested this?

thanks in advance