BiLSTM implementation question

[Matsumotorise]

Hello,

I was wondering if the following would be an implementation bug or not.

for _ in range(n_hidden - 1):
            lstm = LSTM(hidden_dim, return_sequences=True)
            x_pre = lstm(x_pre)
            x_post = lstm(x_post)
        lstm = LSTM(hidden_dim)
        x_pre = lstm(x_pre)
        x_post = lstm(x_post)

Does applying the same LSTM to x_pre and x_post mean they share the same weights? I think Keras's Bidirectional wrapper makes the weights distinct in the forwards and backwards operations. Would this be to reduce model size?

Thank you!

[brianhie]

Good question! This is by design, the weights are shared between the forward and backward direction following https://arxiv.org/abs/1902.08661

[Matsumotorise]

Thanks!