Created by: April0402

How should orthogonal initialization be done for LSTM model? I have tried to setting parameters through this way, but failed.

 for i in range(num_lstm_layers):
         parameters.set('lstm_%d' % i, get_orthoganal_matrix(hidden_dim, hidden_dim))

num_lstm_layers is the number of lstm layers, hidden_dim is the dimension of lstm layer get_orthoganal_matrix is the function of getting orthoganal matrix

And this is the error message.

ValueError: Value shape mismatch, expect (128, 128, 4), should (512, 512)

How should I fix this problem??