Why the testing code only use half of the size of gru bias

[benlin1996]

Thanks for your wonderful code. When I train my own RNNoise model, I found that three gru bias sizes are twice as large as the origin RNNoise in rnn_data.c. But I print out the model in python and sure that my gru bias size is correct.

However, I read the compute_gru function in rnn.c and found that test code only use half of the gru bias in the testing, Can anyone explain why this case happen? And why the origin RNNoise model only keep half of the actual size of gru bias.

Thanks

[redasamik]

في ثلاثاء، 27 نيسان، 2021 في 11:43 ص، كتب benlin1996 < @.***>:

Thanks for your wonderful code. When I train my own RNNoise model, I found that three gru bias sizes are twice as large as the origin RNNoise in rnn_data.c. But I print out the model in python and sure that my gru bias size is correct.

However, I read the compute_gru function in rnn.c and found that test code only use half of the gru bias in the testing, Can anyone explain why this case happen? And why the origin RNNoise model only keep half of the actual size of gru bias.

Thanks

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/xiph/rnnoise/issues/177, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOW7T7DQZG3ZOKAWS3H3AO3TKZ2MPANCNFSM43URAQVA .