LSTM and GRU Nan Losses.

[mohammadjoshaghani]

Hi, @amirabbasasadi LSTM and GRU creates Nan losses in meta-learning pipeline when using updated requirements.txt
this is because of the last lines of the decoder. I didn't get why it is: Y = 0.5 * (out[:, :, :self.output_size] + out[:, :, self.output_size:]) !
but maybe it would be more precise if it is: Y = 0.5 * (out[:, :, :self.output_size] + out[:, :, self.output_size-1:]). However, this seems it causes issues when time series features are >=2.

[amirabbasasadi]

Hi, @mohammadjoshaghani
Could you please check this error for the case of bidirectional GRU?
The line that you have mentioned is responsible for averaging the result of two directions since the shape of output_tensor for a bidirectional LSTM is (L, N, D∗H_out​)) where D is the number of directions. I guess the problem is associated with one-directional GRU or LSTM. That line should be applied only when directions==2.

[mohammadjoshaghani]

Thanks amir. I get it. Yes it works properly now.

…

On Tue, Dec 6, 2022 at 10:24 AM Amirabbas Asadi ***@***.***> wrote: Hi, @mohammadjoshaghani <https://github.com/mohammadjoshaghani> Could you please check this error for the case of bidirectional GRU? The line that you have mentioned is responsible for averaging the result of two directions since the shape of output_tensor for a bidirectional LSTM is (L, N, D∗H_out​)) where D is the number of directions. I guess the problem is associated with one-directional GRU or LSTM. That line should be applied only when directions==2. — Reply to this email directly, view it on GitHub <#3 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AQ6SV5QDD7HYL4XBZZ7Y5VDWL3PKHANCNFSM6AAAAAASUUV7PQ> . You are receiving this because you were mentioned.Message ID: ***@***.***>