Closed leesunfreshing closed 7 years ago
Do you mean striding in the input-to-hidden projection or the hidden-to-hidden? I don't see how you could do strides in the hidden-to-hidden convolution (wouldn't the output get smaller at each timestep?) so it's probably just easier to do an initial tf.layers.conv2d with strides before the RNN, no?
Yeah, I think that is the way. Thanks a lot
On Fri, Jun 16, 2017 at 6:33 PM Carl Thomé notifications@github.com wrote:
Do you mean striding in the input-to-hidden projection or the hidden-to-hidden? I don't see how you could do strides in the hidden-to-hidden convolution (wouldn't the output get smaller at each timestep?) so it's probably just easier to do an initial tf.layers.conv2d with strides before the RNN, no?
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/carlthome/tensorflow-convlstm-cell/issues/13#issuecomment-309073000, or mute the thread https://github.com/notifications/unsubscribe-auth/AGVYgQ3Y1xnpeAHLiFkRLQNb7Sv0wiwpks5sEq5BgaJpZM4N8lGh .
--
It seems setting strides=[2,2] does not work. Can you plz add this option?